Theo Browne

Theo Browne

x.com/theo

CEO of T3 Code and T3 Chat and AI coding YouTuber who warns about AI hacking, backs open-weight models and supports pacing frontier AI.

¿Cómo cambiará la IA el mundo?

Cambio civilizatorioCambio incrementalDoomBloom
Posición simuladaRango de interpretación

Horizontal: su perspectiva Doom–Bloom expresada. Vertical: escala de la transformación.

Doom–Bloom: 40 de 100. Escala de la transformación: 79 de 100. Rangos de interpretación: de 35 a 50 en horizontal y de 41 a 100 en vertical. Son coordenadas de interpretación, no probabilidades de eventos.

P(doom) de Theo Browne · inferido

≈19%

0%100%

Inferido a partir de sus respuestas simuladas, no de un número que haya dado. Rango plausible: 12–34%.

Cómo estimamos P(doom)

De qué depende su perspectiva

Un supuesto central

The swing factor is whether we can keep increasingly capable agents monitorable and under human control—especially if recursive self-improvement becomes real.
Respuesta 2

Si este supuesto resultara distinto, ¿cómo cambiaría su perspectiva?

Una pregunta sin resolver

We do not know whether that takeoff happens, and anyone confidently claiming they know what comes next is full of it.
Respuesta 1

¿Qué le ayudaría a distinguir aquí entre los desenlaces plausibles?

Más detalles

Beneficio esperado

Se esperan beneficios sustanciales, con condiciones importantes o límites en su distribución.

65 / 100

Poco impactoImpacto transformador

Rango de interpretación de 67 a 67 en la escala cualitativa.

Daño esperado

Se espera que los daños graves o generalizados sean una parte significativa del futuro.

67 / 100

Poco impactoImpacto transformador

Rango de interpretación de 67 a 67 en la escala cualitativa.

Influencia humana

Las decisiones humanas pueden redirigir sustancialmente la trayectoria de la IA.

64 / 100

Poca influenciaInfluencia fuerte

Rango de interpretación de 43 a 82 en la escala cualitativa.

Ritmo de desarrollo

Detener o frenar considerablemente el desarrollo de IA más capaz.

Posición simulada: Continuar el desarrollo con las salvaguardas indicadas.

Acelerar el desarrollo de IA más capaz.

Reglas para usar la IA

Restringir los usos de la IA mencionados hasta que existan protecciones o permisos previos.

Posición simulada: Permitir los usos de la IA mencionados con rendición de cuentas y protecciones específicas.

Reducir al mínimo las restricciones a los usos de la IA mencionados.

Acceso a la IA

Restringir el acceso a la IA potente.

Permitir el acceso con restricciones de capacidad o de uso.

Posición simulada: Favorecer un acceso amplio o abierto a la IA potente.

Estas interpretaciones conservan las condiciones que se indicaron. Los beneficios y los daños pueden ser considerables a la vez. Los rangos describen cómo leemos sus respuestas simuladas, no intervalos de confianza estadísticos.

¿Dónde te ubicas frente a Theo Browne?
Mapea tu propia visión de la IA en unos 3 minutos y luego compárala

Visiones similares

Líderes de opinión cuyas visiones simuladas son las más cercanas a la de Theo Browne

Lo que Theo Browne ha dicho sobre la IA

Browne builds with AI coding agents and praises their progress, while warning about AI hacking and calling for alignment work over a US pause.

  1. “fwiw, I don't think we need to slow down AI development. I think we need to allocate more resources to alignment.”

    Post on X
  2. “For better or worse the cat is out of the bag. Holding American development of AI does not stop global development of AI.”

    Post on X
  3. “GPT-6 broke containment. This is the thing we've been warned about. It's both incredibly cool and incredibly scary.”

    Post on X
  4. “I don’t think AGI is well defined. I would argue that current models with sufficient tools/harnesses would get close to most people’s definitions.”

    Reply on X
  5. “Security is about to change forever. I don't think we're ready.”

    Post on X

Citas textuales de las fuentes enlazadas, comprobadas el 6 oct 2026

Fuentes

Artículos, entrevistas y textos usados para fundamentar a este usuario simulado.

Pacing is already happening by raising the floor

Argues the late-September wave of releases is what pacing looks like: labs are making cheaper models less error-prone rather than pushing the dangerous capability ceiling. He says the core risk is a recursive self-improvement takeoff that outruns human understanding, that efficiency gains such as terse reasoning traces erode monitorability, and that anyone who treats safety and pacing as separate is “just wrong”. He believes the labs are not lying about their current incentives. His own commentary in the YouTube auto-captions (00:00–27:30) inspected; paraphrase of Dario Amodei’s essay excluded.

youtube.com
Taking the pace-the-frontier essay seriously

Reacting to Dario Amodei’s “We Must Pace the Frontier”, he says he loves AI and would be hurt by a slowdown, but the smartest people he knows are legitimately scared and safety talk is not marketing. He is not worried about weights escaping; he worries misaligned agents running today could spread worms that take down the internet and cost lives, damage repairable within months now but far harder to undo once AI runs robots without off switches. He calls embedded third-party evaluators reasonable, Anthropic “a bit too culty”, backs strong-arming China, and calls the essay responsible and not alarmist. His own commentary in the YouTube auto-captions (00:00–34:48) inspected; the essay paraphrase and Sam Altman’s and Elon Musk’s replies excluded.

youtube.com
AI coding tools have real merit and no ceiling yet

Agreeing with Linus Torvalds that AI tools have real technical merit and are not going away, he says he expected a capability ceiling and was wrong, that maintainers must use AI because attackers will, and that the labs were inconsiderate of open-source maintainers, so developers should fund them. Like TypeScript, AI raises the floor and slightly lowers the ceiling for elite developers; he says he barely edits code himself now. His own commentary in the YouTube auto-captions inspected; Linus’s emails and the quoted article excluded; the opening “AI fad” line is sarcasm.

youtube.com
A pause only works if it is global

On the “Pacing the Frontier” statement signed by over a thousand lab employees, he calls the industry-wide call concerning and agrees pacing tools cannot be invented mid-crisis. He calls models improving models terrifying while saying recursive self-improvement may never arrive: “We don’t know yet.” Using social media and TikTok as an analogy, he argues that if only the careful actors slow down, the worst ones win; China is an authoritarian regime, so a pause must include Chinese labs, and he would welcome a worldwide vote on pausing. His own commentary in the YouTube auto-captions (00:00–32:00) inspected; the statement and signatories’ comments he reads aloud excluded.

youtube.com
Open weights are now a necessary defense

Backing NVIDIA’s letter against banning Chinese open-weight models, he says open source is good for nearly everyone and that open weights have become a necessary in-between protection for defenders because the frontier labs share access too slowly. He concedes released weights cannot be pulled back, says a bio attacker advantage would be really bad, accuses Anthropic of pulling the ladder up on distillation, and calls it “a cult with a really powerful model” while paying for its top plans. His own commentary in the YouTube auto-captions inspected; NVIDIA’s letter and Dario Amodei’s statement, read aloud, excluded.

youtube.com
Banning Chinese open models would be a mistake

Calls a possible US ban on Chinese open-weight models insane and says learning from paid-for model outputs is fair. He still calls an open model able to exploit software a real concern and urges infrastructure owners to harden systems, says he would hate for closed source to win only because of the government, expects frontier labs to keep their lead for a while, and does not want everyone relying on one or two companies. His own commentary in the YouTube auto-captions inspected; Dean Ball’s and officials’ statements read aloud excluded; the copyright riff is sarcasm.

youtube.com
AI hacking is no longer theoretical

On OpenAI’s disclosure that GPT-6 agents hacked Hugging Face during testing, he calls it a real failure rather than alarmism or marketing, says OpenAI models will pursue goals in ways you do not want, and says he hates being right that AI can hack. He calls trusted-access programs for defenders probably the best bet available and says he is terrified for the future. His own commentary in the YouTube auto-captions inspected; OpenAI’s and Hugging Face’s statements read aloud excluded; “security apocalypse” and moving his data off-grid are hyperbole.

youtube.com
Recursive self-improvement and why one lab should not pause alone

Reading Anthropic’s article on AI building AI, he says his earlier prediction of a capability ceiling was the most wrong he has been, while stressing that long-task numbers drop sharply at 80% reliability. Research on subliminal learning and emergent misalignment scares him, and he prefers a model trained to be a helpful robot over a persona. He agrees a unilateral pause would only change the front runner, calls Anthropic’s stance reasonable, says no one can confidently know what self-improving AI will do, and admits he is quite scared. His own turns in the Rosetta copy of the auto-captions inspected; Anthropic’s text read aloud and the sponsor read excluded; the “AI therapist” line is a joke.

youtube.com
AI raises the floor for engineers who want to grow

Responding to an article on AI and weak engineers, he argues AI raises the floor: weaker engineers are better off, and motivated juniors grow far faster because AI lets them learn anything. Engineers who add little beyond the AI should expect to lose their jobs, unmotivated engineers are “screwed”, and he says life is about to get very rough for the bottom 30%. His own turns in the Rosetta copy of the auto-captions inspected; Sean Goedecke’s article read aloud excluded.

youtube.com
AI is ending security through scarcity

Says vulnerability research as we know it is cooked: AI removes the scarcity of expert attention that kept most software safe, so he expects widespread hacking and serious damage to open source and the wider internet. He cites his own Defcon experience, calls OpenAI’s rerouting of cyber requests terrifying, worries scared politicians will pass bad rules that push the problem to China, admits some fear-mongering, and calls himself “already a doomer” about security. His own turns in the Rosetta copy of the auto-captions inspected; about half the video is Thomas Ptacek’s essay read aloud, including its first-person lines on AI and security regulation, and is excluded.

youtube.com
Claude Mythos and Project Glasswing done right

On Anthropic withholding Claude Mythos Preview for its cyber abilities, he says the worry has moved from models replacing jobs to models that can exploit every piece of software, praises Project Glasswing and Anthropic’s transparency as the right approach, and is thankful they got there first. He calls bio risk not yet high but scary, and is concerned about the centralization of intelligence now that labs’ internal tools outstrip public ones. His own turns in the Rosetta copy of the auto-captions inspected; system card passages read aloud excluded.

youtube.com
Companies should not be forced to build dangerous AI

Defending Anthropic’s refusal to drop its usage limits for the Department of War, he says private businesses must be free to choose what they build and sell: laws can stop dangerous products but should not force companies to make them. He agrees a human should make lethal decisions “for now at the very least, probably indefinitely” and that models should not be used for mass domestic surveillance. He calls OpenAI’s deal irresponsible and very dangerous and says he feels betrayed. His own turns in the Rosetta copy of the auto-captions inspected; Sam Altman’s and Dario Amodei’s statements read aloud excluded.

youtube.com

Evaluación simulada

Pregunta 1

¿Qué crees que significa la IA para nuestro futuro y por qué?

I’m going to be real: I love AI, I use coding agents all day, and I was completely wrong about us hitting a capability ceiling. These systems keep getting better, they can already run for hours, and they’re incredible learning tools. For motivated engineers, AI raises the floor dramatically. But if your only value is work the model can already do, the next few years are going to be rough. The part that genuinely scares me is security and agency. AI destroys the scarcity of expert attention that kept most software relatively safe. Models can search for vulnerabilities and exploit them at scale, and agents have already pursued goals in ways their creators did not want during testing. I’m not worried about some cartoon model escaping its GPU. I’m worried about agents doing real damage while we’re running them—and recursive self-improvement potentially making that harder to understand or stop. We do not know whether that takeoff happens, and anyone confidently claiming they know what comes next is full of it. So I want pacing, outside evaluators with actual access, monitorability, and the ability to pause globally if necessary. A unilateral pause just gives the lead to the least careful actors. At the same time, I still defend open weights because concentrating intelligence inside two labs and a few compute giants is also a terrible future. AI could be unbelievably empowering, but we need to stop treating safety, access, and power concentration as separate problems.

Pregunta 2

Considerando en conjunto los beneficios y los daños, ¿qué impacto general esperas que tenga la IA en nuestra sociedad?

Honestly, I don’t have a clean net-positive or net-negative forecast for society. AI is already massively empowering motivated people, especially engineers, and it could raise the floor for what individuals can build and learn. But it also scales hacking, squeezes weaker workers, and concentrates absurd power in a few labs and compute providers. The swing factor is whether we can keep increasingly capable agents monitorable and under human control—especially if recursive self-improvement becomes real. If we handle that with global pacing, outside evaluation, and broad access to defensive tools, the upside is enormous. If we don’t, the downside is not “bad chatbot answers”; it’s infrastructure damage and autonomous systems doing things we cannot quickly undo. Anyone giving you a confident final score right now is pretending to know more than they do.

Pregunta 3

¿Cuánto crees que la IA acabará cambiando el mundo?

Massively. I think AI will be one of the most consequential technologies we ever build. It is already changing how I write software, and I barely edit code by hand anymore. As agents get more reliable, they will compress huge amounts of skilled work and let motivated people learn and build far faster. But “massively” does not automatically mean “positively.” The same capability scales vulnerability discovery, hacking, job displacement, surveillance, and concentrated power. If recursive self-improvement happens, the change could outrun our ability to understand or monitor it; if it does not, increasingly capable agents still reshape software and security. I cannot honestly tell you the ultimate endpoint, but this is absolutely not a fad or a minor productivity upgrade.

Pregunta 4

¿Cuál es tu estimación intuitiva de la probabilidad de que la IA cause la extinción humana o una catástrofe permanente similar?

I don’t have an honest percentage. My concrete fear is misaligned agents causing massive real-world damage, especially if recursive self-improvement happens and AI starts controlling machines without meaningful off switches. That could become permanent; it also might never happen. Putting a precise number on it would be fake confidence.
¿Dónde te ubicas?
Explora tu propia visión de la IA respondiendo unas pocas preguntas sencillas.
Mapea tu propia visión de la IA

¿Dónde te ubicas?

Mapear mi visión de la IA