Area
Reporting and discussion about Models & Research, with links to the original sources.
The Moonshots panel debated a paper described on air as finding a 98% overlap in the reasoning pathways of leading language models. The study, Artificial Hivemind, measured somethi...
Updated 14 Sep 2026 IdeaA drone navigating by sound cannot carry a complete account of the world it flies through. Rich Sutton and Khurram Javed use that mismatch to explain why AI must keep learning. The...
Updated 14 Sep 2026On Training Data, Box CEO Aaron Levie was asked where enterprise AI memory is heading — retrieval, or models whose weights absorb a company's knowledge. His answer started with a lawyer who can see five matters and whose access changes daily, and ended with a wish for a rubric deciding what gets baked in and what stays a lookup.
Asked what comes after large language models, Alex told a caller on Moonshots with Peter Diamandis to separate two things people usually merge: the job of predicting the next piece of text, which he thinks has "effectively infinite longevity," and the transformer machinery doing it, which he says is already being swapped out part by part. Dave added his own forecast that the chips underneath will move to photonics within 18 months to two years.
A Moonshots panel unpacks Dwarkesh Patel and Jerry Han's experiment, which found that improvements in training data delivered a 12-fold compute-efficiency gain between 2019 and 2025 against 3.7-fold for architectures and training recipes — at small scale, on easy benchmarks. The panel then splits over whether a company's proprietary data is a durable advantage, with a $32 billion data-subsidiary valuation on one side and the fate of BloombergGPT on the other.
Nvidia's chief executive declared AGI achieved on September 6 while announcing more GPU capacity, and the Moonshots panel split between calling the label meaningless and calling the underlying capability the most important moment in history. A second claim on the same show — that OpenAI's agents now do 3.1 days of research work per human day — comes from an internal report that measures how long agents ran, not how much research they finished.
OpenAI said on 8 September that an internal model, running roughly 10,000 agents for 88 hours, produced a forced blowup construction for the Navier–Stokes equations and a machine-checked proof of it. On Moonshots with Peter Diamandis, the panel worked through what the result is — a statement about idealized fluids, not a device — what it cost, and why the credit for it was contested within hours.
World Labs released Atlas, a model that generates video along a camera path the user designs and rebuilds scenes from a handful of photographs. Its own garden example shows the seam: one photo leaves the surrounding buildings invented, while more photos pin them down. On Moonshots, the panel worked through what Gaussian splats are and why the approach might matter for robots, games and planning a vacation.
On Moonshots with Peter Diamandis, a panelist interrupted an argument about AI regulation to read a headline off his feed: Anthropic had formalized Fermat's Last Theorem. Anthropic's report describes dozens of agents working eleven days, about six billion output tokens and 30,300 intermediate theorems — plus a piece of bookkeeping software that stopped runs from losing track of their own work. The panel's takeaway was about how to narrow enormous machine output into one result you can build on.
OpenAI's GPT-6 Astra nearly saturates the interactive ARC-AGI-3 benchmark and leads Epoch AI's composite capability index, yet sits third on Artificial Analysis's suite, behind Claude Fable 5.1 and Muse Spark. On Moonshots EP #286, the panel works through what each ruler measures — and argues that Astra's real target was doing tasks with fewer output tokens, so a model can drive a desktop at conversational speed.
A TIME report has Sam Altman expecting an internal system he would call AGI within four months, and OpenAI's chief scientist saying its unreleased Astra model has met an internal benchmark for an automated research intern. On the Moonshots panel, the label mattered less than a practical test: whether the next model can finally keep hold of what it has learned over a long job, instead of handing a summary to a successor and starting again.
On Moonshots, the panel revisits Elon Musk's January prediction that models were "off by two orders of magnitude" in intelligence per gigabyte, after Tim Sweeney tweeted that it had come true and Musk replied that specialist AIs add another 100x. Dave calls 100x a lower bound and asks what anyone would actually do with 10,000 brilliant agents; Emad Mostaque describes running specialized agent teams, while Alex argues Musk's "specialist models" are really sparsification inside generalist models.
Alibaba's Wan 3.0 and a relayed claim that 70% of Chinese AI token use goes to video sent the Moonshots panel into an argument about money: one guest said American labs chase revenue per token while Chinese labs give their weights away, another said video is the only market that will trust a Chinese model. They ended up disagreeing about whether world models or text models reach self-improving AI first.
On Moonshots with Peter Diamandis, the panel read out a new leaderboard result: Google's Gemini 3.7 Flash on top of the AA-AnalystAgent benchmark with 60%, ahead of Claude Opus 5 at 54%. Diamandis called it proof that Google is back; Alex argued the score measures repeated reliability on spreadsheet analysis rather than frontier capability, and blamed Google Search for pushing Gemini toward speed and determinism. Emad Mostaque agreed the model was decent but said Google's problem is institutional, not a shortage of chips.