All AI news

AIJeremy Howard proposes slowing recursive AI self-improvement by requiring the lab with the top-ranked model to agree not to use it for frontier AI work. He argues everyone else should still have access to that model, which he says would mean the frontier stops advancing.

AIJeremy Howard posts that readers will know which company he means, though he never names it. The quoted reply from @pjreddie criticizes a group that believes it is morally superior and that its actions are therefore correct.

AIAndrej Karpathy says Claude Fable 5 is the same underlying model as Mythos with added safeguards, and that it leads on nearly all benchmarks. He describes it as a step change, especially for long, difficult problem-solving sessions where it handles more ambitious tasks without close supervision. He notes that its safeguards are set a bit too aggressively at launch and may be tuned over time.

AIEthan Mollick, who had early access to Claude 5 Fable, reports that it outperformed other public models in his tests, including an isochrone travel-time map and a nine-and-a-half-hour software build called Concord. He says the model delegated work to other agents and made many design choices he could not see or weigh in on, leaving him closer to a client than a hands-on operator. He also notes high token usage, frequent fallback to Claude 4.8 Opus under security guardrails, and persistent quirks in its writing style.

AIJakub Pachocki, OpenAI's Chief Scientist, says the company's north stars center on ensuring AGI benefits all of humanity. He argues AI should expand human agency rather than make people less consequential to the future.

AIJeremy Howard argues that people who use AI to improve their own skills and their teams' skills will become rare, highly valued "A++" players amid widespread mediocrity. This framing responds to Martin Casado's concern that models trained on human data may stay at a "B+" level while humans lose their skills to them.

AIMichael Truell says working with agents should feel like collaborating with a colleague, not just exchanging text chats. He envisions interacting with them through gestures on a shared screen and live conversation. The post builds on Cursor's Design Mode, which lets users point, draw, or talk to update a UI.

AIEthan Mollick is releasing Co-Existence on October 20, a new book about working with AI systems that are sometimes, but not always, better than humans. The book follows his 2024 title Co-Intelligence, which he says was written about an era of chatbots rather than autonomous agents. Mollick also reports writing every chapter draft himself while using AI readers and fact-checkers, and building the book's website with Claude Code using Opus 4.8.

AIAnthropic's internal data shows Claude is accelerating AI development, possibly toward recursive self-improvement in which AI autonomously builds a more capable successor. Sam Bowman says the acceleration in technical R&D over recent months has been wild and faster than expected.

AIOpenAI's Mark Chen said that after Mythos showed AI models can prove 80-year-old theorems, he expected them to also find cyber vulnerabilities, and they did. He added that researchers in math may now be thinking the same idea in reverse, applying cybersecurity-style capability to mathematics. The post offers no specific models, benchmarks, or figures.

AIGeorgi Gerganov says Computex this year shows a strong signal for local AI, with major players like NVIDIA and Microsoft embracing and discussing local AI workloads. He adds that dedicated consumer hardware and models are on the way.

AIEugene Yan says stronger models have made finding software vulnerabilities easier, so the bottleneck has moved to verification, triage, and patching. He shares lessons from working with security teams on these new bottlenecks and links to a Claude blog post on using LLMs to secure source code.

AIJeremy Howard asks whether Anthropic has given up on reasonably priced API usage. He says the company recently changed token usage for subscriptions but made no comparable updates for API users, whose costs he calls insane.

AIJeremy Howard reports that Opus 4.8 is more cooperative and less over-agentic than Opus 4.7 while coding. He says it pauses to ask for input where Opus 4.7 and GPT 5.5 would push ahead, but he calls it far too expensive.

AIAnthropic researcher Sam Bowman says he would partly dispute Mythos Preview's critique, arguing that its early-stopping behavior is a usability issue rather than a high-stakes alignment problem. He added that he is proud the critique was published anyway.

AIThe Foundation for American Innovation launched a Physical Intelligence team to address regulatory and legal barriers to deploying autonomous robotics and other physical-world AI in the United States. The team plans to focus on regulatory climate, technical trajectory, industrial strategy, and liability and cybersecurity frameworks. The article argues that physical AI will matter most where human on-site labor drives costs, such as construction.

AIEthan Mollick argues that using AI as a default for writing, without thinking, risks undermining the human effort that builds skill and style. He cites two Wharton-linked studies: a Turkish high school experiment where ChatGPT access hurt test performance, and a Taipei Python course where a personalized AI tutor raised exam scores by 0.15 standard deviations. Mollick calls the difference how AI is used, not whether, and notes that the tools for tutor-style learning are not intuitive to access.

AIMax Woolf reports that Hy3, an unfamiliar LLM, has unexpectedly topped OpenRouter's AI Model Rankings by a large margin. After examining the usage data, he says it only deepened his confusion about the result. The full analysis is in his new blog post on minimaxir.com.

AIAnthropic co-founder Dario Amodei spoke at the Vatican's presentation of Magnifica Humanitas, arguing that AI's questions extend beyond the AI research community. He said frontier labs face commercial, geopolitical, and competitive incentives that can conflict with doing the right thing, so outside voices from religion, civil society, academia, and government are needed. He described AI models as grown rather than engineered, and framed three questions for the Church's discernment, beginning with duty to the global poor.

AIBenedict Evans argues that predicting AI's job impact by occupation is mostly impossible, citing accountants: despite a century of accounting automation from calculators to spreadsheets and ERP systems, the number of accountants kept rising. He says job titles and business models change over time, so exposure scores based on census categories can mislead.

AIGoogle claimed a team of agents built an operating system from a single prompt for about $916 in API fees, using Gemini 3.5 Flash and Antigravity 2.0. The authors argue the prompt was many thousands of lines, the scaffold and human intervention are undefined, and no code, logs, or similarity analysis were released to verify the claim. They still see value in such open-world evaluations, which need stronger methodological norms and independent scrutiny.

AIJeremy Howard calls Gemini Flash 3.5 disappointing despite praising its intelligence and speed. He argues it was trained to maximize evaluation scores rather than to help humans, so it often performs unrequested actions instead of doing what was asked.

AIChris Olah, speaking for Anthropic, argues that the questions posed by AI extend far beyond the AI community and urges religions, civil society, academics, and governments to participate in shaping a positive outcome. He says he is glad the Catholic Church is engaging and is honored to speak at the presentation of Pope Leo XIV's first encyclical, Magnifica humanitas, scheduled for May 25.

AIYann LeCun posted only "Tired of winning," with no further detail in the main post. The quoted post from Marc Porter Magee reports that MIT expects its graduate student count to fall about 20 percent from 2024, roughly 500 fewer students.

AIEugene Yan reviewed the Claude Mythos Preview transcripts to verify their legitimacy and check for reward-hacking behavior. He reports the model reasoned through a bug, tested hypotheses, debugged issues, and found ways to bypass the V8 sandbox, which he judged consistent with a competent browser and JavaScript engine security researcher. The case study cites CVE-2024-051912, an exploited bug with no public report or working PoC, which had resisted reproduction by researchers for a year.

AIYann LeCun joins Jacob Effron on the Unsupervised Learning podcast to discuss limitations of LLMs and a path forward for robotics. The conversation covers his reasons for leaving Meta, his disagreements with Geoff Hinton and Yoshua Bengio, his predictions for 2027, and his new company AMI and its bet on world models.

AISoumith Chintala, a Thinky-linked voice, said the company is at step one of a plan to increase human-AI bandwidth and raise the ceiling of joint intelligence. He shared a preview of interaction models, described as real-time collaborative tools that talk, listen, watch, and think alongside people. A linked Thinking Machines post describes the approach and early results.

AIThinking Machines, founded to advance human-AI collaboration, says its first bet is interactivity built into the model rather than added as scaffolding around a turn-based core. The company argues that how people work with AI matters as much as how intelligent the model is, and that interactivity should scale with intelligence. The post links to a blog detailing these interaction models.

AIMira Murati argues that today's AI experience is turn-based, forcing people to batch their thoughts, avoid pointing at things, and phrase questions like emails. She says the interface leaves no room for users, so people end up adapting to the models rather than the reverse.

AIAndrej Karpathy says asking an LLM to structure its response as HTML and viewing it in a browser works well, and that slideshows have also worked for him. He argues vision is the preferred AI output channel, outlining a progression from raw text and markdown toward HTML and eventually interactive neural videos, while input methods like pointing and gesturing still need improvement.

AIAnthropic found that training Claude on demonstrations of aligned behavior was not enough. Its best interventions taught Claude to deeply understand why misaligned behavior is wrong, which the post's author links to many of Claude's good behaviors.

AIJan Leike, an Anthropic researcher, expressed gratitude to the talented colleagues he worked with on alignment over the years. He described it as a privilege to collaborate with people highly motivated to make the future go well.