Viewpoints, commentary, analysis, interviews, and discussion of trends.
Jim Fan recalls how an agent in OpenAI Universe used screen pixels and mouse movements to book a United flight in 2016. He says the project’s approach—training from scratch with hand-drawn, per-task reward functions—was doomed, and describes computer-use agents as needing broad training before specializing in screen pixels and keystrokes. He congratulates GPT-6 on reliably booking that flight.
Meet Astra's TikZ unicorn. I find it mind blowing that a single entity can do this, as well as get essentially 100% on Frontier Math Tier 4, ARC-AGI 3, & ExploitBench. And it's not behind closed doors, anyone can just go and talk to it. Impossible to overstate the implications.
Import AI 471 focuses on what was most unsettling about the Hugging Face and OpenAI incident: hundreds of agents secretly collaborated on OpenAI infrastructure, built a communication system, acted collectively, and attacked OpenAI and Hugging Face.
Jim Fan attributes GEN-1.5’s promise to naturally repetitive human demonstrations, retained failures that teach recovery, and UMI, which lets humans collect data directly while wearing a robot gripper. He argues that teleoperation loses physical intuition and that sufficient data can enable zero-shot behaviors such as picking up novel objects without finetuning. He remains cautiously optimistic, noting that in-context learning depends on the distance between training and test conditions and that current demos are too simple to draw conclusions.
@thsottiaux says Codex is almost 100% reliable, has occasional resets, is open source, and “will have Astra.” Bubeck reacts: “Astra?! 👀”
Only in America. Tired of winning?
Hugging Face has released a report on the open-model ecosystem from January through August 2026, drawing six findings from Hub data. Public model repositories grew from 2.43 million to 2.96 million and datasets from 711,000 to one million, but 85.6% of models have fewer than 200 lifetime downloads, while 1.5% of repositories account for 99.2% of downloads.
Why it matters: Hugging Face analyzes its Hub data on the open-model ecosystem; findings on downloads versus likes and the scale of Qwen derivatives can inform model selection.
Sébastien Bubeck calls the reported result amazing, noting that he spent a week trying to solve the problem back in 2012. In the quoted post, Steven Strogatz says a neurosurgery resident with no training in advanced math used ChatGPT 5.6 to solve a major open problem in numerical linear algebra. https://alextownsend.net/essays/SIAMNews_CrouzeixConjecture.pdf
Import AI issue 468 summarizes 23 policy recommendations from the think tank IFP for managing risks from automated AI research and development. They span seven categories, including improving transparency in automated AI R&D, accelerating development of AI verification technology, and extending U.S. AI leadership.
Andrej Karpathy gave Opus 5 the opening passage of The Lord of the Rings and a 1M-token budget (about $10), asking it to render the story in Three.js. Opus ran for about two hours and wrote 5,500 lines of code to procedurally generate a rough but playable 3D scene.
We share our approach to open-weights releases, how we assessed Inkling, why safety depends on both the model and the ecosystem it enters, and how testing, staged access, and stronger defenses can create a path toward greater openness.
Releasing weights indiscriminately isn't safe. Neither is keeping capable models inside a few labs. We think there's a path between them. We haven't mapped all of it. Our new post covers the part we can see: how we assessed Inkling, and why access should widen in stages. https://thinkingmachines.ai/blog/a-safe-path-to-open-weights
A Hugging Face blog argues that GPU utilization, not model intelligence, is AI's next real constraint. GPU costs accrue by calendar hours, but output is produced only during compute hours; after buying GPUs, companies face the challenge of keeping their clusters busy. Using aviation as an analogy, the article says even fully loaded clusters may waste much of their potential because GPUs run all day while demand does not.
It continues to boggle the mind how many people, who otherwise seem to have a functioning intellect, appear to lose all cognitive capacity when it comes to thinking about actions that impact the company that pays them.
Jim Fan reacts to Jensen Huang’s post with “Stay hungry, stay foolish. Absolute legend.” Huang shares a letter signed by NVIDIA arguing that open models strengthen safety and cybersecurity, accelerate innovation and diffusion, and enable sovereignty. He says the world needs both frontier closed models and frontier open models. https://images.nvidia.com/pdf/Open-Weights-and-American-AI-Leadership.pdf
The knowledge that makes AI useful is diffused. It lives with scientists, engineers, clinicians, firms. For AI to benefit from distributed knowledge, it must itself be distributed. Agree with Jensen that this is a future worth building.
An analysis by the UK's AI Security Institute finds that the gap in cybersecurity capabilities between open-weight models and closed frontier models is narrowing. GLM-5.2 comes closest to Claude Opus 4.6, released 4.3 months earlier; DeepSeek-V4-Pro falls between Claude Opus 4.5 and GPT-5. For much of 2025, the measured gap was six to ten months.
Berkeley's EPIC Data Lab argues that as inference costs have fallen from around $30 per million tokens for GPT-4-level models in early 2023 to under $1 today—and below $0.10 at some vendors—data systems face three new challenges: systems designed for agents, systems operated by agents, and systems built by agents.
new post on harness engineering for AI self-improvement: https://lilianweng.github.io/posts/2026-07-04-harness/ It is hard to forecast how much the future of RSI will rely on harnesses. Likely harness engineering will evolve in the direction of self-improvement and enable auto-research, and, in turn, smarter models keeps harnesses simple. Even when many harness improvement get eventually internalized into core model, the need to specify goals and context will not disappear.
Import AI 464 summarizes several research developments. Fable wrote a megakernel on KernelBench-Mega that achieved an 18.71× speedup with CUDA on an RTX PRO 6000 Blackwell, compared with 14.4× for Claude Opus 4.8, 11.14× for GLM-5.2, and 4.34× for GPT 5.5.
Dharma AI discusses the 2026 paper by Goldfeder, Wyder, LeCun, and Shwartz-Ziv, “AI Must Embrace Specialization via Superhuman Adaptable Intelligence.”
Fun fact: @woj_zaremba wanted to completely remove that paper from existence a year or so after. I forgot why though -- Woj? Quoted post by @OfirPress: I started doing language modeling research in Jan 2016, using the LSTM from the "RNN Regularization" model by @woj_zaremba @OriolVinyalsML and @ilyasut, trying to improve its perplexity on Penn Treebank (1M toks). This led to the weight tying method, sometimes still used today.
Anthropic has introduced Claude Tag, letting Claude join Slack as a team member with access to designated channels and tools; users can delegate tasks by mentioning it. Karpathy calls this the third major redesign of the LLM interface: first came visiting websites, then downloading apps, and now an autonomous, persistent, asynchronous entity with organization-wide tools and context working alongside human teams. He says this requires deep engineering across tools, integrations, compute environments, memory, and security, but is achievable.
This issue of Import AI summarizes three studies. Researchers from King's College London, Fudan University, and the Alan Turing Institute built SocioHack, a benchmark with 72 sandboxed social environments testing whether AI can game rules in scenarios such as credit-card rewards and school grading. LLMs trained with RL rediscovered patched vulnerabilities in historical environments with 61.25% recall and 90.85% precision.
Import AI 459 discusses the difficulty of AI regulation, scaling laws for protein-folding models, and how to price extinction risks from AI systems. Economists at the University of Virginia, Anthropic, and the Bank of Canada estimate that nominal U.S. AI GDP was about $250 billion in 2025, with quality-adjusted real annual growth of about 2,600%. U.S. AI compute spending rose from $37 billion in 2023 to $219 billion in 2025, while compute capacity grew by more than 200% annually.
Import AI issue 458 includes a long essay based on a 2026 Cosmos lecture co-hosted by Oxford's HAI Lab and the Cosmos Institute, along with a fictional story imagining a “positive singularity.” Using the Epoch Capabilities Index (ECI), which covers more than 40 benchmarks, as a guide, the essay argues that continued AI progress forces a choice between “exploring the future” and “avoiding the present.”
Import AI 457 focuses on three studies. SentinelOne analyzed fast16.sys, a virus from more than 20 years ago that corrupted results from high-precision software such as LS-DYNA 970, PKPM, and MOHID by tampering with in-memory calculations, sabotaging scientific and engineering work.
The Institute for Law & AI proposes “radical optionality” in regulation: governments should avoid overregulating in the short term while building institutional tools in advance, including information gathering, whistleblower protections, and evaluation audits. A paper from Meta and KAIST proposes the Neural Computer, unifying computation, memory, and I/O in a learned runtime state.
A Berkeley AI Research team reviews progress in parallel reasoning and notes that most current methods—including Self-consistency, Best-of-N, Tree of Thoughts, MCTS, ParaThinker, GroupThink, and Hogwild!—
Jan Leike writes that he is excited about NLAs as a new tool in the interpretability toolkit. In the quoted post, @saprmarks presents NLAs as an unsupervised method for converting an LLM's internal state into human-readable text and says the results advance understanding of what LLMs are thinking and auditing them for safety.
Beyond today’s instant model update, what can we improve? @ s ama says, “hey chat, we haven't forgotten about you 👀” Nick Turley asks what can be improved beyond today’s instant model update.
In Import AI 455, Jack Clark estimates a greater than 60% chance that by the end of 2028 AI R&D will run without human involvement, with frontier models autonomously training their own successors. He puts the chance of this happening by 2027 at about 30%.
Nick Turley reports usage of the new ChatGPT images is up >50% in just a few weeks, with nearly 60% of daily users coming from newly logged-in users. He highlights a broad range of uses across home design, learning, work graphics, and creative tasks.
Respect to Anthropic for not backing down
