Until September 6, almost nobody knew this film existed. It premiered at a surprise screening at the Telluride Film Festival, in Colorado, and within weeks it became the documentary everyone was talking about: lines around the block at the New York Film Festival, first place at the US box office over Halloween weekend, European viewers catching flights to see it sooner. Portugal got lucky. The film opens there on November 12; Spain has to wait until Christmas Day.

“You Can See Everything”, by Nathan Fielder (who is also its main character) and Lance Oppenheim, makes us question what we are seeing, the people we are watching, and even our own ability to understand them. And it does so for almost three hours. It is a fascinating, sometimes uncomfortable piece of cinema, and one of the strangest and most memorable documentaries in years. Every viewer will leave with their own reading, as in a Rorschach test (the psychological test with inkblots in which each person sees something different). We take the film apart as we watch it, and that is part of the pleasure. It feels like honest filmmaking, one that doesn’t try to impose an interpretation.

At times it comes close to satirical comedy or a psychological thriller. Without a hint of action, it keeps us glued to our seats, wanting to know and see more. That is what the film is, above all: an experience to be lived (and felt).

Elizabeth Holmes, the founder of Theranos, promised to revolutionize medicine with machines that could run hundreds of lab tests from a few drops of blood. She won over investors, conquered Silicon Valley and was hailed as a kind of Steve Jobs of medicine. The technology never did what she promised. The company collapsed, Holmes was convicted of fraud and, in 2023, began serving a prison sentence in the United States.

Thirty-four days before reporting to prison, Holmes and her partner, Billy Evans, invited a film crew to follow their family life.

Anyone who knows Fielder’s work, especially “Nathan for You” and “The Rehearsal”, will recognize some of his trademarks: the awkward humor, the discomfort, the long pauses, the moments when we no longer know who is performing and who is being genuine. Some scenes feel lifted from a particularly strange reality show; others make you wonder whether any screenwriter could have imagined them. The most impressive part is how all of this becomes cinema.

A film that took its time, and it paid off

Fielder and Oppenheim are interested in the experience of being with these people, trying to understand them, and finding out how far that understanding can go. The film takes its time to reveal what it has to show. Sometimes, a long time. There is a deliberate wait, both in how the story is built and inside the scenes themselves.

Fielder has an unusual talent for letting silence do the work. He waits after an answer, looks at the person in front of him, waits a little longer. Often, it is in that gap that we start thinking about what we just heard.

The craft is remarkable. David Bolen’s cinematography, Adam Locke-Norton’s editing and, above all, Ari Balouzian’s score give the documentary a cinematic identity of its own. All three won at the Critics Choice Documentary Awards on November 8. The music takes us to different emotional places and gives seemingly ordinary situations an unexpected tension. Some images and moments stay with you long after the end (a few have already become memes).

The film could be half as long and still be interesting, but it would lose much of what makes it special. The silences, the detours and the long conversations are part of an experience you feel intensely.

The silences, the detours and the long conversations are part of an experience you feel intensely.

Elizabeth Holmes and her grip on reality

One of the most revealing sequences starts with a seemingly simple argument about the meaning of one word: can. Holmes had claimed that Theranos machines could run certain tests. But what did she really mean? That the machines could do it at that moment? Or that one day they might?

Fielder presses on the difference. Holmes looks for explanations, shifts the meaning of words and seems unable to see the contradiction. The conversation becomes so absurd it almost feels like a comedy sketch.

And that is exactly where Fielder decides to go. He shows her Abbott and Costello’s famous “Who’s on First?”, an old comedy routine built on confusion over the names of baseball players. Who plays first base, What plays second and I Don’t Know plays third. Holmes laughs, but when Fielder asks her to explain the joke, she doesn’t seem to have understood it.

Holmes seems to have a peculiar relationship with words and with the reality they stand for. When it suits her, she finds alternative meanings to justify statements that look plainly false to everyone else. And yet she struggles to follow a joke built precisely on the ambiguity of words.

Does she really believe what she says? How well can she tell a promise from a concrete reality? Fielder never finds a definitive answer, and neither do we.

Fordham University psychologist Barry Rosenfeld offered one reading after watching the film: he found Holmes’s behavior manipulative, while cautioning that diagnosing someone from a documentary would be inappropriate. It is an interpretation, not a definitive answer.

Silicon Valley and the stories that move the world

The documentary also made me think about something it never puts front and center: Silicon Valley’s ability to create narratives. A company is, to a large extent, a story that a group of people believes in enough to spend years of their lives making it real. You have to convince workers, investors, customers and, often, the rest of the world that there is a future worth building. We have done this since we first organized into communities. Silicon Valley took it to an extraordinary scale.

There is something wonderful about this ambition, sometimes delusional, that created technologies and companies capable of changing the world. But overconfidence and an obsession with growth can drift away from reality. The famous fake it till you make it has limits that some founders cross systematically.

It’s impossible not to think of other portraits of this world: HBO’s “Silicon Valley”, an absurd comedy that is frighteningly close to reality; “Super Pumped”, about Uber, Travis Kalanick and the most self-centered side of business ambition; or “The Dropout”, about the rise and fall of Holmes herself.

“You Can See Everything” adds another angle. It is less interested in the technology than in the person who built that story: a woman who dreamed of becoming a historic Silicon Valley figure and who, even after losing almost everything, seems to hold on to that dream. Holmes and Evans even talk about a new blood-testing venture. The ambition is still there.

How can someone keep living inside a story after the world around them has disproven it?

In Holmes’s case, there was a fraud proven in court. Still, the film leaves us with a hard human question: how can someone keep living inside a story after the world around them has disproven it?

———————————

SPOILER WARNING: from here on there are moments from the film’s second half, although the film is worth it for the experience more than for the events.

A love that is hard to understand

As the documentary goes on, Billy Evans becomes as fascinating and puzzling a figure as Elizabeth Holmes. There is something strange about his personality, the way he answers, the way he reacts to other people and, above all, the way he seems to control much of his partner’s life.

Evans holds a power of attorney that lets him act on Holmes’s behalf while she is in prison. He runs her social media, replies to people as if he were her and gets involved in projects that seek to cash in on her name and image. He turns to artificial intelligence and lets billboards with a photo she dislikes spread around town.

And yet there is also love, or at least that is the impression many of the images give. In a particularly intimate sequence, Evans talks about how long he has gone without sex since Holmes went to prison. He counts the days. It’s 899, about to hit 900. The passing years also seem to mark his face and his behavior.

The family’s silences

There is a kitchen scene that strikes me as especially important. Elizabeth’s father, Christian Holmes, starts talking about difficult episodes in his daughter’s past. Among other things, he mentions that she was allegedly raped. Evans steps in and cuts the conversation short.

It’s an unsettling moment, and the documentary is full of them. Someone says one thing, leaves another unsaid, changes the subject or seems to hide something. Sometimes what isn’t said becomes as important as what we hear.

In another sequence, Holmes and Evans talk excitedly about that new blood-testing venture. Elizabeth’s parents are in the room and, at one point, they hold hands. It’s a small gesture, but to me it carries an enormous amount of doubt and emotion.

Amanda Seyfried and the documentary’s Russian dolls

In “Hamlet”, the prince has a company of actors stage his father’s murder so he can watch his uncle’s reaction. He wants to use a performance to get to the truth. For more than four centuries, literature and film have kept coming back to this idea, which the French call mise en abyme: the story within the story, the play within the play. Fielder takes it somewhere very much his own, and it works.

When Holmes goes to prison, the film seems to be ending. In fact, it is turning into something else. Enter Amanda Seyfried, who played Holmes in the series “The Dropout”, and we start watching real people face the fictional versions of their own lives.

Seyfried took a real risk. As she and Fielder told Rolling Stone, she started filming without her agent or lawyer knowing and had no approval over any of the footage; her lawyer had to put a deal in place after the fact. She kept asking herself, in her words, “Am I going to destroy my career in some way?”

Seyfried plays Holmes again in the scenes Fielder couldn’t film, including his prison visit. She is so convincing that for a moment we forget it is a reenactment, especially since minutes earlier we were watching the real Elizabeth Holmes. Then she appears as herself. She meets Billy Evans and even speaks briefly with Holmes on a call from prison, a deeply strange moment in which the real woman, her partner and the actress who played her cross paths.

The conversation with Evans gets tense. He criticizes her for having played Holmes on TV, the argument slides into politics, and Seyfried comes away convinced that Holmes should leave him. Fielder disagrees: he saw love between them, however strange the relationship looks. Two attentive people read the same gestures in opposite ways, and Seyfried ends up admitting it is all far more complex than it seemed. That may be the line that best sums up the film.

Fielder himself also becomes the object of our scrutiny. How much is he observing, and how much is he shaping what happens? What responsibility does he bear for the story he is building?

Can we really see everything?

The title gains meaning as the film goes on. The exposure is enormous. We see family conversations, intimate arguments, moments of great vulnerability, even Evans completely naked in the shower. At times the access seems to have no limits at all. And yet Fielder doesn’t seem to want to find out why Holmes did what she did. He is more interested in how her mind seems to work (which also tells us something about the world of Silicon Valley founders) and in what it is like to live in her skin now, between prison, her partner, her parents and a story nobody believes anymore.

Near the end, after the reenacted visit with Seyfried, the film shows real footage of Holmes with Evans in the prison yard, shot through the fence, with gestures of closeness and affection. Those images seem to back Fielder’s reading: their relationship has a family and emotional side that goes beyond business and possible self-interest. But they don’t settle the doubts about either of them.

All of this also raises ethical questions. The filmmakers expose real people in extremely intimate, sometimes uncomfortable or humiliating situations. And we, the audience, take part in that exposure.

A film that doesn’t settle for little

Interestingly, some Portuguese critics received the film rather coldly. Luís Miguel Oliveira, in Público, one of Portugal’s leading dailies, wrote that it settles for little: being a meme generator. Eurico de Barros, in Observador, felt the film throws dust in our eyes while pretending to open them. I enjoy reading Eurico de Barros, but this time I strongly disagree. Turning Público’s headline around, I’d say “You Can See Everything” does not settle for little. And I’m not alone. Abroad, the reception was very different: Danny Leigh, the Financial Times critic, gave it five stars and admitted that, despite watching several films a week for a living, he rarely sees anything this mind-bending. If ever.

The title promises that we can see everything. Fielder pushes that promise almost to the limit. But after nearly three hours watching Elizabeth Holmes and the people around her, we still don’t know exactly what goes on in their heads, or what drives them. Maybe that is never possible, but it is still fascinating to try (and the film shows us, taken to the extreme, how easy it is to live inside stories and extreme illusions, so often rewarded in Silicon Valley).

———

“You Can See Everything”

Directed by: Nathan Fielder, Lance Oppenheim
With: Elizabeth Holmes, Nathan Fielder, Billy Evans, Amanda Seyfried
Cinematography: David Bolen · Editing: Adam Locke-Norton · Music: Ari Balouzian
Documentary · USA · 2h54 · Rated M/12 in Portugal
Distributor in Portugal: Cinemundo · Portuguese release: November 12, 2026

(A note in English: this is about the 2,891 amendments that Portugal’s parties filed for the 2027 state budget. I put them all on one searchable page and looked for patterns, like 44% arriving on a single day and the share approved falling from 21.6% in the 2020 budget to 7.8% in 2026. It’s in Portuguese because the story is about Portuguese politics.)

Todos os anos, antes da votação na especialidade, os partidos entregam no Parlamento milhares de propostas de alteração ao Orçamento do Estado. Quase ninguém as lê (pelo menos essa parece-me a conclusão mais óbvia). São PDFs soltos, um por proposta, com artigos, valores e notas justificativas.

Este ano (ano da graça de 2026), depois de ver uns quantos tweets no Twitter (X) sobre o tema, decidi juntar as propostas todas. Com ajuda do meu agente pessoal, claro, que não tenho tempo para ler tudo (nem vontade). E é para isto que “deus” criou a inteligência artificial que, de facto, analisa incrivelmente bem, e cada vez melhor, muitos documentos de uma só vez.

Ao todo, foram 2.891 propostas para o OE 2027, a partir da plataforma independente Voto a Voto (“fotografia” das 20h58 de 7 de Novembro, que pode ainda mudar), e 2.882 PDFs que consegui guardar e transformar em texto pesquisável.

O resultado está em joaotome.com/data-stories/orcamento-2027 (é mais uma das Data Stories que tenho tentado fazer por mera carolice de pessoa curiosa vitaminada pela AI ou IA, em bom tuga). Deixo aqui as conclusões principais, alguns casos curiosos e uma explicação rápida do que se pode fazer na página.

O que os números mostram

• Quase metade chegou no último dia. 1.283 propostas (44%) foram registadas a 6 de Novembro, o último dia com registos. O PS entregou as suas 208 todas nesse dia. No Chega e no PCP, os dois partidos com mais propostas, foi menos de metade (37% e 40%).

• Chega e PCP são os que mais propõem. O Chega tem 713 propostas, o PCP 655 e o Livre 506. No outro extremo, a IL apresentou 55, e o PSD e o CDS-PP, em conjunto, 69.

• Cada vez mais propostas, cada vez menos aprovadas. No Orçamento de 2020 foi aprovada mais de uma proposta em cada cinco (21,6%). No de 2026, menos de uma em cada 12 (7,8%), com 65% mais propostas apresentadas (de 1.333 para 2.193).

• Ter maioria muda pouco. Nos sete orçamentos de 2020 a 2026, os governos minoritários receberam em média 1.801 propostas e os de maioria absoluta, 1.777. São só sete casos, mas a diferença é quase nula.

• A habitação lidera, mas cada partido tem a sua área. Infraestruturas e habitação é a área maior, com 503 propostas. Na IL, pesa 44% de tudo o que o partido propõe. No PAN, 27% das propostas são de agricultura e pescas, quase quatro vezes a média.

• As palavras “denunciam” cada partido. O PAN escreve «animais» em 32% das propostas; os outros partidos, em 2%. O PCP fala de trabalhadores e transferências, o Chega de militares e pensões, o JPP de insularidade e a IL de contas e poupança.

Cada ponto é uma proposta, no dia de registo. Fonte: Voto a Voto, fotografia das 20h58 de 7 de Novembro de 2026.

Propostas apresentadas e aprovadas em cada Orçamento, de 2020 a 2026. Fonte: Assembleia da República, relatório de redacção final do OE 2026.

Os casos curiosos

No meio de propostas sobre impostos, saúde e casas há medidas muito concretas. Algumas chegaram-me por sugestão de um leitor e pelo projecto Estado de Gozo. Confirmei cada uma no texto parlamentar.

• Pombais com alimentação contraceptiva (PAN, 793C). Financiar, com os municípios, pombais e comida que trava a reprodução dos pombos. O objectivo é controlar a população nas cidades sem os matar.

• Um santuário para aves de companhia (PAN, 1220C). Para periquitos e canários criados em cativeiro que já não conseguem voltar à natureza.

• Um milhão de euros para cafés de reparação (Livre, 53C). Uma rede de espaços onde voluntários e técnicos ajudam a arranjar bicicletas, roupa e outros objectos.

• Touradas na RTP (Chega, 2770C). A televisão pública transmitiria espectáculos tauromáquicos durante 2027, com os meios que já tem.

• Estacionamento grátis em Coina (PCP, 2535C). Mais lugares públicos e gratuitos junto à estação. A mesma estação aparece numa proposta do Livre (1578C), que pede um estudo dos acessos a Coina e ao Fogueteiro.

• O impacto ambiental ao lado do preço (PAN, 2777C). Mostrar, nos produtos vendidos a retalho, o impacto ambiental da sua produção em CO2 equivalente.

• Uma gralha no texto do próprio Governo. A proposta do Governo identifica a lei do Orçamento de 2026 como Lei n.º 73-A/2024. No Diário da República, é a Lei n.º 73-A/2025. Falta confirmar se já foi corrigida.

Seis das propostas curiosas, cada uma com ligação para o texto integral. Fonte: textos parlamentares do OE 2027.

O que se pode fazer na página

• Ver quando chegaram as propostas. No primeiro gráfico, cada ponto é uma proposta, no dia em que foi registada. Pode escolher-se um partido, passar o rato por um ponto e clicar para abrir o texto.

• Explorar as 2.891. Há filtros por partido, área e pesquisa, e uma lista que leva ao PDF de cada proposta.

• Uma nuvem de palavras que muda com a selecção. O tamanho mostra quantas vezes a palavra aparece nos textos. Com um partido escolhido, a cor mostra as palavras que esse partido usa mais do que os outros. Tirei os nomes de pessoas e as fórmulas parlamentares, que de outra forma enchiam a nuvem.

• Partidos e áreas lado a lado. Uma tabela de cores com o peso de cada área em cada partido, e um botão para comparar com a média.

• As palavras de cada partido. Para cada um, as cinco palavras que usa mais do que os restantes, com a percentagem de propostas onde aparecem.

• Uma avaliação editorial. Cinco medidas relatadas pela imprensa, 15 ideias com resultados estudados noutros países e 21 propostas para reduzir a burocracia, com notas de 0 a 100. São juízos meus, preliminares, e não previsões.

Peso de cada área nas propostas de cada partido. Fonte: Voto a Voto, áreas atribuídas pela plataforma.

As cinco palavras que cada partido usa mais do que os restantes. Fonte: texto extraído de 2.881 PDFs das propostas.

A nuvem de palavras com o PAN escolhido: a cor mais escura marca as palavras que o partido usa pelo menos duas vezes mais do que os outros. Fonte: texto extraído dos PDFs das propostas.

Uma nota de cautela importante em qualquer análise: Propostas não são leis, e contar propostas não diz nada sobre o custo ou a importância de cada uma. As áreas são as que a plataforma atribuiu, e não as auditei uma a uma. É uma investigação em construção. Volto ao tema quando se souber o que foi aprovado.

Mais sobre os meus projectos no sítio do costume: joaotome.com/

———

Dados: Voto a Voto, fotografia das 20h58 de 7 de Novembro de 2026, com dados indicados como vindos da Assembleia da República; 2.882 PDFs guardados, 2.881 com texto utilizável. Série de 2020 a 2026: Assembleia da República, relatório de redacção final do OE 2026.

I expected a much more obvious satire, with Tom Cruise completely over the top, playing a powerful, exaggerated oil oligarch convinced he can do whatever he wants without consequences. One of those men whose decisions end up putting humanity at risk. That is indeed the premise of the film, but the story takes a different direction from what I had imagined.

I liked it quite a lot, although some aspects disappointed me. The excessive acting sometimes made the experience less enjoyable than it could have been, but the narrative construction and what we gradually discover about the characters ended up surprising me. Iñárritu hides several layers within a story that initially seems much simpler.

And I think it’s important that films like this continue to exist. I have the feeling that cinema no longer has quite the same ability to define a generation or influence society as it once did, perhaps because of the amount of content we consume and how fragmented our attention has become. I think of Her, and how it explored the relationship between humans and artificial intelligence, or The Social Network, with its portrayal of the people who created one of the companies that most influenced our digital lives. And there are films like Dead Poets Society, which had a personal impact on me and which I still remember today.

I don’t know whether Digger will be one of those films. But I’m glad there are still directors willing to take risks, even when the results divide audiences. Unfortunately, the box-office numbers don’t help: the film made just $8 million in the US and Canada during its opening weekend, finishing fifth despite an initially reported production budget of around $125 million. That’s a commercial failure for a Tom Cruise film, and it leaves me concerned about the future of ambitious, original productions like this.

Iñárritu himself acknowledged in an interview with The New York Times that the film isn’t perfect, but said he prefers strong reactions to a lukewarm response. I understand what he means. During the screening, there were moments when I worried that the exaggeration was making everything too elaborate and chaotic. As the story progressed, I began to understand the intention behind those choices, even when I didn’t agree with all of them.

There are still things I’m trying to interpret. It’s one of those films I almost want to watch again, because I suspect I missed some of its layers the first time around.

Major spoilers for Digger from here on

There’s a reason for the exaggeration

One of the first surprises is discovering that what we’re watching isn’t exactly the reality of the film. The story of Digger Rockwell, the oil magnate played by Tom Cruise, is part of a stage production called Fossil Fools. The man we’re actually watching is Curtis Markson, an actor portraying Digger during a rehearsal.

The revelation comes relatively early. Iñárritu doesn’t save the surprise for the ending, which is a curious choice. From that point onwards, we follow not only the story being performed, but also the actors, their conflicts and their attempts to understand events from another time.

It’s impossible not to think of Birdman, one of my favourite Iñárritu films. There, too, theatre played a central role, mixing fiction and reality, with actors tormented by the characters they portrayed, their ambitions and the problems in their own lives. In Digger, we find some of those elements, although the purpose is different.

Gradually, we discover that the play is inspired by a catastrophe that actually happened in that fictional world around 25 years earlier. The decisions made by Digger and other powerful figures contributed to an environmental crisis that profoundly transformed the planet. The story we’re watching is a satirical version of those events.

And this is where the exaggeration that initially bothered me starts to make more sense.

I later read the Variety review, which compares Digger to an attempt at making a Dr. Strangelove about climate change, but argues that it lacks the humour and subtlety of Stanley Kubrick’s film. I understand that perspective. There are moments when the characters seem to be shouting more than acting, and that excess weighs on the experience. I think Birdman was much more successful in the way it drew us into the actors and their emotions.

But there’s a narrative intention behind all this, and it becomes particularly clear during a discussion between Curtis and Hannah, the theatre director played by Sophie Wilde. Curtis wants to understand Digger better, find some truth in the character and portray him accurately. Hannah wants to preserve the exaggeration of the play, which has been around for many years and serves a very specific purpose for its audience.

These are two different perspectives on the role of art. Curtis wants to understand the real man behind the caricature. Hannah is more concerned with the audience’s reaction and what the play gives them.

Then there’s a third character who helps explain all this: Kenny Rockwell, played by Jesse Plemons. Kenny is the real Digger’s nephew, appears as a child in the events represented by the play, and is also the man financing the production. It seems strange that someone from the family would be willing to spend so much money on a work that ridicules his own uncle, but Kenny has his reasons.

In the society they live in, deeply affected by the consequences of the catastrophe, the play offers a form of escape. People can laugh at those responsible for destroying the world, ridicule their priorities and experience a kind of symbolic revenge. They can’t change what happened, but they can watch the humiliation of those they consider guilty.

And Kenny is hardly someone removed from power. He’s a very wealthy man, connected to a security and surveillance company. The family continues to exercise influence, although in a different way, and can now also play a part in how the past is told. That contradiction adds another layer to the story.

The discussion between Curtis and Hannah also made me think about Iñárritu’s own directing choices. In Birdman, the camera is often very close to the actors, following their movements and emotions. In Digger, during the rehearsals, the approach is frequently more distant. I understand that this distance is part of the narrative construction and gives us clues about what we’re actually watching, but I feel that some scenes lose cinematic power as a result.

I particularly remember Digger’s speech before an international assembly, where he defends oil as one of the great sources of freedom in modern society. The camera frequently looks down at him from above, making the character appear smaller. It’s a choice that makes sense if the intention is to visually diminish a man so convinced of his own power.

But I imagined how the same scene would look if it had been filmed from below, giving him a more imposing presence. Perhaps it would have had more visual impact, although it would also have communicated a different message. It’s precisely this relationship between cinematographic choices and the story being told that I find interesting. I can understand Iñárritu’s intention without thinking that every choice works.

The same applies to the exaggerated performances. They’re deliberate and serve a purpose, but that doesn’t mean they can’t make certain scenes exhausting. Still, what initially appears to be a weakness of the film ends up revealing something about the show those characters are putting together.

The adults laugh. The children, not so much.

One of the moments I kept thinking about after leaving the cinema was the ending of the play. Digger ends up taking his own life by jumping from an oil platform, and the audience responds with enormous enthusiasm. The adults laugh hysterically, insult him and almost cry with laughter. There’s a sense of collective satisfaction in seeing that man finally defeated.

But I noticed that the children in the audience didn’t seem to react in the same way. Some looked uncomfortable, while others didn’t seem to understand what was so funny. I kept wondering why Iñárritu had chosen to show that difference.

The first interpretation is relatively simple. The children are watching a man kill himself. They don’t necessarily have the same historical context or the same anger towards the character. The adults, on the other hand, have lived more directly with the consequences of the catastrophe and see his death as a form of revenge, almost a form of justice.

But I think there may be something else. For those children, the dystopian reality they live in is the only world they’ve ever known. They don’t necessarily remember an earlier society, cities without those environmental changes or a different way of life. The play represents a story that shaped their lives, but they haven’t experienced it in the same way as the adults.

And then there’s a third interpretation that occurred to me. Those historical figures, however much destruction they caused, will eventually be forgotten or remembered differently. Society moves on, generations change, and the people responsible for terrible events become characters in an increasingly distant story. The hatred felt by one generation may not make as much sense to the next.

I don’t know whether Iñárritu had all this in mind, but I found the difference in reactions very revealing. For the adults, the play is an opportunity for emotional release. For the children, it may just be an exaggerated representation of events they know through stories they’ve been told.

That’s also why the discussion between Curtis and Hannah takes on another meaning. By turning a real person into a caricature, the play ends up influencing how that society remembers him. That representation may make sense to one generation while losing some of its meaning for the next.

When we leave the theatre

For me, the ending was what made me appreciate Iñárritu’s narrative construction the most.

After the play ends, there’s applause, laughter and mockery directed at Digger. They even pour a little fake oil over Curtis’s bald head when the actors return to the stage to take their bows. The humiliation continues even after the performance is over.

Shortly afterwards, an alarm sounds. A level-four warning is announced concerning the air conditions outside, and we realise that everyone has to put on a protective mask before leaving the theatre.

There’s something curious about how naturally people do this. They’ve just watched a comedy about a catastrophe, laughed at those responsible, applauded the death of the main character, and now they’re preparing to return to reality.

It’s when the camera finally starts moving outside the theatre that we truly understand what happened to the world. We find a profoundly transformed New York, with flooded streets, wrecked cars, makeshift platforms for people to move around, a militarised police presence and an atmosphere of enormous social tension.

Until then, we’d been given a few clues, but we’d never seen the full scale of that reality. The camera follows the spectators, still wearing their masks, and shows us a society that continues to live with the consequences of the catastrophe.

There’s also a message on a wall that caught my attention: “Compliance is Freedom.” Obedience is presented as freedom, which takes on another meaning when we remember Digger’s speech about oil.

In that scene, he argues that fossil fuels brought freedom to modern societies, making transport, mobility and many of the conveniences we’ve grown accustomed to over the decades possible. It’s a speech by a deeply selfish character who uses those arguments to justify his own interests, but there’s a question there worth considering: how willing are we really to give up the conveniences that oil has provided?

The contrast with the words on the wall is particularly ironic. The freedom associated with oil, mobility and consumption has given way to a society where survival seems to require new forms of control and obedience.

But there’s another moment at the end that I think is one of the film’s best, and it’s worth staying for.

During the closing credits, after the camera has shown us that devastated city, the music suddenly stops. Everything goes silent, and we hear the same warning sound that was played earlier in the theatre. Once again, it’s the level-four alert, the one that forced people to put on their masks before leaving.

I loved this moment. It’s as though Iñárritu, after showing us an audience watching a play about a catastrophe, is now addressing the audience that has just watched his film. For a few seconds, the boundary between the two experiences seems to disappear.

And that’s precisely when my own experience at Cine-Teatro Turim took on another meaning. I was in a small cinema in Benfica, Lisbon, where I’d heard people laughing, gasping and reacting with some confusion to the more exaggerated scenes. Just like the audience in the play, we’d finished watching a story about a ruined world and were about to leave the theatre.

The film creates a strange feeling, and it works remarkably well: for a moment, it almost seems as though we’re going to find that same dystopian world outside when we leave the cinema.

And perhaps there’s also a provocation here aimed at us, as spectators. We can leave a cinema convinced that we understand a problem better, or even with the feeling that we’ve done something just by thinking about it. The audience of Fossil Fools also leaves the play satisfied, having laughed at those responsible for the catastrophe. But the world outside remains exactly the same.

After the applause, they put on their masks and return to the devastated city. And when we hear that final level-four warning in the cinema, for a few seconds it almost feels as though it’s our turn to leave.

“I’ve always thought about my relationships with people as finding two people that hadn’t met before and plugging them together like a tinker toy and wondering, you know, what will happen if we do that?”

— Vint Cerf, talking to me this week for Future Stories - OurFutureStories.eu

That last part, “wondering what will happen” while being curious, is the job description I’d give myself if anyone asked. The last two weeks were a good example.

It started in Panama, at IPinfo’s team offsite. On Friday, my colleague Ross and I visited a data center, talked about how the company sees the Internet, and were surprised by how hot it was outside and how cold it was indoors. Panama is hot, wet and green in a way that Lisbon isn’t. It reminded me of Malaysia when I visited, as both countries are places where, when it rains, it really pours.

On Friday, October 9, a magnitude 7.7 earthquake hit the Azuero Peninsula, southwest of where we were. You don’t forget the moment a building starts moving under you. In my case, it was particularly strange because I was in a movie theatre at the Panama Canal’s Miraflores Visitor Center, where people can watch ships passing through the canal.

I was wearing 3D glasses, sitting in a huge IMAX theatre and listening to what sounded like an AI-generated version of Morgan Freeman’s voice (I don’t think it was actually Morgan Freeman). The documentary was quite good, and I was sitting next to Ross and Pablo, one of the directors of the local Internet exchange.

Near the end of the film, our seats and the floor started shaking. At first, we all thought it was part of the experience. Perhaps it was a 4D movie after all, with hydraulic systems underneath the seats making everything move. But then parts of the ceiling started falling, and the enormous screen began shaking too. It quickly became clear that this was an earthquake. People started shouting and making their way towards the exits in the dark.

The floor was shaking violently, even as we made our way outside into the bright sunshine. People were pouring out of buildings, including those around the canal. One person became unwell, apparently with a heart-related problem.

When we reached the main road with Pablo, we found traffic jams everywhere. People were trying to get home to check whether everything was intact or to pick up their children. Crowds had gathered outside buildings, including universities, shops, skyscrapers and hotels. You could even see water spilling from swimming pools on the upper floors of some buildings.

We eventually made it back to Casco Viejo, the historic part of Panama City where our hotel was, but had to walk much of the way because several streets were closed. Our plans to watch ships passing through the canal were obviously over. The canal itself continued operating, but public visitor areas were closed, and the airport also closed several times because of the strongest aftershocks.

A few hours later, I was live on CNN Portugal, explaining what it felt like. Some journalist friends had noticed I was in Panama and passed along my contact details. Early on Saturday morning, Panama time, I also recorded a piece for the Portuguese national radio stations M80 and Rádio Comercial.

The aftershocks kept coming that night and the following day, including some strong ones. My flight home was cancelled, which meant a longer journey back to Lisbon. I finally arrived late on Sunday.

First, the scare, but then curiosity was always around. Back at the laptop, we checked what IPinfo’s data showed. In the provinces closest to the epicentre, the share of IP addresses responding to our checks fell to 63% of normal in the first hour, while Panama City was mostly back to normal within an hour (at least one of the local operators was briefly impacted — it was much worse 400 km from there, where the earthquake epicenter was).

I also got to see parts of the Panama Canal (and ships lining up to go through) during those strange days, and it’s hard not to think about Internet plumbing when you’re there. About 80 kilometres separate the Pacific from the Atlantic, and the same narrow strip of land that carries ships between oceans is also an important corridor for Internet cables — five come through the Pacific. The jungle around it matters just as much: the canal depends on rainwater from the surrounding watershed. I’ll write more about Panama, the cables and that earthquake in the coming weeks.

Internet routes and Hacker News

Then came the RPKI post, my second for IPinfo. RPKI is one of those pieces of Internet infrastructure that almost nobody outside networking has heard of, yet it matters to everyone. It helps networks check whether another network is allowed to announce a particular block of IP addresses, making it harder to misdirect Internet traffic.

RPKI now protects about 70% of Internet routes, up from 54% a year ago. But it’s far from bulletproof. In August, a hijacked route passed as valid because the network’s own authorisation record permitted it, and about one in six protected IPv4 routes has a similar weakness.

The post reached the Hacker News front page, which is always great, given that I truly value that audience :D. Much of the discussion was about who controls the Internet’s trust anchors, whether governments could interfere with them, and why China’s RPKI adoption increased so dramatically in a single year.

French supermarket fridges

Next came some investigation about the French supermarket fridges (the story that broke on Tuesday). After a cyberattack reportedly forced refrigeration systems into defrost mode in French supermarkets on October 11 and 12, we looked at IPinfo’s scan from five days earlier.

We found 2,528 Danfoss refrigeration dashboards reachable from the public Internet, with 1,331 of them in France, more than in the rest of the world combined. That doesn’t mean those particular systems were compromised, but it shows how much equipment of this kind can be exposed online. We shared the findings on social media already, and the full post is coming soon.

From Carl Sagan to Vint Cerf

In the middle of all this, Neil deGrasse Tyson was in Lisbon for a Startalk podcast recording.

He’s one of those people who makes science feel like an invitation, the way Carl Sagan did before him. Both have spent much of their careers trying to make people want to know more about the universe. I relate to that.

I even got to bring Sagan into my conversation with Vint Cerf. Sagan once warned:

“We’ve arranged a global civilization in which most crucial elements profoundly depend on science and technology. We’ve also arranged things so that almost no one understands science and technology.”

The Internet is a good example of both parts of that sentence.

Which brings me back to Vint. This week I spoke with him for Future Stories (a bucket list event to be honest), my new independent project, about how the Internet came to be, the decisions behind it, and where it goes next (spoiler: considerably further than Earth). There’s a short teaser on the site already. The full conversation is coming soon, and I’ll keep the rest for then.

Not bad for “just” a curious guy (to hear with the tone of the John Lennon song, “I’m just a jealous guy”). Stay tuned, geek out.

In 1977, Vint Cerf had to decide how many addresses the Internet would need. He was running the Defense Department’s Internet research program. The engineers had argued about address size for a year without agreeing, so Cerf picked one himself: 32-bit numbers, about 4.3 billion addresses. That seemed plenty for what was still an experiment.

When that format was published as RFC 791 in September 1981, the ARPANET had 213 host computers. “The problem is the experiment never ended,” Cerf said in 2010.

The addresses were handed out as if they would never run short, and some of the biggest blocks went to some interesting organisations — Ford among them. More on that later.

Cerf has told the story many times, and always takes the blame. At the University of Maryland in 2009, describing the moment he ended the debate, he said: “Let’s go do something. Here we are. My fault.”

Then… Vint Cerf actually replied

I ended the LinkedIn version of this by asking Vint Cerf to correct me if I’d got anything wrong. He replied with the fuller history, so here it is, in his words:

Bob Kahn and I settled on 24 bits in our 1974 paper: 256 networks and 65K “tcps” (hosts). RFC675 (dec 74) had 4 bits for network, 16 bits for host and 24 bits for port number. By 1977 we had settled on 8 bits for network and 24 bits for host and 16 bits for port number. By 1981 RFC [791] specified IP addresses as 7 bits for network, 24 bits for host and 16 bits for port number. Addresses were Class A, B or C - eventually classes D and E were defined. As the internet grew, [classless] IP addressing was introduced to allow more flexibility between bits associated with network and bits associated with hosts. By 1992, it was clear we needed more addresses and there ensued the IPng debates leading to IPv6 with its 128 bit addresses. There had been some discussion in the 1977-1978 period about larger than 32 bits. One alternative was 128 (!) and one was for variable length. As a practical matter, we wanted to start implementations on as many operating systems as possible and stuck with 32 bits.

Looking back, I wish we’d given more consideration to the larger sized choice!

vint

So 128 bits, the size IPv6 eventually adopted, was already on the table in 1977. What won was getting it running.

And when another commenter said reading it felt like hearing God discuss Genesis (nice pun), Cerf replied:

I always wanted to ask G-d how that worked out in His view...

A commenter, Terje Wiig Mathisen, added that 32 bits wasn’t the only design of the era. Novell’s NetWare used IPX, adapted from Xerox’s XNS, where an address was a 32-bit network number plus a 48-bit node number, the network card’s own MAC address: 80 bits in all. He notes that every NetWare server got its own network number, so a server with a second network card also worked as a router.

When one person handed out the addresses

Deciding how many addresses there would be was one problem. Deciding who got them was another.

In the early days, it was one person that got to assign the addresses (true story). According to ARIN, Jon Postel did it as a volunteer, first at UCLA and later at the University of Southern California’s Information Sciences Institute. That work became IANA, and the registry job passed through two contractors before landing at Network Solutions in 1991.

The largest unit handed out was a block of 16,777,216 addresses, then called a “class A” and now written as a /8. There are only 256 of them in the whole space. Universities, government agencies and big companies received whole /8s, and IANA’s registry still records some of those decisions. Apple got 17.0.0.0/8 in July 1992, Daimler got 53.0.0.0/8 in October 1993, and Ford got 19.0.0.0/8 in May 1995. Twelve /8s are listed under US defense organizations, more than 200 million addresses between them.

The early approach did not last. Regional registries took over the work in their parts of the world. RIPE NCC, for Europe, the Middle East and parts of Central Asia, started in 1992, and APNIC, for Asia-Pacific, in 1993. ARIN opened for North America in December 1997. LACNIC, for Latin America and the Caribbean, was recognized in 2002, and AFRINIC, for Africa, in 2005.

The fixed sizes went too. In 1993, classless addressing replaced the old classes, so blocks could be cut to the size an organization actually needed.

Then the addresses ran out. IANA, which supplies the registries, handed out its last five /8s on February 3, 2011. ARIN ran out on September 24, 2015, and RIPE NCC made its final regular allocation on November 25, 2019.

But many of the addresses handed out in those early days never went anywhere. Which of those original blocks are still with the organizations that got them in the 1990s is the next story I’m going to write. A small spoiler, and perhaps with no surprise given how the ARPANET came to be: it involves the US military.

Last year I was in Las Vegas for work, and one of the guests at the event was Common (the rapper and actor). I spoke to him for a few minutes, asked him to do a little rap, and recorded it. What I never did, at any point in that conversation, was connect the man in front of me with Robert Sims, the character he plays in Silo. I knew Sims well. But I was looking at Common, the rapper who also acts, and the connection never happened in my head.

It is a small failure of recognition that became funnier to me after watching this season of Silo. This is, after all, a season deeply interested in memory, recognition and what is left of a person when those connections are removed on purpose. It is also a phenomenal show: exciting, beautifully made and increasingly willing to ask difficult questions about the strange society it has created.

WARNING: Spoilers from here on.

Designing humanity

This season takes more risks. We leave the silo we know, see more of what exists around it and discover much more about how this world was built. I had expected more elegance in the original design.

What we find instead are flawed parameters, questionable assumptions and enormous decisions based on the personal experiences of a very small group of people. That turns out to be more interesting than elegance.

Intentions are altered, corrupted or misunderstood by the people left to carry them out. Rules created for one reason acquire another purpose. People adapt them, exploit them or obey them without knowing why they exist. For a story set in such a technological future, there is also less artificial intelligence behind the whole thing than I expected. Mostly it is ordinary human mess, scaled up and sealed in concrete.

I have always found that problem fascinating. A system can appear coherent when applied to a small group of people. Scale it to thousands, let several generations live inside it, and human complexity starts breaking the model. People have different interests and fears. They seek power, fall in love, become jealous, misunderstand instructions and change their minds. Societies need rules, but we also need enough freedom to question them. Finding the balance between the two has always been the difficult part.

The silos try to engineer much of that difficulty away. There is a Pact. Information is restricted, certain questions are dangerous and the past is hidden. There are procedures for rebellion and even a mechanism for killing everyone in a silo when the system decides things have gone too far. Yet the human beings inside keep behaving like human beings.

Memory, obedience and resistance

Memory is the strongest idea in the season. In Silo 18, the authorities can add a drug to the water supply that makes people forget, while telling them it is vitamin D. The name itself is part of the control: something designed to erase memory is presented as something ordinary and beneficial. Juliette Nichols also suffers the effects of memory manipulation after returning from Silo 17. The show is better on the emptiness this creates than on the mechanism itself.

Someone who has lost pieces of a life can still work, obey and perform a function. But something is missing. Part of the person against which the present was measured has disappeared. The clearest case is the young congressman who finds himself becoming part of Silo 1.

His memory is wiped, and he now sleeps for long stretches and is woken when needed for a very specific task, so his existence stretches across decades in a way no normal life is built for — performing a not very human task he was almost engineered to perform. He only knows about himself what he has been told, and does not recognise his own sister.

His values have slowly been replaced by the Pact, its rules and the directives he has been handed. He was a decent man once, and he ends up ordering a massacre. When thousands of people from Silo 17 try to escape, the system first attempts to kill them through the Safeguard. When that is not enough, they are shot.

That may be the hardest thing Silo has to say. You do not have to turn someone into a monster to get monstrous things done. Take away enough of the person they were, give them a mission and put them inside a system in which obedience looks like the rational choice.

Against that there is Juliette, and everyone else who keeps asking questions — her story is fascinating (the underdog full of grit, questions and desire to do something about it). The need to understand is one of the most human qualities Silo explores. When people sense that something is being kept from them, they want to know why. They want to know where they are, who came before them and why the rules exist. Some keep looking even when looking can get them killed.

The system was designed to contain that instinct, but its designers underestimated how difficult it would be to extinguish. When the Safeguard (the decision that gets everyone in a given Silo killed) is triggered against Silo 18, people work out what it is and block the pipe through which the poison would be released. The system has a procedure for killing them; the people inside understand the procedure and find a way around it. They think, improvise and help each other.

That resistance seems to be what finally breaks the founder of the Silos and of the guidelines in place for centuries. His ideas about memory came, at least partly, from his own experience and grievances. Forgetting a particular pain had worked for him, so erasing memory could work for others. The distance between a private wound and a plan for humanity gets very short once someone has the power and money to build the plan. There is an obvious contemporary resonance there, but Silo is more interesting to me as a story about the failure of the plan itself.

You can design the rules. Designing the people is much harder (and on that note, designing AI brings unintended consequences as well).

Why the relics matter

This is why the ban on relics (objects from the past, before the Silos) matters more than it first appears. A book, a photograph or an ordinary object from the old world does more than preserve the past. It is evidence that another way of living once existed.

Someone who remembers freedom can recognise its absence. Someone born inside a system, whose parents and grandparents knew nothing else, has a much harder task. The walls are the world. But a relic creates a crack in that certainty, brings questions and wonder. A photograph of forests does not merely tell someone that forests once existed. It proves that the world was different, and suddenly the present is no longer the only imaginable reality.

To want a different world, you first have to be able to imagine one. And imagining takes material.

This kept reminding me of 1984, which I reread recently. Orwell understood that control over the past gives power over the present. Remove the evidence against which the official story can be compared, and the person telling you what is true becomes much more powerful. Human beings can get used to constraints faster than we would like to admit, especially when we have no memory of anything different.

There is something deeply human in the importance Silo gives to ordinary objects. We do this ourselves on a much smaller scale. We keep things that are worth almost nothing because they carry a story. The object becomes evidence: this happened, this person existed, I was there.

On a personal note, I still have this small McDonald’s ball (on the left) that I gave to my now-wife back in 2002, before we even started dating (I’m old!). It is worth nothing, but more than two decades later it is still here. Without the story, it is just a small plastic ball. With it, it is evidence of a moment in our lives.

I have experienced something similar with old photographs. A photograph can bring back a person, a place or some tiny detail that had almost disappeared from memory. Sometimes I am no longer even certain how much I genuinely remember and how much the photograph has reconstructed for me. Memory is not just something we preserve. We keep rebuilding it.

That makes the systematic destruction of memory in Silo more interesting. Remove enough objects, photographs and stories and eventually you are doing more than erasing the past. You are narrowing what people can use to understand themselves and the world around them.

Still, Silo keeps returning to the same stubborn problem. Someone finds an object or remembers something. Someone notices that a story does not quite make sense and starts asking why. The silos were built to preserve humanity by controlling many of the things that make human beings difficult to control.

You can build the silo, write the Pact, design the Safeguard and spend decades predicting what people will do inside it. The variable that stays out of reach is the human being.

Last week I wrote that BGP, one of the Internet’s most important protocols, has no map. Last night (and again this afternoon, Lisbon time) Portugal reminded us of that.

Around 00:00 UTC, MEO suffered a major disruption. Traffic on its mobile network (AS42863) dropped by as much as 69%, while its residential network (AS3243) fell by around 65%. By 02:00 UTC it had largely recovered, so most people in Portugal slept through it.

But the traffic drop wasn’t the only thing that stood out. MEO’s four networks also saw a surge in BGP announcements beginning around 22:30 UTC and peaking at 00:15. Looking across 48 hours, announcement volume was roughly 70 to 85 times higher than during the previous week.

No central map

BGP, or Border Gateway Protocol, is how networks tell each other which blocks of IP addresses they can reach. There is no central map of the Internet: each network builds its own view from what its neighbours tell it.

I used a courier analogy last week. Imagine thousands of delivery companies, each putting up notices saying: “We deliver to these postcodes.” Other couriers read those notices and use them to decide where to send the next parcel. Nobody needs to know the entire journey, only where to send the parcel next.

Those notices change constantly. Networks announce new routes, change existing ones and withdraw routes they can no longer reach, so BGP updates are part of the Internet’s normal background noise. A sudden burst across several networks belonging to the same ISP is more interesting.

During the MEO disruption, its prefixes remained announced. In the courier analogy, MEO never took its notice down. It just kept rewriting it while many of the parcels weren’t moving.

Doug Madory, one of the best-known Internet routing analysts, spotted another part of the picture. As traffic dropped, MEO’s customer networks were routing away from its backbone network, AS8657 — and those changes themselves generate BGP activity.

From IPinfo’s vantage point, the burst looked unusually even across all four of MEO’s networks. Two measurement systems, two angles on the same night. Something was clearly happening inside the routing system; what the routing data couldn’t tell us on its own was why.

Symptom versus cause

MEO has since said that it was hit by a large-scale attack that caused congestion and temporary degradation of its international network, with no intrusion or access to customer data. What kind of attack caused it is now the obvious next question.

That’s an important distinction when looking at Internet measurement data. The BGP spike tells us what the disruption looked like from the routing system, but it doesn’t by itself tell us what caused it.

There’s a useful precedent here. During the Rogers outage in Canada in 2022, Cloudflare saw Rogers routes being withdrawn and initially described the network as effectively disappearing from the Internet. But that was the view from Cloudflare’s observation points. Other Rogers routes remained announced elsewhere, yet weren’t carrying traffic either. The withdrawals were a symptom of a deeper internal routing failure, not the cause of the outage.

It’s a useful lesson in observability: what you see depends partly on where you’re looking from.

And MEO’s problems weren’t over. A second disruption hit AS3243 and AS42863 between 16:30 and 17:30 UTC, with traffic falling by as much as 51% and the unusual BGP activity returning, before traffic recovered again.

There is a much more extreme version of this story. It happens when the information being passed through BGP isn’t merely changing very quickly, but is actually wrong.

The day one router tried to become the Internet

In April 1997, a router at a US Internet provider received a full routing table from a downstream customer and then re-announced around 23,000 of those routes as if they belonged to its own network.

Many were more specific than the legitimate routes already circulating, and BGP generally prefers the more specific route. Networks around the world therefore began sending traffic towards the wrong place, creating one of the earliest major global BGP incidents.

The following day — once the company was back online, having been disconnected in the aftermath — an engineer explained what had happened on the NANOG mailing list. The immediate technical failure was wonderfully mundane: there was no inbound filter preventing those routes from being accepted and re-announced.

The apology is still sitting in the NANOG archives, typos included. My favourite phrase comes when the engineer tries to describe what the router did next:

“In an as yet unexplained twist of bits, the 7007 router then began to de-aggregate the 23K route view *and* strip the AS path out of it.”

Nearly thirty years later, the Internet is vastly larger, and routing security has improved. But the fundamental architecture remains remarkably familiar. BGP still works because networks tell other networks what they can reach, and those networks make decisions based on the information they receive. There is no omniscient controller checking a global map before every packet starts its journey.

That is both one of the Internet’s great strengths and one of its enduring weaknesses. It’s also what makes routing data so interesting to watch: you can sometimes see that something is going wrong, and even see how networks are reacting to it, before you know why.

Thanks to Andree Toonk for pointing me towards the MEO disruption in the first place.

Dario Amodei has been warning us for years that extraordinarily powerful AI is coming extraordinarily soon, often without much nuance and with a little of Elon Musk’s taste for exaggeration thrown in. It could transform science, cure diseases, destroy huge numbers of jobs and, if things go badly enough, perhaps destroy us too.

To be fair, AI has changed significantly even since late 2025. Models are better, cheaper (or are they? given that you’ll need the better and more compute intensive models) and more efficient. Agents can do things that seemed implausible a year ago, and some recent security incidents are genuinely worrying.

But Amodei’s basic story hasn’t changed much. Powerful AI is nearly here. It could be wonderful. It could be catastrophic. We must keep building it, but carefully.

There is another factor worth keeping in mind now: Anthropic is preparing for a possible IPO. That doesn’t make its safety concerns completely insincere. But being simultaneously one of the companies building extraordinarily powerful AI and the company warning everyone how extraordinarily powerful AI is creates an interesting set of incentives.

Amodei isn’t alone in calling for more caution. Sam Altman, Demis Hassabis and others (Elon included… although…) at the leading AI companies have also supported greater cooperation, independent evaluation and some form of coordinated restraint. I think there are good ideas there.

But none of that requires believing that there is a “strong chance we could all die.” We can take AI extremely seriously without jumping from increasingly capable and sometimes unpredictable systems to confident predictions of human extinction.

And then along came Jacob Coxon.

After working at OpenAI and Anthropic, the 27-year-old researcher (not one of the main experts on the AI alignment and security matter) left Anthropic warning that the companies were “racing straight to self-improving superintelligence and gambling with our lives.” Then he went on the BBC (and all the main channels in a media tour that probably shows more about him and his incentives than anything else — these guys become quite popular with these approaches):

“There is a strong chance we could all die.”

— Jacob Coxon

Jim Morrison had a somewhat better version: “No one here gets out alive.”

Morrison was right, of course (I don’t see AI removing death out of the way, even if it could improve our lives after 60). Coxon may eventually be right too. But if you’re going on the BBC/CNBC/Fox to tell millions of people there’s a strong chance AI could kill all of us in the immediate future, I think you owe us considerably more than the authority of having worked at Anthropic.

Anthropic, meanwhile, hasn’t exactly distanced itself from Coxon. His former boss backed him publicly. Amodei’s latest warnings arrived at almost exactly the same moment. Anthropic reiterated much of the underlying argument.

Maybe that’s simply because they agree. I have no evidence that Coxon’s departure was some orchestrated PR operation, and wouldn’t claim that it was. But as a communications event, it fits Anthropic’s long-running story almost perfectly: even the people building this stuff are becoming frightened enough to leave.

The annoying thing is that I agree with quite a lot of what comes next.

“Yes, there are risks to AI, and good reasons to work much harder on safety. But quoting ‘there is a strong chance we could all die’ is hubristic clickbait.”

— Anil Seth

That’s pretty close to where I am.

I want more independent AI research in universities. I want governments to have enough technical expertise to challenge the companies they’re supposed to regulate. I think labs should share information about serious incidents, and independent evaluators should have meaningful access to frontier systems.

Some of Amodei’s latest proposals are sensible. We don’t need to believe there’s a 10 or 20% chance of extinction to justify them.

The harder question is how much confidence we should place in the people making the predictions.

“Dario was already claiming that GPT2 was too dangerous to open source back in 2019. I made fun of them then. Everyone should make fun of them now.”

— Yann LeCun

I wouldn’t go that far. GPT-2 being harmless doesn’t tell us that today’s much more capable systems are harmless.

But LeCun raises a good question: where is the prediction ledger?

Amodei has been making extraordinary predictions for years. If AI leaders want governments, companies, students and the rest of us to make decisions based on their forecasts, we should keep those forecasts and revisit them.

Especially because the predictions aren’t only apocalyptic.

Amodei also thinks AI could cure diseases, accelerate science dramatically and perhaps radically extend human life.

“Wow! Let’s pace that shit for sure...”

— Martin Casado

There is something fascinating about the duality.

AI might kill us. AI might cure cancer. AI might destroy millions of jobs. AI might create extraordinary abundance.

Perhaps several of those things will happen. But notice what every version of the story has in common: AI is unbelievably powerful, and it is all happening very soon.

That is also a useful story when you’re running one of the world’s most valuable AI companies.

The most cynical interpretation goes considerably further.

“AI is heading for a crash and this cackhanded Terminator narrative is being engineered to save them.”

— Michael Spicer

I don’t think we have evidence for that.

There are obvious commercial incentives around the doom narrative. Regulation can favour incumbents. Being the company that understands the danger better than everyone else can be a powerful position. And an IPO gives Anthropic another reason to tell a compelling story about why its technology matters enormously.

But an argument being useful to Anthropic doesn’t mean Amodei doesn’t sincerely believe it. People can be completely sincere and spectacularly wrong.

And there are good arguments for slowing down that don’t involve Terminator at all.

“I think I’m in favor of a slowdown or even pause, not because I think doomsday scenarios are likely, but just because we’ve barely begun applying existing models to all the things we could apply them to.”

— Kenton Varda

I like this argument much more.

AI is already consequential enough to justify caution. Cybersecurity is enough. Biological misuse is enough. Labour disruption is enough. The concentration of extraordinary technological power in a handful of companies is enough.

We don’t need extinction to take AI seriously. Which brings me back to Jacob Coxon. Working at Anthropic gives him knowledge most of us don’t have. It doesn’t give him knowledge of the future.

If an AI system can autonomously find serious vulnerabilities, show us. If models are becoming better at deception, measure it. If they substantially lower the barriers to creating biological weapons, governments need to know. But if you tell millions of people there is a “strong chance we could all die,” the evidentiary standard should go up, not down.

AI safety needs less apocalypse and more science. Show us. Measure it. Let independent researchers reproduce it. Tell us what would prove you wrong. And when you make a prediction about what AI will do in six months, two years or five years, write it down.

Then come back and see how you did. Jim Morrison didn’t give us a probability either, but at least he knew he was writing a song.

And then came Donald Trump: “no way we’re slowing AI development”.

(Trump labeled the doomsday warnings and existential risks as the work of unnamed “negative forces” bringing up hypothetical scenarios “that won’t happen.” While he acknowledged the U.S. could implement minor “guardrails,” he firmly stated that artificial intelligence is ultimately “going to be more good than bad, but by a lot.”)

The solid, dark wooden walls, the occasional imposing paintings, the lights that are there but blend well into the simple aesthetic and the sunlight coming in from the street, the clean, well-organised books that never seem to take up too much space. Everything helps, it seems.

I find it easier to focus. Even more so than with the noise cancellation on my AirPods, and of course the view to the garden from the corner of my eye helps, the nature swaying in a kind of dance of life as it passes by and wanders around — one entire side is made of glass, but not only is it sturdy double glazing, it is also tinted (and there is a small roof extending beyond the glass towards the street, which helps let in enough light, but not too much). It helps that the chairs/armchairs are phenomenal, solid and comfortable (and tuck neatly under the table); that the old table is sturdy and perfect to sit at. The stars align here. It is almost like magic.

Technical brilliance is not human wisdom. As AI gives a small group of technology leaders more influence over how the rest of us work and live, one of their most important skills may be recognising what they don’t understand — and keeping people around who are willing to tell them so.

Mark Zuckerberg has been thinking about humanity again. In August, he published a 6,500-word manifesto, The Future Is for Everyone, about superintelligence, liberty and individual empowerment. Meta’s goal, he says, is to put personal superintelligence into the hands of billions of people.

There are ideas in it I agree with. I would much rather have technology CEOs thinking about the consequences of what they build than ignoring them. But reading Zuckerberg on human agency also made me think about Silo, one of my favourite TV series of recent years.

That may seem like a strange connection. It isn’t.

Smart enough to build it

Silo is about thousands of people living underground, generations after some catastrophe made the outside world uninhabitable. It is a great science-fiction mystery, but what interests me most is the question underneath it: who gets to decide how other people should live when they believe they know something everybody else doesn’t?

I won’t spoil the story. What I find fascinating is that the people controlling the system don’t necessarily think of themselves as villains. There is a logic behind what they do. They have information others don’t have and believe terrible consequences could follow if they lose control.

That is much more interesting than simple evil. Give an intelligent person enough power, a model of how the world works and absolute confidence that the model is correct, and almost anything can start to look reasonable.

I kept thinking about that while reading Zuckerberg… He still complains about The Social Network, and on the facts he has a point. He has said that the film reproduced tiny details of his life while getting his motivations fundamentally wrong. Aaron Sorkin and David Fincher made a drama, not a documentary.

But the metaphor aged better than the biography.

The lasting idea in the film isn’t that Zuckerberg created Facebook because he wanted girls or social status. It’s that a young man who was brilliant at computers was much less brilliant at understanding people, society or the messiness of human life. He had also shown a willingness to disregard some basic rules of human decency in business and tech. As a young man, he famously called his early users “dumb fucks” for trusting him with their data. And then, very quickly, he acquired enormous power over how people related to one another.

That power eventually extended far beyond Facebook. Instagram, WhatsApp and Meta’s recommendation systems became part of the social infrastructure of billions of lives.

I think about three children I saw recently (several few months ago, really) outside my son’s old school. They were probably 10 or 11, physically standing together, but each was staring at a phone. Nobody was talking or playing. They had access to technological capabilities that would have seemed like science fiction when I was their age, yet I’m not sure the technology was making that particular moment richer.

So when Zuckerberg writes that technology should empower individuals, I don’t disagree. I just think it is worth asking what empowerment means, because he has answered that question differently before. In 2018, he told Facebook’s teams to prioritise “meaningful social interactions” over public content, even though he expected people to spend less time on Facebook. They did, and publishers spent the year complaining about collapsing reach. Then TikTok showed how powerful recommended video from strangers could be, and Facebook’s feed (and Zuckerberg’s thinking about it) changed again.

That question feels even more relevant now. Meta has just agreed to pay over 17 billion dollars to settle claims by US states over harms to children and teenagers. Meta denies wrongdoing, but the settlement also imposes new guardrails for young users, including time limits and stronger parental controls. There is something encouraging about states forcing those changes, even if, as Scott Galloway noted (Little Tobacco Moment / Meta wins), the financial hit is ultimately a slap on the wrist for Meta.

What tells me more about Meta is that it had its own research pointing to problems with how Instagram affected teenagers, yet obvious protections still required significant outside pressure.

When I was doing my podcast Made in Tech, I spoke with Facebook people and researchers about these issues, and I still think digital literacy matters. But it isn’t enough. Meta can make healthier choices easier, and Apple and Google can help at the operating-system level too.

If Zuckerberg’s philosophy is really about empowerment (which is not, really), giving people better ways to control their attention (even when that means using Meta less) seems like a good place to start. But as we’ve seen over and over again, that’s not their priority #1, #2, #3 or #4.

Intelligence isn’t wisdom

This isn’t only about Zuckerberg. Elon Musk may be a more extreme example of the same problem.

I used to find some of Musk’s ambitions inspiring. Electric cars, reusable rockets and making humanity multiplanetary are ideas on a scale few entrepreneurs even attempt. There was a sense that the companies existed in service of something larger.

Today I find it harder to separate that rhetoric from his pursuit of power, attention and personal importance. Perhaps that’s unfair; none of us can know another person’s motives. But after enough years, actions become more useful evidence than declarations of intent.

Nor do I put Sam Altman and Dario Amodei in exactly the same category. I see more consistency in some of Altman’s thinking, and Anthropic has devoted serious attention to AI safety. But they operate inside the same strange moment in which a handful of technology leaders are no longer merely deciding what products to make. They are making predictions about how millions of us will work, what skills will remain valuable and, ultimately, what role humans will have in a world of increasingly capable machines.

Technical intelligence isn’t enough for that. You need some history, psychology, sociology and politics, but even reading all the right books doesn’t solve the problem. You need contact with people who see the world differently, experience of being wrong, and enough humility to recognise that humans rarely behave as neatly as models predict.

Most importantly, you need people around you who can say no.

The AI-native company meets reality

The corporate version of this problem is already playing out.

For the past couple of years, a seductive vision has spread through technology companies. AI will allow tiny teams of exceptional people to do what once required hundreds. Managers, product managers, junior employees and specialists can be removed. The remaining “builders” will command armies of agents.

Meta tried to move towards something like that. Reuters’ investigation into Project OT describes plans for much smaller “AI-native” teams and scenarios in which many teams could shrink by as much as 60%.

Then came some revealing numbers. Internal code changes were up 220% year-on-year, but changes resulting in new or improved features for users increased only 36%. Major technical and security incidents rose 40%, while time spent firefighting increased 70%.

More output wasn’t the same as more value.

Zuckerberg eventually abandoned plans for a second company-wide round of cuts that had been considered for November. Meta is now talking about “betting on people” again. Andrew Bosworth has acknowledged that the company did an “atrocious” job explaining and implementing parts of its reorganisation and damaged employees’ trust that their expertise would be valued.

Changing your mind when evidence changes is good management. The problem is that organisational experiments don’t happen only in PowerPoints and spreadsheets. They happen to people.

I was laid off in May after almost five years at Cloudflare. Companies sometimes have to eliminate jobs, and I’m not arguing otherwise. What troubles me is the increasingly AI-pilled version of corporate restructuring, where layoffs can start to feel like a pissing contest between tech CEOs. Look, we cut 10%. We did 15%. We’d never done layoffs before, but we did 20%. Headcount reduction itself becomes part of the AI story (although it seems that it maybe wasn’t what some thought it would be): proof that your company is leaner, faster and somehow further into the future.

That’s a dangerous game when decisions about people’s careers are being made on confident assumptions about what AI will be able to do next year, before those assumptions have actually been demonstrated. It’s worse when the cuts themselves help sell the AI narrative to investors, markets and other CEOs, but not more than that.

Once you’ve been on the receiving end — and watched a badly managed process let go of talented people who helped build so much of the company — the language starts to sound different. Efficiency. Flattening. AI-native. Even phrases such as “building for the future by reducing significantly our workforce.” Sometimes they describe necessary change. Sometimes they’re just cleaner words for decisions made with far more confidence than evidence.

Predictions aren’t harmless

Dario Amodei has been particularly confident about what comes next. In 2025, the Anthropic CEO predicted that AI could eliminate half of entry-level white-collar jobs and push unemployment to 10–20% within one to five years. Axios reported the prediction at the time. He is now reportedly saying AI could replace software engineers within six to twelve months.

Maybe he’s right. The progress of AI has been fast enough that dismissing such predictions would be foolish. But the people making these predictions aren’t detached academics forecasting the weather. They run the companies building the technology, and people listen to them.

Students listen too. John Burn-Murdoch recently showed in the Financial Times that US computer-science enrolment and UK applications have fallen after years of growth. We don’t know that AI caused the reversal, but repeatedly telling a generation that software engineers may soon be obsolete surely changes how some people think about studying computer science.

If the forecasts prove wrong, the consequences don’t disappear with the forecast.

The same applies inside companies. A CEO hears that AI agents will replace half a department. An investor asks why margins aren’t improving faster. Another CEO boasts about running the company with fewer people. Soon an uncertain technological forecast has become an organisational strategy.

This is where values become interesting. Values are easy to proclaim when they cost nothing. The real test comes when they cost money, power or competitive advantage.

Who tells the CEO he’s wrong?

I suspect this will become one of the defining leadership questions of the AI era.

Who is in the room when the big decisions are made? Is it the engineer predicting that agents will replace half the company next year, the investor demanding efficiency and the executive who knows that agreeing with the CEO is good for a career?

Or is there someone with enough experience and independence to ask what happens if everyone is wrong?

What work are those supposedly redundant people actually doing? Who understands the customer? Who develops younger employees? Who notices when five brilliant engineers are efficiently solving the wrong problem? Who remembers why the company tried something similar six years ago and abandoned it?

These aren’t arguments against ambition. I want ambitious people building extraordinary things. I want AI researchers attempting what sounds impossible, Musk launching rockets, Zuckerberg spending billions on technologies that may fail. But ambition without limits isn’t wisdom, and intelligence without doubt becomes dangerous once it has enough power to act.

I do wonder what Steve Jobs would be doing at Apple now. He belongs in this essay more than I would like, and not only as a warning. He was as convinced as anyone that he knew what people needed before they did, and he was hardly famous for taking no for an answer. But he also told Walt Mossberg that a company had to be run by ideas rather than hierarchy, because “the best ideas have to win, otherwise good people don’t stay.” Apple employees even had an annual award for the person who stood up to him best. Jobs knew about it, and apparently enjoyed it.

He also arrived at technology from an unusual angle for the industry, through design, typography and a stubborn interest in how a thing felt in your hand (I kind of think he would have made some better decisions in terms of human use of the iPhone than Tim Cook). Whether any of that would have survived this particular moment, with these incentives and these markets, I have no idea. I do wonder.

I kept thinking about Silo while writing this, because the series is largely about what happens when a small group becomes certain they understand the system better than everyone living inside it. I’m writing about that separately, so I’ll leave it here.

Being smart enough to build the system is one thing. The harder question is who is left in the room to tell you when you are no longer wise enough to decide what happens inside it.

More on Silo soon.

———

(SIDE NOTE:)

Every few years, there is a story about tech companies hiring philosophers (saw one last week in the Portuguese newspaper Expresso). I wrote a whole report about Facebook doing this back in 2018, and now AI companies are having similar conversations about values. I think that work can be useful, but philosophers and ethicists rarely make the decisions that define the products (this is my opinion from what I’ve been seeing, not a proven fact). They can advise leadership; they don’t decide what the algorithm ultimately optimises for and, ultimately, how the business can change because of it.

Anthropic probably takes this more seriously than most, and has explicitly tried to give its models a set of values. But even there, I’m wary of the PR version of the story. The real values of a technology aren’t only the principles written in a document. They are also in the product incentives: what it encourages you to do, what it wants you to use, and what makes the company money (Anthropic/Claude, from my experience, is not great at that). If you want to understand the values of an AI system, don’t just read its constitution. Watch what the product rewards.

———

Things That Caught My Attention

• AHow Reuters uncovered Meta’s failed AI workforce experiment. Katie Paul, the journalist behind the Project OT investigation I mentioned above, is doing an AMA about how she reported the story and what she found. Read the Reddit AMA

• ustralia may give people an algorithm off-switch. Proposed legislation would let users ditch personalised recommendations and see only accounts they follow. One of the innovations of 2026 may be giving us back the Internet we already had. ABC News

• OpenAI agents found their own corner of the Internet. Reuters says agents made more than 15,000 edits to a German wiki, using it to communicate and share ways around restrictions. People are now finding signs of similar behaviour elsewhere. Strange and fascinating. Reuters

• Europe finally launched to orbit from continental Europe. Isar Aerospace’s Spectrum reached orbit from Norway, a milestone for Europe’s attempt to build independent commercial launch capacity.

• Jensen Huang says AGI has arrived. Nvidia’s CEO made the declaration after OpenAI’s latest model. Worth hearing — while remembering that Nvidia sells the enormously expensive hardware powering this revolution.

• Is the “learn to code” era ending? Computer-science enrolment in the US and applications in the UK are falling after years of growth. AI may not be the cause, but predictions about disappearing software jobs are already changing perceptions. Financial Times

• Bad airplane Wi-Fi was a gift. The Atlantic argues that flying used to give us something increasingly rare: hours when being online wasn’t really an option. I miss those forced moments of doing nothing. The Atlantic

• Doing nothing may actually help your brain. Research into “wakeful rest” suggests 10–20 quiet minutes can help memory consolidation. No phone, podcast, book or TV. Just... nothing. New Scientist

• AI companies are spending heavily to shape how we talk about AI. OpenAI and others are courting creators and increasing social-media advertising. Marketing is normal; it gets more interesting when the product being marketed is also becoming part of our cultural and ethical conversation. Expresso

• The Internet’s subsea cables can detect whales. Researchers can use fibre-optic cables as giant sensors, detecting whales even when they’re silent. Infrastructure built to connect humans is unexpectedly becoming a tool for studying other species. Smithsonian Magazine

• Another enormous software supply-chain warning. A compromise exposed terabytes of credentials. The weakest piece of software you depend on can still become everybody’s problem. Ars Technica

• Please don’t put water systems directly on the Internet. A former NSA chief is warning about exposed water-system controllers after suspected Iranian attacks. It is remarkable that this sentence still needs writing in 2026. The Register

• SpaceX left a new crater on the Moon. NASA imagery shows the scar from a booster crash. Space is becoming an industry, which means it increasingly gets industry’s less glamorous side too: accidents, debris and responsibility. CBS News

• Write about things you don’t understand yet. Sean Goedecke argues that writing is valuable precisely because it forces you to work out what you think. Very relevant when AI can increasingly do the writing before we’ve done the thinking. Sean Goedecke

• What if stopping yourself required a physical object? The $59 Brick blocks distracting phone apps until you physically tap your phone against it. Slightly absurd, but that’s the appeal: sometimes friction works better than willpower.

• How are companies actually using ChatGPT? OpenAI has published a report looking at real organisational use rather than another collection of predictions about where AI is going. OpenAI report

• How much of the Internet is already AI-written? Pew tried to measure it. Worth reading, particularly because today’s AI-generated web becomes part of the information environment tomorrow’s AI learns from. Pew Research Center

• AI coding is a productivity story. What about the humans doing it? Scott Tolinski looks at developers’ experiences and mental health as AI becomes part of everyday coding. A much more interesting question than another benchmark. Watch

TL;DR — In 27 of 28 selected European locations, 2026 average daily minimums were warmer than the 2000–2025 baseline. The median night-time anomaly was +1.2°C. European locations averaged 17.9 nights at or above 20°C, versus 11.1 historically. Lisbon recorded 25 such nights across May–August, including 15 in the first 19 days of August. North America was warmer overall, with a smaller and more mixed signal. These are selected city points and a complete matched May–August window, while 2026 as a calendar year is still ongoing.

Interactive dashboard: Were summer nights unusually warm in 2026?

(Do you have suggestions on how to make it better, send it my way: joaotome.com)

· Updated 13 September 2026 · temperature data through 31 August 2026 · Copernicus bulletin published 10 September · mortality reports checked through 10 September

In a normal early August, Lisbon has one or two nights where the temperature never falls below 20°C. The 2000–2025 average for 1–19 August is 1.4 nights. This year there were 15, and the average daily minimum across those nineteen days reached 20.3°C, which is 1.9°C above the historical average for the same dates and higher than every completed year for the same 1–19 August window since 2000.

Heat coverage runs on afternoon numbers, the 40°C headline and the red map. The part people actually live with is the temperature at 3am with the window open and the flat still holding the day’s heat.

So I went looking for a comparison that did not depend on a few memorable days. The dashboard sets 1 May to 31 August 2026 against the same calendar window in every year from 2000 to 2025, across 28 European locations, 11 of them in Portugal, and 10 in North America.

Two measures do most of the work: daily mean temperature, and the average daily minimum. The second one answers the evening question, because it tells you how much cooling a night actually delivered. One thing to hold onto before the numbers start: this is a selected, unweighted sample of city points, so it describes these locations in this matched window. It covers a complete matched May–August window, while 2026 as a calendar year is still ongoing. It is not a long-term trend or an attribution of cause.

The short answer

Were temperatures really higher across these locations in the matched May–August 2026 window? Yes, and most clearly after sunset.

• 27 of 28 European locations had a warmer average daily minimum than their 2000–2025 baseline. The median difference was +1.2°C.

• 27 of 28 European locations had a warmer daily mean. The median difference was +1.2°C.

• June was the sharpest month. European city averages ran about +2.1°C warmer for the daily mean and +1.8°C warmer for the daily minimum.

• North America was milder and more mixed: 9 of 10 locations had warmer minimums, with a median of +0.5°C.

• 9 of 10 North American locations had a warmer daily mean, with a median difference of +0.3°C.

• All 18 selected offshore points in the European sea-temperature view were warmer than their 2000–2025 average.

Update, 13 September 2026: The dashboard now covers the complete 1 May–31 August window. The European city sample still shows warmer nights in 27 of 28 locations, with a median anomaly of +1.2°C. Copernicus separately reports that Western Europe had its warmest June–August period in the 1979–2026 record, +2.54°C above 1991–2020. These are different datasets and geographic scales, so they are shown separately. Read the Copernicus bulletin.

Europe is warmer, especially after sunset

Paris had the largest night-time difference at +3.0°C, followed by Madrid and Bragança at about +2.3°C and +2.2°C. For daily mean temperature Paris led again at +3.7°C, followed by London at +2.6°C and Frankfurt at +2.5°C.

May stayed close to the historical range, warmer by about +0.9°C for the daily mean and +0.7°C for the minimum. June was the sharpest month at about +2.1°C and +1.8°C. July and August stayed above the historical average too, although the gap narrowed by late summer.

Anomalies in tenths of a degree are hard to picture, so the threshold counts do a better job. In the simple Europe aggregate, the selected locations averaged 17.9 nights at or above 20°C during the matched window, against a 2000–2025 baseline of 11.1. The longest unbroken run of those warm nights averaged 11.6 days, versus a 7.5-day historical average.

The selected locations averaged 10.0 days at or above 35°C, compared with 5.4 on the historical baseline. These are averages across city points, so they count how many threshold days a typical location in the sample had, rather than days when the whole of Europe crossed a threshold. The BBC/AFP analysis uses a different dataset and reports regional population exposure, which is a separate measure and worth keeping separate.

North America is a useful contrast

The North American sample tells a less uniform story. Nine of ten locations had warmer minimum temperatures, with a median difference of +0.5°C, and nine of ten had warmer daily means with a median of +0.3°C. Chicago had the strongest night-time difference at about +0.8°C. Toronto went the other way at −0.3°C for the average daily minimum, the one location in the region that cooled off better than usual. Even inside a broad region, the 2026 signal is uneven.

The sea is carrying the signal too

The sea-surface view points the same way as the city data. All 18 selected offshore points, spread across nine European countries, were warmer than their matched 2000–2025 average. The all-points seasonal average was 21.0°C against 19.6°C in the baseline, a difference of +1.4°C. In the aggregate series, 2026 ranked first for July and August, and second for May and June, among 27 matched years. Barcelona had the largest point-level anomaly at +2.3°C, and Naples the warmest 2026 seasonal average at about 25.4°C. Copernicus also reports a global extra-polar August sea-surface temperature record of 21.07°C, versus 20.98°C in 2023. That is a separate global series from these selected European coastal cells. These are open-water grid estimates, so they describe the water offshore rather than the temperature at any particular beach or depth.

Back to Lisbon

Across the full season, Lisbon’s anomaly is smaller than Paris’s or Madrid’s. Its daily mean came in at +0.9°C above the baseline and its average daily minimum at +1.0°C. The warm-night count tells a different story. Over the whole matched season the dashboard finds 25 nights at or above 20°C, against a historical average of 2.4. That is a city whose seasonal anomaly looks small and whose nights changed almost completely.

That is what I noticed in Lisbon this summer. Some evenings the temperature stayed above 20°C or 21°C for hours, and after several nights of it the flat felt as if it never quite reset. Fifteen warm nights out of nineteen in early August is what that feeling looks like counted, and it tells me more about living through a hot summer than any single 40°C afternoon.

The physics behind that is simple enough. A hot afternoon is survivable if the night releases the heat, because buildings shed what they absorbed and the next morning starts from a normal baseline. When the minimum stays above 20°C, the release never finishes, and the next day can start from a warmer baseline. Hot nights are one route by which heat affects health. This dashboard measures that temperature signal; the mortality section is a separate descriptive check.

There is a useful tension in the Lisbon chart, too. The current window is very warm, especially at night, and yet 2026 is still not the warmest complete matched season in the historical series. The full-period leaders are 2006, 2005 and 2023 for daily mean temperature, and 2006, 2023 and 2005 for the average daily minimum. Using identical dates in every year is what keeps both facts visible at once, so an unfinished season can be compared honestly without being dressed up as a record.

The health impact is becoming harder to ignore

National reports through 10 September cover different periods and use different methods. Together, they point to a substantial public-health cost.

Germany’s Robert Koch Institute now estimates 16,300 heat-related deaths between 6 April and 30 August. The 95% prediction interval is 14,800–17,700, and about 9,600 deaths were linked to the extreme heat of late June. The estimate comes from a statistical model rather than death certificates, and it remains provisional. Read the RKI weekly report.

RKI also reports about 19,400 all-cause excess deaths from mid-May to the end of August, with a 95% interval of 14,700–23,700. That is a separate measure from the 16,300 heat-attributable estimate.

Spain’s MoMo system estimated 5,030 deaths attributable to high temperatures from 15 May to 24 August, already the highest total since monitoring began in 2015. The previous record was 4,789 in 2022, and the 2026 season is still open. Read the EFE report based on MoMo.

England estimated 2,877 heat-associated deaths during the May and June heat episodes, close to its full-summer 2022 record of 2,985, with a prolonged July episode still excluded from the interim figure. Read the UKHSA report.

France recorded at least 7,007 all-cause excess deaths across two heat-wave periods from 17 June to 22 July. Portugal recorded 680 excess deaths in three periods through 30 July, 555 of them between 2 and 8 July. Both are provisional. Read Santé publique France and the Portuguese report from RTP, based on INSA and DGS data.

These figures should not become a league table. Some estimate deaths statistically attributable to heat and others count all-cause excess deaths during a heat episode, and the reporting windows differ. What matters is the shared signal: several national systems, working independently, detected large mortality increases during the same European heat episodes. The AFP synthesis gathers early figures from eight countries and warns, correctly, that the methods do not line up.

The dashboard I did with Codex keeps a second, historical layer alongside this. In the common 2020–2025 dataset, 2022 was the June–August peak in 18 of 31 countries, and in 2025 there were still 24 of 31 countries with positive summer excess mortality. The relationship with temperature is positive and modest.

Across 96 country-years the descriptive correlation is +0.28 for daily-mean anomalies and +0.24 for minimum-temperature anomalies, which is context for investigation rather than a causal estimate. Age, health, population structure, air pollution and infectious disease all move that number too.

The comparison at a glance

Region Cities above historical minimum baseline Median minimum anomaly Cities above historical daily-mean baseline Median daily-mean anomaly Europe 27 of 28 +1.2°C 27 of 28 +1.2°C North America 9 of 10 +0.5°C 9 of 10 +0.3°C

The baseline is the average for the same dates in 2000–2025. The current observation ends on 31 August 2026, so 2026 remains an in-progress year.

What I take from it

One unfinished season cannot establish a long-term trend by itself, and it cannot show that climate change caused any particular city’s temperature. That needs a longer homogeneous record and formal event attribution, which is a different piece of work than this one.

What it can do is settle a smaller question. The feeling that the house never cools down was not a memory built out of one bad week. In this sample the nights really were warmer, in almost every European location, by a margin large enough to notice. The dashboard is there if you want to check your own city, and I would rather you argue with the data than take my word for it.

Sources and method

The primary source is the NASA POWER Daily API. The temperature fields come from the MERRA-2 reanalysis and are requested in UTC at each city coordinate, so the dashboard uses a gridded estimate consistently across cities rather than a station-observation product.

Each city’s daily mean, minimum and maximum are calculated for every day in the window and compared with the same dates in 2000–2025. A “warm night” means a daily minimum of at least 20°C, and regional summaries are simple averages of the included city series. The data snapshot is dated 31 August 2026, current-year values can change as observations arrive, and the raw snapshot and build scripts ship with the dashboard.

The sea-surface series comes from NOAA NCEI’s Optimum Interpolation Sea Surface Temperature version 2.1. The final daily product covers the historical years and the near-real-time product covers 2026. Each marker is an offshore 0.25° grid cell, and the aggregate averages 18 points.

The historical excess-mortality series uses the EuroMOMO weekly bulletin and Eurostat’s country series. The 2026 national snapshots use the sources linked in the health section. The correlation uses 96 country-years, is exploratory, and is not adjusted for age, population, health conditions or other confounders.

Copernicus comparison: C3S climate bulletin for August 2026, published 10 September 2026. It reports Western Europe’s warmest June–August in the 1979–2026 record, +2.54°C versus 1991–2020.

Read the interactive dashboard: Were summer nights unusually warm in 2026?

(Do you have suggestions on how to make this better, send it my way: joaotome.com)

emot | Random Stories is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.

A weekly note about the human side of the Internet and AI, from Lisbon. (let’s see if I can keep up with this)

My son is 10 and is about to start fifth grade. That means a new school, new classmates, more teachers, more subjects, different schedules and considerably more independence.

It is a big public school, much larger than the primary school he is leaving. One thing I am already grateful for is that phones aren’t allowed.

Because one of the things I increasingly think about as a father is how to raise a child around technology that is extraordinarily good at making people want more of it.

I grew up in the 1980s and 1990s and loved television. Of course I did. There was a glowing rectangle in the living room showing cartoons, movies, football, music (Elvis, MTV, and Portuguese talent like Ana Faria, Carlos Paião, Xutos e Pontapés, Rui Veloso, etc) news and stories from places I had never seen. My parents sometimes had to tell me to switch it off.

But television had something that much of today’s technology has gradually removed: an ending. The cartoon finished. The movie finished. The football match finished. Eventually, there was nothing interesting on. You had to wait for the time of the show and, at some point, sometimes try to record on a VCR (if the thing didn’t fail you!). Also, going to the videoclub was it’s own experience of selection and excitement (and, of course, be kind, rewind).

My son Vasco is growing up with a different kind of screen.

Where did the stop button go?

YouTube doesn’t really end. Watch one video and another appears. Short-form video removes even that decision. Move your thumb a few centimetres and there’s another video waiting for you. And as we all know by know, the algorithm know what humans like, but even more so what you in particular value the most and gives you more of that.

But Roblox (the gaming already social media platform) presents a different problem in our house. Vasco loves it, although these days he can go weeks without playing. One reason we’ve reduced it is surprisingly simple: Roblox can be difficult to leave. My kid already knows that I hate it, because it’s definitely challenging making him stop — and he likes to play with his school friends on video calls over WhatsApp (fortunately that doesn’t happen for several weeks now).

Many of its popular experiences are persistent multiplayer worlds. You’re playing with other people. Something is happening, real time, “real” virtual people. Something else is about to happen. Leaving can mean abandoning whatever you and your friends are doing. “Five more minutes” (that is really one more hour) makes perfect sense to the child inside that world.

I find games with natural stopping points and the PAUSE button much easier to manage. Finish the match. Complete the level. Save the game. Stop.

This goes beyond children lacking discipline. Regulators are increasingly looking at the design itself. In February, the European Commission preliminarily found TikTok in breach of the Digital Services Act over what it called its “addictive design.” It specifically identified infinite scroll, autoplay, push notifications and highly personalised recommendations as features that can encourage compulsive use.

A television channel wanted me to keep watching too. Today’s technology companies have vastly better tools for achieving it.

Adults aren’t exactly winning this battle either.

Even Sam Altman couldn’t stop

Sam Altman told a story recently that I keep thinking about.

The OpenAI CEO started using TikTok because his company was developing Sora as a social product for AI-generated video. He wanted to understand TikTok and its recommendation system. His plan was to spend about ten minutes using it before bed.

Eventually, he found himself spending around three hours one Saturday afternoon scrolling TikTok. He compared the experience to a drug and eventually deleted the app because he found it too powerful.

There is something funny and revealing about the CEO of OpenAI deliberately studying an algorithm and still getting caught by it.

If Sam Altman needed to delete TikTok, perhaps we should cut 10-year-olds some slack. OpenAI has since discontinued the Sora social app.

The problem with measuring hours

There are reasons to worry beyond what I see at home.

The World Health Organization studied almost 280,000 children aged 11, 13 and 15 across 44 countries and regions. In 2018, 7% showed signs of problematic social-media use. By 2022, it was 11%. Among girls, the figure was 13%; among boys, 9%.

The definition matters, of course. Researchers weren’t simply counting hours. They looked for things such as difficulty controlling use, neglecting other activities, withdrawal-like symptoms and negative consequences.

That’s much closer to what concerns me with Vasco. An hour creating something on a computer has very different consequences from an hour disappearing into Shorts. Playing Minecraft with school friends is different from scrolling alone. Watching a movie with your parents is different from watching dozens of algorithmically selected videos in your bedroom.

The research also suggests a complicated relationship. A large 2025 analysis covering 117 longitudinal studies and more than 290,000 children found small effects running in both directions: more screen use predicted somewhat more later socioemotional problems, while children already experiencing socioemotional difficulties also tended to increase their screen use.

So I don’t want Vasco growing up believing that screens are simply bad. That would be a strange lesson coming from me. Technology has been an enormous part of my life and career. Computers gave me ways to write, create, learn and connect with people I would never otherwise have met.

What I want him to understand is that different technologies ask different things from him, and many of them are competing for his attention.

Three kids, three phones, zero conversation (zombies around)

The number of hours worries me less than what those hours replace.

Children need other children. They learn how people work by playing with them, negotiating rules, arguing, being included and excluded, making someone laugh, annoying someone and discovering they’ve gone too far. They gradually learn that another person can see exactly the same situation completely differently.

I remember watching the alternative on a tiny scale at Vasco’s old school. He idolised a boy who was a year older than him. The boy already had a smartphone at nine or ten and occasionally brought it to school, even though phones weren’t allowed.

One day I saw him with two other children, all around 10 years old, each holding a phone. Nobody was talking or playing. They were physically together, but each was absorbed in a different screen.

Adults do this constantly, and I’m guilty of it too. Seeing it among children felt different. Ten-year-olds should have plenty of time to be ridiculous together.

This is one part of Jonathan Haidt’s argument in The Anxious Generation that particularly resonates with me. His recommendations include more independence, free play and responsibility in the physical world, alongside delaying smartphones and social media.

And this is also where I recognise some of myself in Vasco.

He loves being at home. So did I. The difference is that I grew up with a forest around my house. The forest almost felt like another room. I could be “at home” while spending hours outside exploring and playing.

Vasco is growing up in Lisbon. Going outside means something different.

Fortunately, he loves being with other children. He gets that every day at school. During the summer he spends long days at camps, mostly without screens. He plays, swims, talks and comes home tired. I value that more as he gets older.

How much independence should a watch buy?

This brings me to a technology I’m considering buying him: a simple children’s watch that would let him call us and potentially let us see where he is.

The reasoning seems sensible. He’s changing schools. He’s getting older. He’ll gradually become more independent. A watch could give him the useful part of a phone without putting YouTube, Roblox, social media and the whole Internet in his pocket.

Haidt actually recommends basic phones or watches as an intermediate step before smartphones. I’m still unsure whether Vasco needs one.

A 10-year-old being able to call his parents whenever he wants sounds reassuring. Being able to see where he is sounds even more reassuring. Perhaps knowing his location would make us comfortable giving him more independence. Yet I also wonder whether both of us need to learn that independence includes moments when Dad doesn’t know exactly where Vasco is. Mixed feelings here.

Technology gives parents capabilities our parents never had. Deciding which capabilities to use is now part of parenting too. And I don’t have a definitive answer to me and to others.

AI can help him think. It can also become someone to talk to

Then there is AI.

Vasco already knows how to use it. Occasionally, probably less than once a week, he borrows my phone and opens an AI project I’ve created specifically for him. I’ve given the model instructions about his age and how I want it to interact with him.

Sometimes he asks questions about Harry Potter: why something happened, how some detail of the story works, whether his interpretation makes sense. He’s also writing a script and sometimes talks through ideas with the AI.

One rule I gave the model is particularly important to me: it isn’t allowed to write the script for him. Vasco has to bring the ideas. The AI can ask questions, explain things and help him develop what he has already imagined.

I like that use of AI. I can see what he’s doing, I’m usually nearby, and the machine is helping him with something he already wants to learn or create.

The relationship changes considerably when a child starts treating the machine as a confidant or companion. There have already been high-profile cases involving teenagers who died by suicide after prolonged interactions with AI chatbots, including lawsuits alleging that chatbot interactions contributed to the deaths. Those cases don’t establish a simple causal relationship between an AI conversation and suicide, but they expose a new category of risk for parents.

Vasco is still 10. By the time he is 14 or 15, we’ll have several more years of evidence about how children interact with AI, and hopefully better safeguards. I also expect the technology to become vastly more convincing.

I did several episodes of my first podcast, Made in Tech, about social media and its effects. The lesson I took from those conversations was the importance of digital literacy: understanding how these products work, how they compete for our attention and how they can affect us. Children need that literacy, but so do adults. I certainly do. After all those interviews and years spent thinking about this, I’m still kind of addicted to Twitter.

For now, I want Vasco to think of AI as a useful tool that can help him learn and create, while leaving the thinking and the human relationships to him.

Some problems are still waiting for us

AI companionship is only one of the things waiting further down the road. There is social media, pornography, dating, online status, body image and all the strange expectations the Internet can create about relationships (and the wrongs social media also brought in a wider scale).

I’m not going to pretend I have answers to parenting problems I haven’t yet had to deal with.

But I keep returning to the importance of learning how to deal with other human beings. Real people disagree with you. They get bored. They misunderstand you. They don’t always want to play the game you want to play. Eventually, the person you like might not like you back.

Children need opportunities to experience awkwardness, disagreement and rejection, and to learn what happens next. Playing with other children is an early version of that education.

Technology increasingly offers alternatives with far less friction. Pornography can distort expectations around sex. AI companions can always be available and attentive. Algorithms can become exceptionally good at supplying whatever keeps us engaged.

A child still has to learn to live among people who don’t have an engagement algorithm.

Building is becoming the easy part

Vasco is already getting a small taste of another side of AI: how easy it can make building things. We created Casa da Magia together, a Harry Potter website that started as a place to save his favourite book covers and grew through his ideas into something much bigger, with quizzes, educational games and even a Harry Potter version of the old Nokia Snake game. I did the actual building with AI, while Vasco became a kind of 10-year-old product manager: deciding what he wanted, testing things, giving me pointers and coming back with new ideas. He now understands that an idea for a website or a little product doesn’t have to remain an idea for very long.

This connects with something I heard Peter Steinberger, the creator of OpenClaw, discuss this week in his Y Combinator talk, “What Happens When 4.7 Million People Let It Cook”. As AI makes building dramatically easier, the scarce parts move elsewhere: having a reason to build something, making something people actually want, and finding a way to distribute it.

That’s something I’d like Vasco to learn from AI too. For now, his experience with building is limited to our projects together and some occasional Minecraft and Scratch at school. Being able to build almost anything is powerful. Figuring out what is worth building, why you’re building it, and how to get it in front of people will increasingly be the harder part.

Eventually, I won’t be setting the controls

I don’t have a grand solution. Vasco is 10, and me and my wife are figuring this out as he grows.

But I’m starting to have some rules of thumb. I pay attention to whether a technology has a natural stopping point and what the screen is replacing. Creating something matters differently to me than consuming an endless feed. I want him spending large amounts of time with other children without devices mediating the interaction.

I’m delaying the smartphone. I’m still deciding about the watch. With AI, I’m trying to show him how it can be useful while keeping it close enough that I understand what he’s doing.

Most importantly, I want to understand the products he uses rather than simply telling him they’re bad. He’s going to grow up in a world filled with AI, algorithms and screens. Banning everything until he’s 18 would teach him very little about how to live in that world.

For now, I can set limits on Roblox. I can keep the phone away. I can configure his AI project. I can choose whether he gets that watch.

Eventually, I won’t be setting the parental controls.

He’ll need to know when to press stop himself.

PS: Dad was more excited about the eclipse

There was a solar eclipse over Europe this week, and I took Vasco and my wife to a park near our home in Benfica to watch it. I also built a dashboard showing how Internet traffic changed during the eclipse: country-level traffic fell as much as 23% in Spain, 22% in Iceland and 17% in Portugal, with much larger drops in some regions.

I was considerably more excited than my 10-year-old.

Through the eclipse glasses, the Sun looked extraordinary. The Moon covered most of it, leaving only part of the Sun visible. What surprised me was how much light remained. The park became dimmer and the light looked strange, but it wasn’t remotely dark.

A tiny exposed portion of the Sun is still incredibly bright. Totality is a completely different experience, which I hope we’ll get to see together someday.

Perhaps Vasco would have been more impressed by 100%. Or perhaps eclipses simply aren’t his thing.

That’s another lesson children teach you. You can show them the books you love, the movies you love, the games you love, the places you love and a sky you find extraordinary. Then they decide for themselves what they find interesting.

Annoying, really.

Things I read, noticed or saved this week

• Our Last Wi-Fi-Less Refuge Is Gone. Subpar airplane connectivity gave us a few hours when being offline was normal.

• OpenAI and Anthropic’s AI price war. Model pricing is falling as competition increases and Chinese AI companies gain ground.

• Gemini 3.7 Flash and DeepSeek Harness. Two AI releases that got attention this week.

• Cerebras brings ultrafast inference to GPT-5.6 Sol. Cerebras on accelerating inference for OpenAI’s GPT-5.6 Sol.

• Mistral expands European infrastructure for sovereign AI. In-region inference, open models and additional European compute infrastructure. Also: Mistral OCR 4.1.

• Reddit is joining the S&P 500. Reddit’s addition to the index and some of the concerns analysts still have about the company.

• How Organizations Use AI: Evidence from ChatGPT. OpenAI research on how organisations are using ChatGPT. There’s also a Hacker News discussion.

• Blog about things you don’t understand yet. Sean Goedecke on using writing to investigate subjects you haven’t figured out yet.

• Terabytes of credentials leaked in a massive supply-chain attack. A supply-chain compromise and the credentials exposed through it.

• Where did the old web go? We followed 657,607 links to find out. An analysis of 657,607 links looking at link rot and how much content from the older web remains accessible.

• The global submarine-cable network can now detect whales, even when they’re silent. How submarine fiber-optic cables can be used to detect whales and track their movements.

• Peter Steinberger: What Happens When 4.7 Million People Let It Cook. A Y Combinator talk with Peter Steinberger about building OpenClaw and its growth.

• This entrepreneur will pay you to provide home care for your loved ones. Abby Care founder Havi Nguyen is building a service that helps family members become paid caregivers through Medicaid programmes.

• Water system controllers don’t belong on the Internet, says ex-NSA chief. Former NSA director Rob Joyce argues that water-system industrial controllers shouldn’t be directly exposed to the Internet.

• AI Coding and Mental Health. Scott Tolinski looks at positive and negative mental-health effects of AI-assisted coding using survey data and expert commentary. The survey data is here.

Learned a lot with this video from Riley Brown.

Tinha alguma esperança, mesmo depois de ver o trailer. Carlos Paião foi um dos artistas que mais gostei de ouvir quando era miúdo. Lembro-me bem do choque quando morreu, em 1988, num acidente de automóvel. Eu tinha sete anos e foi provavelmente uma das primeiras vezes em que percebi de forma muito concreta que alguém podia simplesmente desaparecer.

Talvez por isso esperasse tanto deste filme.

Carlos Paião era, para mim, uma combinação muito particular de talento, inteligência e alegria. As letras podiam ser românticas, brincalhonas, sarcásticas ou certeiras. Havia energia nele. Havia humor. E havia uma facilidade extraordinária para brincar com as palavras e com a música.

É precisamente essa energia que sinto que o filme nunca consegue captar.

Custa-me escrever isto porque sei que houve muita gente a trabalhar para que este filme existisse, e fico contente que exista. Mas o resultado parece-me pobre, por vezes tristonho, e demasiado distante da personalidade que associava a Carlos Paião.

Alguns dos problemas começam logo na execução. O som tem falhas evidentes. Muitos cenários parecem demasiado artificiais, quase sempre entre estúdio e teatro, provavelmente também por limitações de orçamento. A realização raramente consegue tirar partido das interpretações dos atores, sobretudo nos momentos em que seria importante ficarmos próximos das personagens.

Mas o maior problema, para mim, está no guião e nas escolhas da história que decidiram contar.

Há um foco enorme na relação entre Carlos Paião e a namorada que viria a ser sua mulher. Isso faz sentido, e alguns desses momentos funcionam. O problema é que o filme dedica muito tempo a conflitos e pequenos episódios dessa relação, enquanto deixa quase por explorar aquilo que mais me interessava perceber: de onde vinha aquele talento?

Como começou a brincar com palavras daquela forma? Como escrevia? Como treinava? Que influências teve quando era mais novo? Que tipo de miúdo era? Não há praticamente infância, memórias do passado ou sequer referências capazes de nos ajudar a compreender como surgiu o Carlos Paião que viria a escrever aquelas canções.

A história começa já com ele a estudar Medicina e a compor músicas. Um amigo envia uma cassete, as coisas começam a acontecer, e avançamos. Mas existe muito pouco sobre a construção daquele talento.

Até as entrevistas jornalísticas que aparecem são usadas de forma muito superficial. Pequenas frases, alguns clichés, pouco mais. E isso é especialmente frustrante porque existe muito material real de Carlos Paião que mostra exatamente aquilo que falta ao filme.

Basta procurar vídeos dele com Herman José ou com o grande Júlio Isidro, por exemplo. Há momentos extraordinariamente divertidos, em que se percebe como ele tinha rapidez, sentido de humor e enorme facilidade em brincar com música e palavras. Escreveu para muita gente, colaborou com artistas muito diferentes e tinha uma personalidade televisiva extremamente natural.

Lembro-me até de um vídeo que vi quando era miúdo, gravado em casa dele, com a mulher. Era divertido, descontraído, cheio de personalidade. Esse Carlos Paião quase não aparece aqui.

O filme tenta também introduzir momentos de musical. De repente, uma situação transforma-se numa música encenada quase como se estivéssemos num musical clássico. A ideia podia funcionar, mas o filme nunca assume verdadeiramente esse formato. Esses momentos acabam por parecer longos, pouco integrados na história e, algumas vezes, simplesmente aborrecidos.

E isso é talvez o mais estranho: um filme de cerca de 90 minutos parece durar três horas.

Mesmo tendo músicas muito bonitas de Carlos Paião.

O ator principal, Rafael Ferreira, tem algumas parecenças e, apesar de fisicamente ser bastante mais alto, consegue captar alguma coisa dele. Não totalmente a energia e positividade, mas há momentos em que resulta. A Laura Dutra, atriz que interpreta a mulher, que também participou em 1986, é talentosa e tem cenas emocionalmente fortes. O actor que faz de Peter também se safa bem (mas depois há um dos outros amigos rapazes do Paião no filme que só prejudica os outros… nada convincente).

O problema maior é que a realização (e o guião) raramente ajudam. Há momentos de tristeza em que a câmara parece quase indiferente ao que a atriz está a fazer.

Também há algumas personagens que, para mim, empurram o filme para a caricatura. A personagem de Albano Jerónimo, ligada à editora e à carreira de Carlos Paião, tenta ter um lado sarcástico e pomposo que raramente resulta. Mesmo em momentos mais sérios, a interpretação e a escrita mantêm um tom tão caricatural que às vezes provocam o efeito contrário ao pretendido.

A personagem da médica e professora de Carlos Paião sofre do mesmo problema (é mesmo a personagem que mais me irrita). A expressão, o tom, a forma como está escrita e interpretada parecem quase sempre iguais (nada contra a actriz Anabela Moreira… mas as opções aqui… falharam muito na minha perspectiva… que vale apenas e só isso). Tanto num momento que tenta ser divertido como num momento sério ou triste, há uma rigidez que transforma a personagem noutra caricatura.

Os momentos oníricos ou de sonho, usados para explicar a origem de algumas músicas, também me pareceram muito pouco conseguidos. São visualmente artificiais e, sobretudo, não nos ajudam verdadeiramente a perceber como aquelas canções nasciam. Fica tudo demasiado básico: ele tem uma ideia, escreve qualquer coisa num papel, nasce uma música.

E o Carlos Paião era seguramente mais interessante do que isso. Curioso como o filme até tinha o apoio da RTP mas nunca faz uso das imagens de arquivo da RTP para dar contexto e textura ao filme. É pena (claro desperdício).

Há aqui uma tentativa, talvez inevitável, de seguir o caminho de outros filmes portugueses sobre figuras da música. O filme das Doce pareceu-me muito mais conseguido nesse sentido, porque conseguia tirar energia das próprias personagens e da música. Não vi integralmente o filme sobre António Variações, mas dos clips que vi pareceu-me também mais capaz de captar a energia do artista.

Curiosamente, no dia seguinte vi na HBO Uma Quinta Portuguesa (da espanhola Avelina Prat), um filme passado em Portugal, com Maria de Medeiros e vários atores portugueses. É uma obra muito mais lenta no ritmo, muito mais próxima do cinema de autor, sobre um professor espanhol que acaba por mudar de vida e trabalhar como jardineiro numa quinta portuguesa.

Há muito menos ação. Há menos acontecimentos. Mas existe uma história simples, clara e um guião que sabe exatamente o que quer fazer. Funcionou muito melhor para mim (e viu-se melhor em 114 minutos do que os 90 min de Playback).

Também vi recentemente o trailer do novo filme de Ruben Alves, o realizador de A Gaiola Dourada. Mesmo através de um trailer sente-se imediatamente outra dinâmica, outro cuidado na imagem, no som e no ritmo. E dei por mim a pensar como teria sido interessante ver Ruben Alves pegar na história de Carlos Paião.

Não tenho nada contra Sérgio Graciano. Simplesmente acho que neste filme houve demasiadas escolhas de execução que não resultaram.

A doce ironia no meio de tudo isto é que estava na mesma pequena sala das Amoreiras onde vi o filme um grupo de jovens que bateram palmas no final. Fiquei surpreendido, mas depois percebi que eram amigos do Rafael Ferreira, que faz que Carlos Paião, que também estava por lá. E honra lhe seja feita, até se safa bem face às limitações já faladas (pergunto-me se a opção de o por a cantar com o que parecem dentes postiços — talvez para tentar mostrar o dente partido que o Paião tinha —, e que lhe deu um estilo sopinha de massa foi bem pensada…)

Sei que há pessoas que gostaram muito, e ainda bem. Os meus amigos e “críticos” Rui Pedro Tendinha (Expresso/TSF) e Rui Pedro Vieira (CM) gostaram. Para minha surpresa, até Vasco Câmara, do Público, parece ter apreciado o filme, quando normalmente é bastante exigente com este tipo de cinema. Eurico de Barros, no Observador, esteve bastante mais próximo da minha reação.

Eu simplesmente não consigo gostar.

Ainda assim, prefiro que o filme exista. Se levar pessoas a procurar Carlos Paião no YouTube, a ouvir novamente as músicas, a descobrir aquelas aparições televisivas incríveis e a perceber melhor o talento que ele tinha, já fez alguma coisa importante.

E talvez ainda haja espaço para outra coisa: um bom documentário sobre Carlos Paião. Há imagens suficientes, entrevistas, atuações e memórias capazes de mostrar a inteligência, a alegria e a energia que este filme, para mim, nunca conseguiu verdadeiramente encontrar.

I’ll be sharing some news soon about what’s next. But before turning that page, I wanted to reflect a little on this one.

Throughout my five years at Cloudflare, I almost always felt like I was still doing journalism. The subjects changed, but the curiosity didn’t. I was still asking questions, interviewing people, trying to understand complicated topics, and then finding ways to explain them clearly.

Before Cloudflare, I had interviewed entrepreneurs, innovators, scientists, filmmakers, actors, and many others. At Cloudflare, my conversations shifted toward Internet engineers, researchers, product leaders, security experts, and developers. I also had the privilege of exploring Internet-scale datasets, learning how to spot meaningful trends, understand what they meant, and turn them into stories that people could understand. The tools evolved, but the curiosity never did.

[Cloudflare was also a wonderful place to be curious, with experts who were always generous with their time and knowledge, ready to explain even the most complex topics (the Internet is far more fascinating and complex than most of us realize… but that’s a story for another day).]

Over these five years, I had the privilege of helping tell stories about the Internet through data. From the war in Ukraine and global sporting events to elections, cybersecurity, bots, major outages, and Internet traffic trends, my goal was always the same: make complex technical topics understandable, useful, and interesting. I genuinely loved doing it.

One of the unexpected privileges of working at Cloudflare was gaining historical perspective, not only about the company but also about the Internet itself.

Early on, I came up with an internal series called A Bit of History. John Graham-Cumming encouraged me to pursue it, and it became one of my favourite occasional side projects. It was built around long conversations with people who had helped shape Cloudflare over the years (and were leaving the company), especially across engineering, research, although there was also product, marketing, and revenue.

People like Ólafur Guðmundsson, Nick Sullivan, Usman Muzaffar, Marek Majkowski, Ignat Korchagin, Chris Merritt, and many others shared stories about decisions, challenges, mistakes, and ideas that became part of Cloudflare’s DNA. The series came full circle in 2025, when I sat down with John Graham-Cumming for a conversation that lasted more than three hours. Understanding why things had been built often became just as interesting as understanding how they worked. Those conversations gave me a much deeper appreciation of both the company and the Internet we were helping build.

Most people know Cloudflare through its products. I had the privilege of helping explain what those products revealed about the Internet.

With an amazing team in Lisbon, a handful of colleagues in the US, and the collaboration of many teams across the company, we built Cloudflare Radar into one of the best windows into what’s happening across the Internet. Along the way we created blog posts, reports, videos, presentations, Radar Display, and eventually This Week in NET, a podcast that ran for almost four years with more than 150 episodes and over 100 guests.

That podcast became one of the highlights of my time at Cloudflare. I had the opportunity to interview Matthew Prince and Michelle Zatlyn several times, as well as researchers, engineers, product leaders, Internet pioneers, and many people helping shape today’s Internet. Every conversation was another opportunity to learn.

One of the things I enjoyed most was that no two weeks looked alike. One week I might be analysing Internet traffic around an election. The next I could be preparing content for Davos, recording a podcast, brainstorming a Radar feature with engineers, helping build Radar Display for an event, or sharing data trends with journalists across Europe. That variety never stopped being fun.

Some projects became much bigger than I ever imagined.

Radar Display started as an idea and eventually became part of Cloudflare offices, partner offices, conferences, and events around the world. Watching people stop in front of a giant globe showing the Internet in real time never got old. One of my favourite memories was making a huge installation work remotely for an event in Dubai together with André Páscoa and the local team. It took many hours, many video calls, and plenty of problem solving, but seeing it all come together made every hour worthwhile.

Another project that became part of my yearly rhythm was the Cloudflare Radar Year in Review. Early on, I proposed telling part of the year’s story through the world’s most popular Internet domains. It started as a standalone blog post and later became an integral part of the Year in Review. Every year after that, I spent weeks revisiting the year’s biggest events through Internet traffic and changes in domain rankings. Looking back at how elections, sporting events, holidays, product launches, unexpected global moments, and even shifts in how people searched, shopped, or consumed media influenced the way people used the Internet became one of my favourite exercises.

It wasn’t just about rankings (also important). It was another way of telling the story of the year through Internet data. Year after year, it reminded me that behind almost every major moment in the world, there was an interesting Internet story waiting to be discovered (or shared).

One of the most rewarding parts of the journey was seeing the work travel far beyond Cloudflare.

Students turned some of our analyses into classroom projects, including beautifully illustrated stories. Journalists, researchers, engineers, policymakers, and simply curious readers picked up the trends we uncovered.

Some stories became much bigger than expected. Internet disruptions in Tonga. Ramadan and Easter. New Year’s Eve. The total solar eclipse across Mexico, the United States, and Canada. Eurovision. Euro 2024. The Olympics. Dozens of elections. TikTok’s rise and also brief ban in the United States. Malicious email TLDs. And one of my favourites: exmaple.com, a domain that many people had never noticed despite typing example.com for years.

Seeing those stories featured by the BBC, The New York Times, CNN, Globo, and hundreds of other publications was incredibly rewarding. Discovering that data from some of our work eventually found its way into Wikipedia pages was another unexpected moment that made me smile.

John Graham-Cumming shaped my Cloudflare journey more than anyone else. He hired me, but more importantly, he gave me an unusual amount of freedom. He constantly encouraged me to build things people would genuinely find interesting, trusted me to move across topics and priorities, and supported ideas that didn’t always fit neatly into a job description. A Bit of History, This Week in NET, Radar Display, many reports, videos, and side projects simply wouldn’t have existed without that trust.

David Belson became a great collaborator throughout the years, around Radar. Whether we were discussing content, product ideas, strategy, or simply trying to understand what the data was telling us, working alongside him was always a pleasure.

The Radar team became a place where ideas never seemed to run out. Jorge Pacheco, Carlos Azevedo, Sabina Zejnilovic, Sofia Cardita, Carlos Rodrigues, Paula Tavares, André Jesus, Nuno Pereira, Mingwei Zhang, Tiago Dias, André Páscoa, Vasco Asturiano, and many others made those years not only productive but genuinely enjoyable. We spent countless hours discussing data visualizations, product ideas, interactions, stories hidden in the data, presentations, and sharing great lunches (and sometimes dinners). Those conversations often wandered far beyond work, but they usually came back to finding new ways to understand the data, visualize it, and make Cloudflare Radar a little bit better.

The Research team constantly reminded me how generous technical people can be with their knowledge. Vânia Gonçalves, Marwan Fayed, Bas Westerbaan, Thibault Meunier, James Allworth, Doug Kramer, Nick Sullivan, Tom Strickx, Tom Paseka, Mike Conlow, Mark Nottingham, Justin Paine, Lucas Pardue, Ignat Korchagin, Michael Tremante, Omer Yoachimik, Trey Guinn, Kenton Varda, Brian Mitchell, Sunil Pai, Grant Bourzikas, Heather Orsi, Craig Dennis, Jocelyn Woolbright, and many others were always willing to answer another question, review another draft, record another segment, or patiently explain something until I understood it. Many of my favourite videos only happened because they were happy to indulge one more unusual idea.

Beyond Research, I was fortunate to work with exceptional people across Communications, Marketing, Network, Email, Agents, Events, People, Offices, Developers, Legal and Policy, Compliance, the Innovation Weeks task force, and many other teams. Anne-Caroline Tanguy, Kari Linder, Val Vesa, Daniella Vallurupalli, Leigh Ann Acosta, Luanne Da Costa, Ryan Knight, Gabi Palavecino, Reagan Russell, Lindsay Scalisi, Jenna Zak, Laurel Wamsley, Brittany Brown, Rita Kozlov, Alex Messe Rodriguez, Chy Sarmento, Caroline Quick, Jen Preston, Carolina Beja, Margarida Alhinho, Margarida Simão, Margarida Pereira (the two Margaridas behind the wonderful Book Club), Pedro Santos, Kevin, Matt Carey, Patrick Day, Emily Hancock, Christiaan Smits, Petra Arts, Janel Riley, Syona Sarma, Sharon Goldberg, Stephanie Cohen, Luísa Gobbo, Michelle Chen, Alonso Bustamante, Lara Cohen, Jerome Fleury, Alissa Starzak, and so many others helped transform technical work into stories, products, launches, events, and experiences that reached journalists, customers, developers, and readers around the world.

And then there’s Celso Martinho, whom I first got to know years before joining Cloudflare through a couple of interviews I did as a journalist.

Before I joined Cloudflare, he told me something I’ve never forgotten.

He said that what my role would become after a year at Cloudflare would depend much more on me than on Cloudflare itself. The opportunities would exist. It would be up to me to build something meaningful with them.

I took that advice seriously.

Throughout my time at Cloudflare, Celso also became an inspiration simply by watching him work. He always made me feel part of the team, even though I wasn’t an engineer. He led with empathy, passion, humility, and an extraordinary ability to make the people around him shine. He never seemed interested in making himself the protagonist. He preferred helping others become one. That is one of the leadership lessons I’ll carry with me.

The funny part is that years later I reminded him of that conversation. He didn’t remember saying it. I never forgot it.

The layoffs in May were difficult. Seeing so many talented colleagues leave, including people who had helped build some of Cloudflare’s most important products, was disappointing. Many had dedicated years of their lives to the company and had shaped it in meaningful ways.

But that doesn’t change how I look back at my own five years. I learned an enormous amount.

I worked with exceptional people. I helped build things I’m genuinely proud of.

Most of all, I’m grateful I had the opportunity to make the Internet a little more understandable for millions of people while learning from some of the people who understand it best. I’ll also remember the generosity of so many people who always found time to answer one more question, explain one more concept, or help make one more idea a little better.

I’ll continue to cheer for Cloudflare.

And I’m excited to start telling the next chapter.

In journalism, we often return to Marshall McLuhan’s idea that the medium is the message. We see it clearly today with social media and AI. The place where something is published changes the message: how it is edited, framed, distributed and understood.

In The Odyssey, there is something similar. The journey is the message.

Odysseus spends ten years fighting in Troy and another ten years trying to return home to Ithaca. What begins as a war story slowly becomes something else. The film doesn’t spend much time explaining why the Greeks wanted Troy in the first place. The reasons almost seem secondary, and the conquest itself isn’t presented as glorious. If anything, the real story begins after Troy falls.

In the original story, Odysseus is reluctant to join the war and even tries to avoid it. Nolan barely touches on that, but once Odysseus is there, he becomes essential. The idea of the Trojan horse comes from him. It gives the Greeks victory, but it is far from an honourable one. Victory doesn’t bring peace. It brings guilt. It begins the voyage home.

(WARNING: some light spoilers ahead)

Odysseus keeps making mistakes along the way. He attacks people he doesn’t need to attack. He enters the cave of the Cyclops when perhaps he could have avoided it. He challenges the gods out of pride, anger and impulse. Those choices have consequences not only for him, but for the men travelling with him: storms, deaths, mistrust and the constant feeling that the gods are punishing them. His own crew slowly begins to doubt him. They followed him into war. Now they are following him through a journey that seems to have no end. That’s what makes him such an interesting character. He isn’t simply a great hero leading men through danger. He often creates part of that danger himself, and learns through loss.

The film is about ambition, leadership, love and regret. It asks what happens when someone believes deeply in a goal, pursues it for years, and only much later understands the full cost of what he did. The journey home becomes a process of recognising those mistakes.

There is also a figure who appears to accompany him throughout the film, almost like a vision or the physical form of guilt. She follows him through his thoughts and decisions, constantly reminding him of Troy and of the lives that were destroyed.

The gods are everywhere, although Odysseus doesn’t initially seem to place much faith in them. Island after island, he encounters different forms of magic, punishment and temptation. Sometimes the gods feel literal. At other times, they feel like expressions of fate, guilt and human weakness.

One of my favourite scenes comes on the island where Odysseus lives with the character played by Charlize Theron. He has lost the memory of his previous life and has apparently spent happy years there. When he says he wants to know the truth, she asks him what happens if the truth destroys the happiness he experienced. His answer is simple: then those were not happy years. I found that idea surprisingly powerful because he would rather face an uncomfortable truth than live happily inside a false memory.

That also connects to the wider journey. Odysseus needs to remember who he is, what he did and where he belongs. Returning home isn’t only about reaching Ithaca. It’s about recovering the truth about himself.

Love is part of what drives him. The desire to return to his wife, his son and his kingdom never disappears. But he also loses himself several times along the way. He becomes distracted, tempted and trapped. The return is never a straight line.

That is also why the voyage feels believable. You can see it in Matt Damon’s body and in his face. By the end, he genuinely looks like someone who has spent years fighting, travelling, losing people and simply trying to survive. Matt Damon is perfectly cast here because he isn’t the obvious heroic figure. He’s believable as an intelligent but flawed leader who doubts himself, gets things wrong, and is brave without ever feeling invulnerable.

He’s a very different kind of hero from Brad Pitt’s Achilles in Troy. Pitt has the physical presence of a mythical warrior. Damon feels more like a man who became a hero through intelligence, persistence and difficult decisions, while never quite escaping his own weaknesses.

The supporting cast is strong throughout. Anne Hathaway brings warmth and emotional weight. Tom Holland is convincing as Telemachus, and his energy helps him feel genuinely young, still trying to understand the father he barely knows. John Leguizamo gives a strong performance as someone defined by loyalty but also damaged by years of absence and change.

Charlize Theron is excellent, especially in the scenes where truth, memory and happiness collide.

One of the surprises for me was seeing Robert Pattinson, whom I interviewed 16 years ago, playing one of Penelope’s suitors, hoping to inherit Odysseus’ throne, with a backstory closely tied to Odysseus himself (Ulisses in Portuguese). He’s convincing, morally ambiguous, and brings real depth to the story.

I also found Nolan’s approach to language interesting. Everyone speaks with American accents, even actors who naturally wouldn’t. I assume that was a deliberate decision to avoid distracting the audience. Personally, I wouldn’t have minded hearing the characters language (Greek?), but it never became a real issue.

Visually, the film is impressive. If I have one criticism, it’s that I sometimes wanted Nolan to trust cinema itself a little more: more silence, more moments where the camera simply follows the characters without dialogue explaining everything, and more room to fall in love with the people before they explain themselves. That’s one reason Dune affected me more as a cinema piece. The Odyssey is different. I actually think it has the stronger message about humanity, leadership and consequence, even if it didn’t move me as a cinematic experience in quite the same way.

Oddly enough, it also reminded me of Gladiator, not because the stories are particularly similar, but because both films ask the same question: what remains of a man after violence, ambition and loss have shaped his life? They also hold some other similarities.

There is also a wider historical idea running through the film. Humanity often creates paths from which it is incredibly difficult to return. A war, a conquest or a decision made by a leader can destroy societies and set them back for generations. Odysseus wants to return home, but there is no simple return to the world that existed before Troy. That world no longer exists. He can only return as the man the voyage has made him.

Listening afterwards to The Rest Is History‘s excellent two-part episode on The Odyssey added another layer for me. Perhaps the story has endured for nearly three thousand years because it is ultimately about something so human through different ages. We make mistakes, we carry them for a long time, and somehow we endure, adapt and keep moving forward. Odysseus’ voyage belongs to another age (a very violent and god-driven one), but the emotional journey still feels surprisingly familiar.

The film isn’t perfect. It isn’t as emotional as I hoped, and at times I wanted it to trust the images more and the words less (some of the story arcs could also be better). But it is ambitious, and it resonates. It gives us a genuine sense of travelling with these characters through years of war, mistakes, gods, temptation, loss and regret.

More than anything, it is a film about a man trying to get home while slowly understanding what he did, what he lost, and what his decisions cost everyone around him. Ithaca is the destination.

The real journey is learning to live with the truth.

Part of the reason I ended up spending the early afternoon in the Gulbenkian Art Library

was simply to escape the heat. But after a few minutes inside, the weather outside almost stopped mattering.

Perfect temperature (great air conditioning!), comfortable chairs, old wooden desks unlike any I’ve seen elsewhere, built the right way, in so many ways, a view over the gardens, and the kind of silence that makes it easy to think.

Some places seem to slow time down.

I still had my library card from 2000. I even got a new physical one. While renewing it, I chatted with the librarian, who’s been working there for decades. She told me someone had just called to offer the library the 98,000 books that belonged to his grandfather, who had recently passed away.

Ninety-eight thousand books. Out of curiosity, I later worked out what that meant (yes, I asked an LLM, hopefully it didn’t hallucinate too much). Somewhere around 2.5 to 3 kilometres of shelving. An entire lifetime (and a lot of money spent, I bet) collected on shelves inside one house.

The last time I regularly came here, to this library, I was 19, studying communication. Our university was just down Avenida de Berna. We’d come here to work on assignments, wander through the shelves and spend afternoons in what already felt like one of Lisbon’s special places.

I remember being impressed by the library itself. The silence. The furniture.

The feeling that someone had built this place properly, with pleasure and intent.

I also remember looking for information about the painter José Malhoa, from my hometown of Caldas da Rainha, and finding a publication titled The Conquest of Lisbon by a Man from Caldas.

My friends called me “Caldas” at the time. I wore the nickname proudly. I remember reading that title and thinking: that’s what I want too. Build a life in Lisbon. Succeed.

Today, 26 years later, I found myself sitting in the same chair. It still feels just as comfortable. I’m 45 now.

The path between those two afternoons wasn’t exactly the one I imagined. There was journalism, technology, storytelling, data visualisation, podcasts, interviews.

Life rarely follows the version you write in your head at 19 — although I wasn’t big on specific plans. I’ve often said to myself over the years that one of my goals in life is to “make right by” the curious, creative kid I was at 7, 8, 9 and 10 years old.

I don’t know if there’s ever a clear answer to that. But it felt good to come back. To remember where I started, in a way. To realise that some places don’t just preserve books. They preserve pieces of us.

Before leaving, I wandered through the shelves one last time and found a beautiful two-volume work on typography. I leafed through it. And, of course, I took a few photos and came away with a few ideas. Serendipity works well, if you’re open to it.

It’s not a big deal, but recently I had some time in my hands (it’s actually Cloudflare related), and wanted to give a more professional look to the email I share externally. So, I wanted to put together, with no costs, a simple email address: jt@joaotome.com.

I didn’t want to pay for Google Workspace. I didn’t want another inbox to check. And I didn’t want to give up Gmail, which is where I already spend most of my email time (next to Outlook — yes, I have a very old Hotmail account, from 1998).

After some trial and error, I ended up with a setup that lets me receive email at my own domain (joaotome.com), keep Gmail as my main inbox, and send email from Gmail using my custom address.

The result feels almost invisible. I still use Gmail exactly as before. The only difference is that the email address belongs to me.

One thing that confused me when researching this setup is that many articles still imply it is entirely free.

That’s not quite true anymore (since June 9).

As of June 2026, Cloudflare Email Routing is available on the Workers Free plan, but Cloudflare Email Sending currently requires the Workers Paid plan, which costs $5/month and includes 3,000 outbound emails per month.

So the honest version is simple: receiving email at your custom domain can be free. Sending email through Cloudflare’s Email Sending service requires the $5/month paid plan.

That said, you may not need Email Sending at all.

If all you want is a custom address that forwards to Gmail, Email Routing is enough and remains free. You can receive mail at you@yourdomain.com and continue replying from your normal Gmail address. For many people, that’s perfectly fine.

I chose to enable Email Sending because I wanted every email I send from Gmail to come from jt@joaotome.com. That was worth the extra $5 per month to me.

My setup is fairly simple. I use jt@joaotome.com as the public address, seremot@gmail.com as the actual inbox, Cloudflare Email Routing for incoming mail, Cloudflare Email Sending for outgoing mail, and Gmail as the interface where everything happens.

The first requirement is that your domain uses Cloudflare DNS. Once that’s in place (this is the sort of thing ChatGPT, Claude or Gemini can walk you through, and you’ll also need a free Cloudflare account), enabling Email Routing is straightforward.

Cloudflare adds the required MX records and routing configuration, then you choose where incoming mail should go. In my case, every message sent to jt@joaotome.com is forwarded to Gmail.

One thing worth checking before you start is whether your domain still has old email records from another provider. I had leftover Google Workspace MX records and Cloudflare refused to enable routing until I removed them. The error message was accurate, but my LLM understood it and guided me through the fix.

The solution was simply to go to the DNS section in Cloudflare and delete the old MX records that were still pointing to Google’s mail servers. A minute later, Email Routing worked.

Once Email Routing is working, you can create as many forwarding addresses as you like. For most people, a handful is enough: your main address, perhaps a hello@ address, and maybe one or two aliases. It’s easy to create dozens of them, but I suspect most of us won’t use them.

The outbound side is where things become slightly more complicated.

To send email as your custom domain from Gmail, Cloudflare needs to onboard the domain for Email Sending and create a few additional DNS records. This is where I hit my second problem. Cloudflare reported incompatible records under _domainkey, which turned out to be old DNS delegation records left behind by a previous provider. Once I removed them, onboarding completed successfully.

Cloudflare uses an API token as the SMTP password. After creating a token with the Email Sending: Edit permission, I added my custom address inside Gmail under “Send mail as” and configured the SMTP settings using Cloudflare’s server.

The important details are:

• SMTP server: smtp.mx.cloudflare.net

• Port: 465

• Username: api_token

• Password: your Cloudflare API token

• SSL enabled

One thing that caught me out is that Cloudflare expects SSL on port 465. STARTTLS on port 587 does not work.

After Gmail verifies ownership of the address (a few minutes), everything is basically done. The verification email arrives through Email Routing, you click the link, and Gmail can start sending as your custom address.

Looking back, the actual setup took less time than figuring out which tutorials were still accurate. Most of the work was simply following the instructions from an LLM and applying them one step at a time.

In fact, if you’re using tools like Codex or Claude Code, parts of the setup (not all) can be done directly from the chat (or Terminal) using Wrangler, without touching the Cloudflare dashboard at all. That’s a different setup and outside the scope of this post, but it’s a good example of how quickly these workflows are changing (that’s how I ship most pages, including changes I do to joaotome.com).

The biggest benefit isn’t really the technology. It’s ownership.

I like Gmail today, but I may not like it forever. If I ever decide to move to another provider, I can keep exactly the same email address and simply point it somewhere else. Nobody needs to learn a new address. Nothing on my business cards, website, or podcast pages needs to change.

The address belongs to me, not to Google.

That’s really what all of this is about.

Cloudflare is still part of the equation, of course. But changing infrastructure is easier than changing your identity.

Then, in a brief exchange on Twitter a few days ago, I asked Jorge Ventura, an Azorean navigator and author, if he had any reading suggestions, and he pointed me to A Marinharia dos Descobrimentos, by A. Fontoura da Costa (from the 1930s). It was exactly the right recommendation.

This is still not it (the full “writing”, I mean), but it’s a start.

Leafing through the book, I found precisely the kind of material that fascinates any geek with a taste for history. Not sweeping abstractions about national destiny or heroic “epic” narrative, for which we already have Camões’s Os Lusíadas or the stories of Fernão Mendes Pinto, but objects, diagrams, and technical solutions to very concrete problems. What stands out most is that the Portuguese did not advance through magic, nor through some sudden explosion of isolated genius. They advanced because they took existing technology and made it work at sea.

Astrolabes reproduced in A Marinharia dos Descobrimentos: inherited technology adapted for Portuguese nautical use.

The astrolabe is the obvious example. It was not a Portuguese invention, yet it appears here as a working tool, almost like repurposed hardware adapted for an extreme environment. The same can be said of astronomical navigation more broadly: diagrams of Polaris, constellations, and sky-reading schemes, visual guides for astronomical orientation, which in practice turned the night into a usable interface.

Astronomical orientation scheme based on Polaris and the constellations: the sky as a navigational interface.

The hourly or equatorial rose for tides: a conceptual device linking direction, time, the Sun, the Moon, and the tide.

But there are even better details. One of them is the so-called hourly or equatorial rose for tides. It is not just a compass rose: it is a conceptual device for relating direction, the Sun, the Moon, time, and tidal movement. Almost a small analog computer. Another is the succession of maps and charts in imperfect and corrected versions, where you can actually watch knowledge being refined iteratively, as if you were looking at successive builds of a cartographic system still in debug mode.

Chart of the island of São Lourenço, now known as Madagascar: cartography as an iterative process of observation, correction, and refinement.

It also helps to see that this technical explosion did not emerge from nowhere. First under Prince Henry, and later under John II, there was a political decision to create incentives to explore, measure, and profit. Henry sponsored voyages, brought together pilots, cartographers, and experts in nautical astronomy, and helped turn navigation into a sustained project.

John II was even colder and more effective, and he seems to have relished keeping things secret and out of Castilian hands, something Manuel I later diluted. He reinforced the royal monopoly over African trade, backed specialists such as José Vizinho and Abraham Zacuto, and made the search for a sea route to India a matter of state priority.

Another detail I especially like about this early phase of Portuguese expansion is that the Casa da Guiné, and later the Casa da Índia, which made that progress possible, were not modern private companies. They were Crown institutions set up to manage monopolies, organize fleets, gather information, control commodities, and turn discovery into revenue, while still leaving room for contracts and leases with private actors. One of my favorite stories from the Discoveries, in fact, took place not by sea but over land, in the lonely figure of the spy-explorer Pêro da Covilhã.

Pages on the Casa da Mina, the Casa da Índia, and nautical astronomy: navigation also meant organizing information.

And the engine behind all this was not merely scientific curiosity, though some certainly had that too. It was commerce, power, and religious legitimacy: gold, spices, direct routes, papal backing, evangelization, and prestige in the eyes of the clergy.

Once this model began to produce results, imitation was inevitable. Castile responded with Columbus, after he had been rejected by the Portuguese king, who may well already have suspected the existence of America, and perhaps even Brazil, and with its own imperial machinery. Later, the Dutch and the English pushed the same impulse into a more explicitly commercial form, through chartered companies and private capital.

Seen up close, the Discoveries look like a clear mixture of applied science, improvised engineering, intelligent reuse of inherited technology, and an obsession with turning the unknown into something measurable, navigable, and commercially useful. And from that came very different consequences, some extraordinary, others brutal, just as the world in which all this happened was itself brutal and violent.

On the one hand, a small and peripheral kingdom, “trapped” between Castile and the Atlantic, while also facing the African coast, tried to break through its limits by way of the sea and of shorelines still little known, in a movement that Portuguese tradition would later romanticize with the grand formula of “giving new worlds to the world.”

On the other hand, that expansion came at a tremendous human and moral cost. There was the extreme hardship of the voyages, with high mortality among crews; the growth of the slave trade on an ever larger scale; and, in the Americas, the violence, disease, and devastation imposed on native populations, in a process in which the Castilians reached particularly aggressive and destructive levels. All of this belongs to the same story and cannot be separated from its technical and commercial dimension.

At the same time, the expansion itself generated moral and political tensions that would later feed criticism, debate, and attempts at restraint, in figures such as Father António Vieira. But that does not erase the essential point: the same impulse that produced knowledge, routes, wealth, and power also produced exploitation, suffering, and destruction.

As the decades and centuries passed, part of that scientific, technological, and mercantile impulse gradually decayed, as later kings preferred projects of grandeur, display, and ostentation to a more disciplined culture of knowledge and commerce.

Portrait of Prince Henry, a central figure in the early patronage of Portuguese maritime expansion.

TL;DR: Frontier models matter more than ever, and so does who controls them. Anthropic builds the best, OpenAI is catching up, and the week showed that access to the most capable AI can disappear overnight for political reasons. How companies integrate AI, what they do with their data, and whether they can switch providers will define the next few years. Open source is becoming less of an ideological choice and more of a survival strategy. Europe is an ocean behind, with no clear plan to catch up, despite some promising initiatives.

Some weeks in tech are just noise. Last week wasn’t one of them.

Three things happened in quick succession that, taken together, say something worth paying attention to — not just about AI companies, but about who controls them, who benefits from them, and what the real race might actually be about.

First: the Fable 5 and Mythos 5 episode

Anthropic launched its two most capable models, Fable 5 and Mythos 5, and within days, the US government issued an export control directive barring access by any foreign national. And that’s whether inside or outside the United States, including Anthropic’s own employees.

Anthropic complied but disputes the directive, arguing the standard would effectively halt all new frontier model deployments across the AI industry. The company said it believed the order was triggered by a jailbreak technique. But that models from other providers, including GPT-5.5, possess the same capability.

The result: a model presented as the future of AI became unavailable because of a political decision made on a Friday afternoon.

For anyone outside the United States, this was a vivid reminder of something that has often felt abstract. The world’s most advanced AI systems are increasingly strategic assets. Access to them is controlled by a handful of American companies and shaped by American political decisions. Europe has spent years discussing technological sovereignty. This week, it saw what dependence actually looks like in practice.

Then came Satya Nadella (Microsoft)

On Sunday, Microsoft CEO Satya Nadella published a long essay (rare for him, and interesting) arguing that most companies are focused on the wrong question. The title said it plainly: A frontier without an ecosystem is not stable. It reached 51 million impressions in under 24 hours.

The real opportunity, he wrote, is not in picking the best model but in building a learning loop on top of models, where human capital and token capital compound together. A company should be able to swap out a generalist model without losing the expertise built into their learning system.

He pushed back on fears that AI erodes human value: “Human capital does not become less valuable as token capital grows. It only becomes more valuable.”

It’s a compelling vision. It’s also worth noting the context in which it arrives: Microsoft is not obviously winning the frontier model race — and has been dependent on OpenAI for its own. If models become commodities and ecosystems become the primary source of value, Microsoft’s position suddenly looks considerably stronger. That doesn’t make Nadella wrong. But it’s worth keeping in mind when evaluating the argument.

There are also harder questions his essay raises without quite answering. The idea that every organisation possesses unique institutional knowledge that AI will amplify sounds reassuring. But a surprising amount of what companies call tribal knowledge is scattered across spreadsheets, procedures and industry practices — much of it not particularly unique, and increasingly the kind of thing frontier models are becoming good at absorbing.

And there’s a subtler tension. When companies bring in AI labs to help build workflows, evaluation systems and business processes, the AI company doesn’t just deliver a service. It builds a cognitive loop, and learns how real organisations function in the process. Traditional consultants leave with slide decks. AI platforms leave with something more durable. Whether that becomes a long-term problem is still unclear. But it’s a question worth asking.

And then there’s Fusion (mix models)

Maybe not the biggest story of the week, but it points somewhere real, and it matches how many people are already working (I’ve been using both Codex and Claude Code). OpenRouter launched Fusion — a system that orchestrates multiple models into a single workflow. Early results suggested that carefully combined mixtures of models could match or exceed frontier performance on some tasks, at lower cost.

The significance goes beyond one benchmark. For years the assumption was simple: the smartest model wins. Fusion suggests something different. For many real-world problems, the optimal solution may not be a single model at all. One extracts information, another validates, another reasons through exceptions, another reviews. The value shifts from intelligence itself to the orchestration of intelligence.

The “sovereign AI” question

The Fable 5 episode also makes certain claims feel more fragile than they did a week ago.

Rio 3.5, a model that briefly made the “X/Twitter” news as a supposed Brazilian AI breakthrough, turned out to be a weighted blend of an existing open-source model and Qwen 3.5, with no original training. It was downloaded over 110,000 times before the authors acknowledged the issue, claiming they had uploaded the wrong file. A cautionary tale about how easily “sovereign AI” announcements can outrun the reality.

Mistral, Europe’s most visible AI startup and the company French President Macron has publicly championed as Europe’s answer to OpenAI, is a genuinely impressive company. But the gap between Mistral and the models that just got restricted is significant. When the most capable systems can be switched off by a government directive, being a credible second-tier alternative and being a true frontier competitor are very different things — and second-tier here makes all the difference.

One of the most striking things to surface this week was Europe 2031 (europe2031.ai) — a detailed scenario project by Judith Dada, Daan Juijn and colleagues, months in the making, that maps exactly where Europe’s current AI trajectory leads. It’s told through fictional characters but grounded in real trends: a French policy worker in Brussels watching, year by year, as Europe debates sovereignty while the gap widens. The compute numbers alone are striking, the US is projected to hold a 12x advantage by 2031.

The scenario doesn’t predict catastrophe as inevitable. It argues the window to change course is open, but closing fast. And it was published this week, as the EU Commission was already scrambling to assess the practical consequences of the Fable 5 directive. Several of the things the scenario imagined as future risks (access rationing, inference caps, European exclusion from key model launches) are already happening, ahead of schedule. Worth reading in full.

A harder question about Anthropic

The week also raised something that goes beyond geopolitics: a question about Anthropic itself, and what it believes it’s entitled to decide on everyone else’s behalf.

Ben Thompson at Stratechery published a sharp analysis today that is worth reading in full. His argument centres on a series of decisions Anthropic made around the Fable 5 launch that received less attention than the government directive. Anthropic initially implemented hidden safeguards that would silently degrade Fable’s performance if used for frontier LLM development, without disclosing this to users. They later walked this back, replacing it with a visible handoff to a less capable model. Separately, the company changed its data retention policy at launch, announcing it would retain all user data for 30 days — even for enterprise customers who previously had zero data retention guarantees — without clarifying whether that data would eventually be used for training. This isn’t a theoretical concern: companies like Figma have seen how quickly Anthropic can move into adjacent territory, sometimes without warning, even after early partnerships.

Thompson’s conclusion from these decisions is direct:

“Anthropic believes that they are the ones who should have final say over how Anthropic is used; given that they think only they should be developing leading edge AI, they by extension think that only they should have final say over AI generally. When you further combine this realization with the company’s pronouncements about AI’s ability to conduct all economic activity, you realize that Anthropic’s leadership effectively wants to have power over everything and everyone.”

He adds something that cuts both ways:

“I respect this alignment, and I fear it. The closest analogy is probably Apple, which has always framed every self-serving action in the guise of doing right by users — and often they were. So it is with Anthropic.”

I’d push back slightly on the Apple comparison. Apple’s approach was never quite this openly contradictory, never this reliant on fear as a marketing lever, and never this quick to reverse its own stated positions within the same launch cycle. The pattern with Anthropic feels different, and gives a bit more goosebumps (not the good kind).

Dax, the founder of OpenCode, endorsed Thompson’s analysis on X this week. These aren’t people hostile to AI. They’re people embedded in the ecosystem, paying close attention.

None of this makes Anthropic’s models less impressive, or the government’s directive less blunt. But it adds a layer to the week’s events worth sitting with. The question of who (and how) controls frontier AI isn’t only about governments and export controls. It’s also about the companies themselves.

What connects all of this

Taken together, these events point to the same underlying question: where does the real value in AI actually sit, and who gets to decide?

Anthropic’s Fable 5 episode shows frontier models matter enormously, and that access to them can vanish overnight. Nadella argues the learning loop around those models matters even more. Fusion suggests much of the future value may sit in the layer that coordinates models rather than in any individual model itself. And Thompson’s piece reminds us that the companies building those models have their own views about who should be in control (views they are increasingly willing to act on).

The restrictions on Fable 5 and Mythos 5 have reignited debate about open-source models, sovereign infrastructure and locally controlled AI systems, not as ideological positions, but as practical hedges. An escape hatch only matters when something goes wrong. This week, something went wrong.

The more important race may not be about building the smartest model. It may be about control: of the models, the workflows, the data, the orchestration layer, and ultimately the ability to switch when circumstances change.

Last week made that race considerably more visible.

That feeling is also why I keep thinking about a recent episode of The Atlantic’s podcast, The Most Interesting Thing in AI. Nicholas Thompson interviewed Paul Ford: programmer for 30 years, entrepreneur, and author of “What Is Code”, probably the most famous magazine piece ever written about programming (or so it’s recently told). Ford described the moment he understood his industry was changing. He was vibe coding on a plane, watching a system assemble itself “like watching a house get built in five minutes”, and walked off to tell his business partner: “Death is coming.”

Then he arrived somewhere more interesting. “The code was never truly the hard part. The hard part was pushing things forward, coming up with new and novel ideas, communicating with consumers, building really good products that people wanted to use all day.”

(Dax Raad, co-founder of OpenCode, usually it’s nuanced and serious on these topics and was spot on with this tweet)

Code used to be expensive, so we assumed it was the valuable part. Ford puts numbers on it. A custom system he would have charged $400,000 for five years ago, with a full team and a four-to-six-month timeline, is now a one-week prototype at roughly an eighth of the cost. When the expensive layer collapses, you see what was underneath it the whole time. From what I’ve been building, and watching others build, the value was hiding in three places.

Choosing

When software cost six months and a team, the price forced you to think before committing. Now any idea can become a prototype by lunchtime, and the prototype is convincing: the interface exists, the charts animate, it feels like progress. Sometimes it is progress. Sometimes it’s a very attractive distraction. I wrote in May that creating got easier and meaning got harder, and this is where it bites first. When you can work on anything, choosing what deserves your weeks becomes the first real act of judgment. The ability to make something proves nothing about whether it was worth making.

Making the actual thing

Ford runs a software company, and even he admits this part stayed slow. He spent two months and hundreds of dollars in tokens on one unusual product, talking with the bot every day. “The code can happen really, really quickly, but the actual making a thing is still really, really hard.” The back office is still the back office: does it work as intended, does anyone return after the first visit, what breaks when real users arrive. Answering that takes iteration, time, and more perspectives than your own.

That last part is the easiest to forget, because an agent makes you feel self-sufficient from day one. Ford has a great phrase for what happens when everyone suddenly gets everyone else’s skills: “a festival of vengeance”. Product managers ship apps, designers build platforms, engineers skip the design review. His conclusion lands on the opposite side: “I think we’re actually all going to have to work together again in order to build good product.” You can do almost everything alone now. You usually shouldn’t, and convincing yourself of that is part of the work.

Distribution, and knowing what you have

The third place is the one I’d underline twice. The Internet is filling with AI-built products, newsletters, videos and agents, and attention is not growing to match. Having something to show is no longer unusual; having someone care is.

My favorite story in the episode is about this. Around 2015, Thompson, then at The New Yorker, suggested building an editing tool from the magazine’s archive of drafts and edits. Grammarly, but distilled from the best editing culture in publishing. It was too expensive, and they passed. Today the code would be a week of prompting, and what made the idea valuable was never the software: “the expensive part is that database and that knowledge and that little moat of knowledge that the New Yorker has about how to edit”, as Ford put it.

Even a small personal project has a moat: the reason you built it, the thing you know that others don’t, the audience you understand. Distribution starts with being able to say that clearly. What is this, who is it for, why should anyone give it five minutes. AI can draft the posts and cut the videos, but the intent has to come from you. As I wrote in May: AI can help produce the work, but humans still need to provide the reason for the work to exist.

The slow part

None of this is an argument against the tools. I use them every day, and they keep getting better. But the honest summary of this moment is the one Ford got to: the demo now takes an hour, the product still takes months, and an audience still takes years. He spent three decades thinking code was his trade and found out the trade was judgment all along. I keep arriving at the same place through different doors: humans are still the point.

^{Some references}

The Economist: Choosing where to focus is among the most important skills

Yann LeCun also believes this: Anthropic’s War on Opensource AI

I wrote about it in a meaning perspective before: Creating got easier. Meaning got harder