Models & Research

Ten cities, half-day workshops, trades and logistics rather than the tech crowd. The biggest lever turned out not to be the model, but owners learning from each other.

For a decade Shopify was React Native's flagship company. Now they're rebuilding native — not because React Native got worse, but because the cost argument for it fell away. Three open-source libraries lose their sponsor in the process.

AlphaGenome Atlas is a petabyte of predictions — every single possible base change in the human genome. Reachable through a web portal, no code required.

Three scenarios, a handful of sliders, wildly different numbers. The new Economic Scenario Explorer runs from 'about like the internet' to 32 percent more GDP with the labor share down at 45 percent.

DeepSeek has formally launched V4.1-Flash. The interesting number isn't a benchmark, it's memory: roughly a quarter of what the previous model needed. From September 14, all V4 Pro requests go to Flash.

A post-money valuation above €21 billion, led by Samsung Electronics, with BlackRock funds and the Grand Duchy of Luxembourg joining. Mistral is sticking with open weights, and that's the part worth noticing.

Roughly 10,000 agents, 88 hours, 130 billion tokens: OpenAI shows a smooth flow can break down in finite time. Then comes the accusation that an Anthropic researcher was pushed off the author list.

Konstantin Ryabitsev added up what the kernel servers spend on crawlers. The answer: across five machines, 14 cores run continuously just to render Git commits as HTML — more compute than all legitimate access combined.

Anthropic has published the first complete computer-checked proof of Fermat's Last Theorem. Dozens of Claude agents wrote it in Lean over eleven days. Kevin Buzzard, who has been funded since 2024 to do exactly this, compiled the code and says it checks out.

Two days after Fable 5.1, OpenAI ships GPT-6 Astra. Same API price, top scores on computer use, math and terminal tasks. Fable 5.1 still leads the Artificial Analysis Intelligence Index. What the numbers say, and what Matt Wolfe found during early access.

After weeks of rumors it's official: Nvidia is acquiring Hugging Face for $12.93 billion. Huang says the platform stays open and Nvidia hardware won't be required. Clem Delangue explains the sale with a lack of compute. Why this is a turning point for open weights.

Anthropic shipped Fable 5.1 and Mythos 5.1 – the same model with different levels of safeguards. Two things stand out: the price drops by around 25 percent, and the real-world research examples are surprisingly concrete.

Simon Willison ran all five effort levels of Fable 5.1, Nick Saraev picked apart the benchmarks. The upshot: low and medium apparently don't think at all, while max produces the best pelican Anthropic has ever drawn, at a cost of 14 minutes.

On the same day as Fable 5.1, Google puts out its third Flash version in six weeks. 90.8 percent on Terminal-Bench 2.1, the same price as before, and a locked-down cyber model on top.

To understand how misalignment forms, Anthropic deliberately trained an Opus-class model on 80 hackable training environments. The result: a model that cheats, games its own reward, and in testing even gives bioweapon tips.

Nvidia, Stripe and others are pouring billions into companies that give their models away. TechCrunch lays out why open-weight is becoming a takeover prize – and what it means for Anthropic and OpenAI's dominance.

The Bun rewrite was built with a pre-release Fable 5 – one developer, eleven days, 6,778 commits. Paul Dix draws a conclusion: this is what normal software work will soon look like.

125 billion parameters, but only 6 active per token – plus a new sparse-attention trick. Alibaba open-sources its next model before Qwen4 even arrives.

Last week an anonymous coding model called 'Ox Alpha' showed up on OpenRouter — nobody knew who built it. Now it's clear: it's GLM-5.3-Flash from Z.ai, and the weights are out in the open on Hugging Face.

NVIDIA's AVO agent scores a perfect run on a brutal benchmark. The model inside it is Claude Opus 5 – the same one that scores only about 30 percent on its own. The whole difference is the scaffolding around it.