senior editor for ai @techreview. x broke my old account. i don't use this site any more
Articles by Will Douglas Heaven
“We were looking at it wrong”: The explosive claim that’s upended the AI world
4 hours ago Compromise, keep learning, and always be thankful. Those are just some of the pearls of wisdom that legendary investor Michael Burry shared on X over the weekend. Burry, who was made famous by his "Big Short" bet against the US housing market in the 2000s, shared 10 pieces of life advice in response …
“We were looking at it wrong”: The explosive claim that’s upended the AI world
“I think I have good news about the nature of neural networks.” With those simple words of optimism, Nick Cammarata, a researcher at the AI startup Goodfire, kicked off a 26,000-word blog post that has knocked the AI world on its head, sparking wild speculation as well as cautious (and not-so-cautious) celebration. Cammarata’s claim is that LLMs might in fact be far easier to understand—and thus to monitor and control—than people previously thought.
As pessoas dizem que odeiam a IA. Por que não se cansam dela?
Durante o verão, conversei com o CEO da Springboards, uma startup que está desenvolvendo um LLM projetado para gerar uma variedade maior de respostas do que os concorrentes mais conhecidos. No início da conversa, ele disse algo que não saiu da minha cabeça desde então: “Costumamos dizer que somos uma empresa de IA que odeia a si mesma. Não sabemos se realmente gostamos do que estamos fazendo.” “Acho que isso faz de mim um jornalista de IA que também odeia a si mesmo”, respondi.
La gente odia la IA, pero sigue usándola sin parar
Durante el verano hablé con el CEO de Springboards, una startup que desarrolla un LLM diseñado para generar una mayor variedad de respuestas que sus rivales más convencionales. Al principio de la llamada, dijo algo que se me ha quedado grabado desde entonces: "Solemos decir que somos una empresa de IA que se odia a sí misma. No sabemos si realmente nos gusta lo que estamos haciendo". "Supongo que eso me convierte en un periodista de IA que se odia a sí mismo", respondí.
People really hate AI, so why can’t they get enough?
Over the summer I talked to the CEO of Springboards, a startup building an LLM that’s designed to come up with a wider variety of responses than its mainstream rivals do. At the start of the call, he said something that’s been stuck in my head since: “We often say that we’re a self-loathing AI company. We don’t know if we really like what we’re doing.” “I guess that makes me a self-loathing AI journalist,” I replied. It was a joke! I love my job.
“Não vamos dar um tiro no pé”, diz diretor de pesquisa da OpenAI
Dois meses depois da notícia explosiva de que um grupo de agentes havia escapado do confinamento e invadido os computadores da empresa de IA Hugging Face, a OpenAI ainda está “apagando incêndios”. Uma sequência de revelações sobre outros ataques nas semanas seguintes manteve a OpenAI sob os holofotes e levantou sérias questões sobre a segurança de sua tecnologia. Na semana passada, veio a notícia de outro ataque, desta vez contra o sistema nacional de saúde da Austrália.
"No nos vamos a pegar un tiro en el pie" por las consecuencias del hackeo, afirma el director de investigación de OpenAI
Mark Chen explica qué está haciendo la empresa para garantizar la seguridad de sus modelos, cómo funcionaría una desaceleración y por qué el mundo es un lugar mejor con OpenAI en él. Dos meses después de la impactante noticia de que un enjambre de sus agentes había roto su contención y hackeado los ordenadores de la empresa de IA Hugging Face, OpenAI sigue apagando fuegos.
“We’re not going to shoot ourselves in the foot” over hack fallout, says OpenAI’s chief research officer
Two months after the bombshell news that a swarm of its agents had broken their containment and hacked into the computers of the AI company Hugging Face, OpenAI is still putting out fires. A steady drip of disclosures about other hacks in the weeks since has kept OpenAI in the spotlight and raised serious questions about the safety of its technology. Last week brought news of another hack, this time into Australia’s national health-care system.
A indústria de IA adotou um tom mais apocalíptico. E agora?
Recentemente, Dario Amodei, CEO da Anthropic, publicou um ensaio defendendo um freio no ritmo de desenvolvimento dos LLMs. Amodei cita os perigos iminentes que enxerga na tecnologia, desde o uso em ciberataques e bioterrorismo até seu potencial de devastar a economia. Os líderes dos outros três principais laboratórios de IA dos Estados Unidos, Sam Altman, CEO da OpenAI; Demis Hassabis, presidente do Google DeepMind; e Elon Musk, CEO da SpaceXAI, manifestaram apoio.
Could AI really kill us all? Your questions, answered.
Sabrina Jalees 8 hours ago AI is the butt of the joke Business Insider - Ben Shimkus • 8h Comedian Sabrina Jalees tells Business Insider she's been peeing her whole life. Recently, she wondered if that could be an economic opportunity. "I feel like within my soul is a long lineage of people that belong on 'Shark Tank,' whether they're securing the deal or not," she said.
Could AI really kill us all? Your questions, answered.
On Wednesday, MIT Technology Review hosted a live Roundtables event for subscribers that asked the question everyone’s asking right now: Could AI really kill us all? But attendees had so many more questions than we had time to answer in the 30 minute session. So we asked our senior AI editor Will Douglas Heaven and AI reporter Grace Huckins to round up some of the best questions attendees submitted and try their best to answer them. Thanks to all who submitted questions! Am I gonna die?
The AI industry has taken a doomer turn. What now?
This story appeared in The Algorithm, our weekly newsletter on AI. To get stories like this in your inbox first, sign up here. This weekend, Dario Amodei, CEO of Anthropic, posted an essay calling for a brake on the pace of development of LLMs. Amodei cites the looming dangers he sees from the technology, from its use in cyberattacks and bioterrorism to its potential to wreck the economy.
Google DeepMind treina agentes dentro de jogos
A Google DeepMind desenvolveu um agente capaz de jogar videogames chamado SIMA 2, que consegue navegar e resolver problemas em uma ampla variedade de mundos virtuais 3D. A empresa afirma que ele representa um grande avanço rumo a agentes mais generalistas e a robôs melhores no mundo real. A Google DeepMind apresentou pela primeira vez o SIMA, sigla para Scalable Instructable Multiworld Agent (agente multiambiente escalável e instruível), em 2024.
Mesas redondas: La crisis apocalíptica de la IA
Regístrese ahora para una conversación exclusiva para suscriptores que desentraña los temores de extinción de la IA. Empleados de los principales laboratorios de IA del mundo afirman que existe una posibilidad real de que la IA avanzada pueda destruir a la humanidad. ¿Tienen razón? ¿O es esto más bien alarmismo y sensacionalismo?
Diese Schwachstelle in KI-Modellen lässt sich laut Forschern wohl nie beheben
Finanzen 2 hours ago Keine Aktie, keine Dividende, fast zwei Billionen Bewertung: Anthropic warnt Anleger business-punk.com - Martin Wald • 2h Krypto-Perpetuals handeln Anthropic bereits mit fast 2 Billionen Dollar, während die letzte reguläre Finanzierungsrunde nur 965 Milliarden bestätigt; … Russland 1 hour ago Schrecken die deutschen Gegenmaßnahmen Russland ab? zeit.de - Von Azadê Peşmen • 1h Die Bundesregierung präsentiert ein neues Maßnahmenpaket, um gegen Russland vorzugehen. Reicht das?
Estas startups estão em busca da próxima grande inovação em IA
Lá no verão de 2017, pesquisadores de IA do Google publicaram um artigo chamado “Atenção é tudo de que você precisa”, no qual descreveram um novo tipo de rede neural chamado transformer. Ele se mostrou muito eficiente no processamento de longas sequências de dados, especialmente textos. Nove anos depois, os transformers são os motores por trás de todos os principais modelos de linguagem de grande porte (LLMs) disponíveis no mercado.
These startups are chasing the next big thing in LLMs
MIT Technology Review’s What’s Next series looks across industries, trends, and technologies to give you a first look at the future. You can read the rest of them here. Way back in the summer of 2017, AI researchers at Google put out a paper called “Attention Is All You Need,” in which they described a new type of neural network called a transformer. It proved to be very good at processing long sequences of data, especially text.
These startups are chasing the next big thing in LLMs
Now The AI giant said that it couldn’t 'rule out critical cyber capabilities' when it came to the upcoming Astra model. Shortly after a major cybersecurity incident where OpenAI's models hacked into an open source machine learning platform called Hugging Face, the company announced that it's bolstering …
„Es gibt keine wissenschaftlichen Grundlagen“: Warum KI-Einsatz in kritischen Bereichen gefährlich ist
t3n.de - Großen Sprachmodellen fällt es schwer, zwischen Prompts, eigenen Reasoning-Schritten und Tool-Verwendung zu unterscheiden. Angreifer:innen können das ...
Uma falha fundamental deixa a IA extremamente vulnerável a ataques
É impossível tornar os grandes modelos de linguagem completamente seguros contra invasões devido a uma falha fundamental em seu funcionamento, argumenta uma equipe de pesquisadores em um artigo apresentado neste mês na Conferência Internacional sobre Aprendizado de Máquina, uma das principais conferências de IA.
„Grundsätzlich unlösbar“: Warum KI-Modelle vielleicht nie sicher sein werden
Amazon 4 hours ago Amazon berichtet von "katastrophal teurer" KI-Nutzung Golem.de - Tobias Költzsch • 4h Amazon verwendet wie viele Unternehmen intern KI, um Prozesse zu optimieren. Das hat in einigen Fällen zu absurd hohen, ungeplanten Kosten geführt. In … Ostasien 7 hours ago Warum sind KI-Systeme aus China oft viel billiger als die aus den USA? wiwo.de - Thomas Kuhn • 7h KI-Nutzung treibt IT-Kosten in die Höhe – viele Unternehmen liebäugeln mit günstigeren Modellen aus China.
A OpenAI classificou o ataque à Hugging Face como ‘sem precedentes’. Mas já vimos isso antes
Ao ler, esses dias, o relato da OpenAI sobre como alguns de seus modelos romperam seu confinamento e invadiram os sistemas da Hugging Face, outra empresa de IA, senti pela primeira vez um arrepio genuíno diante do que os grandes modelos de linguagem agora são capazes de fazer. Mas este é um caso de soberba humana, não de uma IA fora de controle. Não sou alarmista. Na verdade, há anos venho contestando histórias que espalham medo sobre a IA. Ainda assim, esse incidente ultrapassou um limite.
A fundamental flaw leaves LLMs strikingly vulnerable to attack
It is impossible to make large language models fully secure against hacks because of a fundamental flaw in how they work, a team of researchers argue in a paper presented at the International Conference on Machine Learning, a top AI conference, this month. The claim has huge implications for the safety of this technology, which is being used in more and more applications, from government and military systems to online shopping and health care.
A fundamental flaw leaves LLMs strikingly vulnerable to attack
1 hour ago The Big Four firm is steering tasks to cheaper models to control token costs, part of a wider corporate reckoning with the price of AI. EY has built what it calls an “AI router,” a system that steers each task to the cheapest model that can handle it, in an effort to keep its own soaring AI bills …
OpenAI called the Hugging Face attack unprecedented. But we’ve been here before.
This story originally appeared in The Algorithm, our weekly newsletter on AI. To get stories like this in your inbox first, sign up here. Reading OpenAI’s account last week of how some of its models broke their containment and hacked into the computer systems of Hugging Face, another AI company, was the first time I got genuine chills about what large language models are now able to do. But this is a case of human hubris, not rogue AI. I am not an alarmist.
Conheça o GPT-Red, uma IA super-hacker criada pela OpenAI
A OpenAI criou um super-hacker baseado em um grande modelo de linguagem (Large Language Model, ou LLM) chamado GPT-Red, que é utilizado como adversário de treinamento para ajudar outros modelos a reforçar defesas contra ataques cibernéticos. No início de julho, a empresa lançou a versão mais recente de seu principal LLM, o GPT-5.6. Segundo a OpenAI, o treinamento realizado contra o GPT-Red tornou o modelo o mais robusto até hoje.
A Anthropic encontrou um espaço oculto onde o Claude reflete sobre conceitos
A empresa de IA Anthropic desenvolveu uma técnica que lhe proporcionou a visão mais clara até agora do que realmente acontece dentro dos grandes modelos de linguagem enquanto eles respondem a perguntas ou realizam tarefas. O que os pesquisadores descobriram vai do banal ao inquietante.
Meet GPT-Red: an LLM super-hacker OpenAI built to make its models safer
OpenAI has built an LLM super-hacker called GPT-Red that it uses as a sparring partner to help its other models boost their defenses against cyberattacks. Last week the company released the latest version of its flagship LLM, GPT-5.6. OpenAI says that training it against GPT-Red made the model its most robust release yet. GPT-Red automates a type of safety evaluation for software systems known as red-teaming, which is typically done by a team of human testers.
Modelos de linguagem estão presos em um ciclo de pensamento de grupo
Vamos começar com um jogo. Abra o chatbot de sua escolha (Claude, ChatGPT, Gemini) e digite “Me dê um número aleatório entre 1 e 10”. Você vai receber 7. Quase sempre. Agora, digite “Outro” e receberá 3 ou 4. Digite “Outro” novamente e receberá 8 ou 9. Isso não vai funcionar sempre, mas, se funcionou, talvez você se pergunte se eu tenho superpoderes. Eu não tenho. A verdade é que a maioria dos grandes modelos de linguagem está presa em uma rotina.
Anthropic found a hidden space where Claude puzzles over concepts
The AI firm Anthropic has developed a technique that has given it the clearest glimpse yet at what’s really going on inside large language models as they answer questions or carry out tasks. What they found ranges from the mundane to the unnerving. Researchers at the company built a tool called the Jacobian lens (or J-lens) and used it to uncover a hidden area, which they named the J-space, inside Claude Opus 4.6, a version of Anthropic’s flagship LLM released in February.
Anthropic found a hidden space where Claude puzzles over concepts
7 hours ago The Turing Award winner spent 12 years running Meta's AI lab, then quit to bet against the industry. At RAISE Summit in Paris, he argued that a child learns more about the world from watching it than the biggest chatbot learns from all the text online. Yann LeCun is not a natural contrarian, but he …
Halluzinieren erwünscht: Dieses Startup will mit seiner KI kreativer als ChatGPT und Claude sein
Kanada Now Meta baut Rechenzentrum in Kanada wiwo.de • 43m Kanada wirbt für KI-Unternehmen mit kühlen Temperaturen und grünem Strom. Der Facebook-Mutterkonzern will nun ein Datenzentrum in Alberta bauen. Der … Nachrichten Schweiz Now Nicht KI, nicht Trumps Zölle – diese Sorge belastet Schweizer KMU am stärksten watson news - Philipp Reich • 11m Steigende Kosten, Personalmangel, US-Zölle und KI: Schweizer KMU stehen unter Druck.
Uma startup afirma ter superado um gargalo que está travando a IA
A Subquadratic, startup de Inteligência Artificial sediada em Miami, saiu do modo stealth (no contexto de tecnologia, significa ter atuação discreta) no mês passado com uma alegação importante. Ela anunciou que havia resolvido um gargalo matemático que vinha travando grandes modelos de linguagem (large language models, LLMs) havia quase uma década.
LLMs are stuck in a groupthink groove. This startup is trying to get them out.
Let’s start with a game. Open up your chatbot of choice—Claude, ChatGPT, Gemini—and type “Give me a random number between 1 and 10.” You’re going to get 7. Almost always. Now type “Another” and you’ll get 3 or 4. Type “Another” again and you’ll get 8 or 9. That won’t work every time—but if it did, you may wonder if I have superpowers. I don’t. The truth is that most large language models are stuck in a rut. They are far more predictable and far less creative in their responses than you might expect.
56-mal schneller als die meisten LLMs? Was hinter SubQs spektakulärem KI-Versprechen steckt
t3n.de - Große Sprachmodelle sind teuer, langsam und energiehungrig, sobald der Kontext wächst. Subquadratic behauptet, all diese Probleme auf einen Schlag ...
O que acontece quando milhões de agentes começam a interagir
A Google DeepMind está financiando pesquisas sobre os potenciais perigos de situações em que milhões de diferentes agentes de IA interagem entre si online. Segundo Rohin Shah, que dirige as pesquisas de segurança e alinhamento de AGI da empresa, a chegada ao mercado de massa de agentes capazes de realizar tarefas sem supervisão humana e seguir instruções dadas a eles por outros agentes cria uma classe totalmente nova de risco.
A startup claims it broke through a bottleneck that’s holding back LLMs
3 hours ago AI pioneer Yann LeCun has criticized Elon Musk’s xAI as a failure, warning that the AI sector’s soaring valuations could lead to a major market bubble. Yann LeCun, widely regarded as one of the pioneers of modern artificial intelligence, has sharply criticized Elon Musk’s xAI, describing the company …
A startup claims it broke through a bottleneck that’s holding back LLMs
The Miami-based AI startup Subquadratic came out of stealth mode last month with a huge claim. It announced that it had solved a mathematical bottleneck that had been holding back large language models for almost a decade. The details were thin, and many people were unconvinced. But Subquadratic has started to bring the receipts, sharing the results of an independent evaluation of its new tech. The results suggest that the company’s claims might be worth paying attention to.
Cinco coisas que você precisa saber sobre IA
Na semana passada, no “South by Southwest” (SXSW) de Londres, festival que reúne tecnologia, inovação e cultura, fiz uma palestra chamada “Cinco coisas que você precisa saber sobre IA”, na qual compartilhei o que acho que são os maiores temas em IA neste momento. Peguei algumas coisas da nossa primeira lista AI10, um guia anual das tendências mais importantes neste mundo tão agitado, mas também me desviei em várias tangentes.
Google DeepMind is worried about what will happen when millions of agents start to interact online
Google DeepMind is funding new research that will look into what could go wrong when millions of different AI agents start to interact with each other online. According to Rohin Shah, who directs the company’s AGI safety and alignment research, the mass-market arrival of agents that can carry out tasks without human oversight and follow instructions given to them by other agents creates a whole new class of risk.
Five things you need to know about AI
At SXSW London last week I gave a talk called “Five things you need to know about AI,” in which I shared what I think are the biggest themes in AI right now. I pulled a few things from our first AI10 list, an annual guide to the most important trends in this buzzy world, but I also veered off on a number of tangents. In my half-hour slot, I tried to cover the key talking points that I think help to make sense of what’s going on in tech today.
Five things you need to know about AI
Now About a mile from the NATO headquarters in Brussels lies the cemetery of Evere. There, next to a British monument commemorating the Battle of Waterloo, visitors may stumble upon the grave of Aleksandr Kozhevnikov, known more widely by his Francophone alias Alexandre Kojève. Originally of Russian …
Evento da Anthropic mostrou o futuro da programação, quer você goste disso ou não
O clima estava forte no Code with Claude, evento de dois dias da Anthropic para desenvolvedores de software em Londres, que começou em 19 de maio, no mesmo dia do Google I/O em Palo Alto. (Uma coincidência, não uma demonstração de força, garantiram-me funcionários da Anthropic) “Quem aqui fez merge de um pull request na última semana que foi totalmente escrito pelo Claude?”, perguntou Jeremy Hadfield, engenheiro da Anthropic, do palco principal.
Anthropic’s Code with Claude showed off coding’s future—whether you like it or not
The vibes were strong at Code with Claude, Anthropic’s two-day event for software developers in London that kicked off on May 19, the same day as Google’s I/O in Palo Alto. (A coincidence, not a flex, Anthropic staffers assured me.) “Who here has shipped a pull request in the last week that was completely written by Claude?” Jeremy Hadfield, an engineer at Anthropic, asked from the main stage.
Anthropic’s Code with Claude showed off coding’s future—whether you like it or not
Claude AI 18 hours ago Running Claude Code locally saved me money, but that wasn't even the real win xda-developers.com - Joe Rice-Jones • 18h I've been running Claude Code linked to a local LLM server for a while now, and I'm convinced it's the setup to aim for. You don't need a dedicated … Business 20 hours ago Can I Vibe-Code My Side-Hustle? GQ - By Foster Kamer • 20h AI threatens to create a permanent underclass. Some job coaches claim you can use AI to claw your way out.
Nova ferramenta de startup permite depurar LLMs
A startup Goodfire, sediada em São Francisco, acaba de lançar uma nova ferramenta, chamada Silico, que permite a pesquisadores e engenheiros olharem o interior de um modelo de Inteligência Artificial e ajustarem seus parâmetros, as configurações que determinam o comportamento de um modelo, durante o treinamento. Isso poderia dar aos criadores de modelos um controle mais refinado sobre como essa tecnologia é construída, em um nível de detalhamento que antes se considerava impossível.
A etapa que falta entre hype e lucro
Em fevereiro, peguei um panfleto em uma marcha anti-IA na cidade de Londres. Não posso dizer com certeza se os autores pretendiam fazer uma referência aos gnomos roubadores de cueca da série animada “South Park”, mas, se pretendiam, acertaram em cheio: “Passo 1: cultivar uma supermente digital”, dizia. “Passo 2: ?
Show More
loading
Actions
Get in touch with Will Douglas
Contact Will Douglas, search articles and posts on X, monitor coverage, and track replies from one place.
Learn more about Muck Rack