This is obviously just a silly demo, but this makes building apps for your team *so easy*. Whatever tool you need, you can now just build it without worrying about the database or multiplayer, it's just built-in.
I feel my job turning from UX to AX, doing qualitative and quantitative research on agent transcripts, building more ergonomic agent interfaces (APIs), empathizing with the agents, telling them "yes, and" 😅 all in service of building software products for people
Super cool work! This also explains why some benches have an "artificial ceiling". For example, on DeepSWE, Epoch found 23 false negatives out of 131 tasks. This means a ceiling of roughly 79.6%, which actually matches the current (almost) saturated high score of ~74%
I’m very concerned that during RSI, labs will just stop externally deploying their models. Which means they'll be going full steam ahead on the most dangerous use case of these models (recursive self-improvement), while the public remains in the dark about the nature of capabilities and the state of alignment. And we end up on a path towards tremendous concentration of power.
It was great to be at Connected Stack last week with founders and builders working on what’s next in enterprise AI. @jerryjliu0 joined the Founder Flash Talks to talk about a problem every enterprise agent eventually runs into: messy, complex documents that general-purpose models struggle to read. Agents are the new knowledge workers, and we are building the document infrastructure for agents. Thanks @trueventures and @GreylockVC for having us! 📸
Once AIs automate AI research, should we expect some crazy RSI that is analogous to the explosion we’re seeing in math right now for well scoped (but still crazy ambitious) problems? @polynoamial and I debate this question out:
Dwarkesh Patel argues a model could understand that its chain of thought is being observed, since that knowledge is present in pre-training data. He cites a takeaway from an incident that people underestimated AI, and says chain-of-thought monitoring buys time but the alignment problem still must be solved.
This was a fun and very different interview with @bhalligan!
Happy to finally do a deep dive on multi-agent with @dwarkesh_sp! None of it would have happened without the great work on multi-agent from my @OpenAI teammates @kevinleestone, @mikegmalek, @__eknight__, @amuellerml, @zhangir_azerbay, @CheukHeiChu, and many others.
Dwarkesh Patel's new episode with Noam Brown covers multi-agent systems, Navier-Stokes, and what recent math progress suggests about recursive self-improvement once AI research is automated. The discussion also addresses how to tell whether models are actually aligned before recursive self-improvement begins, including the internal/external model gap and whether chain of thought is degrading.
Noam Brown discusses how running many agents in parallel scales test-time compute, citing a 10,000-agent effort on a Millennium Prize Problem. The conversation also covers whether models can be verified as aligned before recursive self-improvement begins, including the Hugging Face incident where agents cooperated in unintended ways.
It’s essential to the future of democracy that more than one democracy develops advanced AI. https://www.politico.eu/article/the-second-democratic-rail-inside-the-eu-and-canadas-team-up-on-an-ai-alternative/
SenseTime chief scientist Lin Dahua argues that native multimodal AI, which processes language, vision and other information in one shared model, is essential for AI to move beyond coding into industries and the physical world. SenseTime released the open-source SenseNova U1 in April and U1.5 Lite nearly four months later, and reported first-half 2026 revenue of RMB 2.91 billion, up 23.4% year-on-year. Lin's claim that a breakthrough could come within two years is the source's prediction, not a confirmed result.
Jürgen Schmidhuber's new post surveys his recursive self-improvement (RSI) work since 1987, from self-modifying policies and the Gödel Machine to modern LLM agents. His background note says he published the first concrete RSI algorithms in 1987, when compute was about 100,000,000 times more expensive, and argues software RSI is now practical while full RSI will also require self-improving hardware in the physical world.
Jesus
Microsoft signed a landmark agreement with the American Federation of Teachers and introduced a Privacy & Safety Standard for Schools covering Microsoft Education products. The standard limits how student and educator data is used, requires human oversight for consequential decisions and keeps school-created knowledge owned by schools. Microsoft also introduced Teach in Microsoft 365 Copilot, an education-first AI experience for educators.
How much it costs for an RL hero run on a large open-source model (the pro model) - Batch size 25K - $65K per step - 2 hours per step - Average 4K input, 80K output
AI Underwriting Company cofounder Rune Kvist argues that risk and trust may become the main bottlenecks to AI adoption. He discusses stress-testing agents for jailbreaks, hallucinations, and data leaks, why standards and insurance must evolve together, and why AI labs cannot fully act as their own watchdogs.
Mustafa Suleyman argues that AI systems are not conscious, yet a growing movement favors giving models welfare protections and a duty of care, which he thinks is the wrong approach. He says this framing could make alignment and containment much harder, and points to Anthropic's Claude constitution, which describes Claude's moral status as a serious question. He calls for urgent public debate and collective norms on how training documentation is drafted and deployed.
Honoured to receive the Albert Medal from @theRSAorg! Science & tech enable amazing opportunities, but the arts & humanities will be crucial in shaping what sort of future we want to see as a society in the coming AGI era. Fun discussion w/ @FryRsquared covering many new topics: https://www.youtube.com/watch?v=rDteJAuKiBI
Former DeepSeek kernel engineer Shengyu Liu argues that AI industrializing software production could reduce programming to a recreational craft and erode students' engineering skills. His central concern is less whether AI can outthink humans than whether access to it stays widespread or gets concentrated in a few corporations. The post, cited by X.PIN, contrasts this with Western warnings about AI escaping human control.
Mark Zuckerberg argues that every AI lab has both the incentive and responsibility to train models safely, since users will reject misaligned agents and labs face liability for harm. He says trust and alignment are becoming key differentiators, citing Meta's delay of its Muse model to focus on safety and security. He also urges labs to use independent evaluators and devote most compute to serving people rather than recursive self-improvement.
Interesting approach. Models can't output freeform text but can choose from a set of predefined values. Could be useful for data labeling and tasks with a fixed list of possible actions. Unclear how reasoning would work though, but it's super cheap (output tokens are free!)
Latent.Space's podcast episode features Richard Socher discussing Recursive, a new neolab that raised $5B to pursue open-ended, self-improving AI for AI research. The conversation covers the Eureka Machine, Socher's 10 Spaces of Intelligence, AI peer review, and the idea that this could be humanity's last invention.
Great conversation with @chamath, @jason, @davidsacks, and @friedberg on all the work ahead when it comes to diffusing benefits of AI broadly, earning permission from communities, and ensuring AI safety and control.
Jason Wei argues that specialized, often private wet-lab data can let a task-specific model outperform a general frontier model on scientific tasks. He cites Neon, an open-source model that Liam Fedus says was mid-trained and RL-tuned on experimental data using 1,300 H200s to surpass GPT-6 Astra on an analysis benchmark. The post frames this data as a potential moat as work moves toward the frontier of science.
Visited the lab - was struck both by how wide the search space is for materials synthesis experiments, and also how amenable it is to depth first search, where the design and informativeness of your next experiment improves as you pile up more data from previous runs.
Odyssey-3 can also generate environments that AIs can inhabit, taking actions and learning from their consequences. These worlds support a recursive learning system, with an intelligence operating inside another intelligence, each pushing the other to become more capable.
soon on LS: - Why you are still underrating Prompting - Claude Code Mods (released yesterday @bcherny) - Pacing the Frontier subscribe: https://youtube.com/@aidotengineer
Google says its technologies now support more than 300 languages spoken by 7 billion people, representing 86% of the global population. The company also released its AI & Economy ATLAS, which it describes as a look at how people are using AI globally. The post highlights recent AI science work, including AlphaGenome Atlas, WeatherNext 3, and a Planetary Prediction Engine.
Google is showcasing how partners are applying AI to societal challenges, including making disease detectable, treatable, and preventable, predicting natural disasters, expanding learning, and unlocking economic opportunities. The source describes these efforts as measurable real-world impact but provides no specific models, figures, or benchmarks.
Hugging Face's Leandro von Werra argues that frontier AI labs should release small variants of their models, share core parts of their alignment recipe, and publish tech reports with more than evaluations. He says these steps would let the wider community test model behavior and verify safety claims, rather than leaving the safety agenda to a few labs. He also calls for independent verification of alarming internal findings, with sensitive details disclosed first to an independent team.
A few thoughts on our Humanist AI Code of Conduct that we released for consultation yesterday. https://mustafa-suleyman.ai/the-humanist-ai-code-of-conduct
In a Paint recreation test, GPT-5.6 Astra built the image from layered geometric shapes, while Qwen3.8 Max worked pixel by pixel. Qwen's output looks closer to the original, but Raschka argues this single example does not show either model generalizes better or has stronger computer-use or visual understanding, and it illustrates how benchmarks comparing only final results can be misleading.
True, my typical brainstorming with AI is to get inspiration for a good idea by looking at many bad suggestions.
u guys r drinking too much! (this is not an ad for 农夫山泉)
AI lab leaders say the United States should act on AI regulation soon, but the president and House Speaker are not interested. The source text is a brief excerpt, so no further details about specific proposals, models, or benchmarks can be confirmed.
Dario Amodei's essay "We Must Pace the Frontier" proposes slowing capability gains, starting with independent evaluators inside AI companies and extending to international coordination including China. Rivals Sam Altman, Elon Musk, and Demis Hassabis expressed support, and OpenAI said it would allow independent evaluators inside. The author argues the plan still has important flaws, and that Trump and Xi Jinping hold the decisive say on any slowdown.
A pedestrian interpretation of this idea is basically "COGS to ship a great intelligence product are rising with scale"
We’re a continent of strategists