Claude formalized Fermat's Last Theorem: 11 days, 13 million lines of Lean, zero human shortcuts
Anthropic announced on September 4 that Claude produced the first complete computer-checked proof of Fermat's Last Theorem, working "largely autonomously" over 11 days. The run, initiated by Anthropic researcher Tianyi Peng, wrote roughly 13 million lines of Lean 4 code — more than five times the size of Mathlib, the community's principal formalized-mathematics library — and proved about 30,300 supporting theorems, of which 29,500 appear in the final argument. The proof follows a simplified version of Andrew Wiles's strategy by Darmon, Diamond and Taylor, rests on only Lean's three standard axioms, and was double-checked through a second independent Lean kernel implementation. The full artifact is public under Apache 2.0 at github.com/anthropics/fermats-last-theorem. For scale: a community effort led by Imperial College's Kevin Buzzard to formalize the same theorem in Lean has been running since 2024, is funded through 2029, and remains unfinished.
The project is as notable for its engineering as for the mathematics. Anthropic says the first multi-agent setup degraded after a few days — agents lost project state and stopped coordinating, with roughly 7% of lines turning into non-boilerplate failures — so the team moved the work onto Prove2Me, an open orchestration platform that schedules proof tasks as a dependency graph, plus a Claude Code-based workflow that ran dozens of agents in parallel. The whole run consumed about 6 billion output tokens from an internal research model roughly comparable to Claude Fable 5.1. Buzzard, after reviewing the result, called it an "extraordinary autoformalization achievement" while noting it tells mathematicians little they did not already believe — the value is that AI can now translate a massive body of literature into machine-checkable form in days. That is the quiet headline for anyone building agents: verification, not discovery, is where the leverage landed, and a repo with pinned dependencies, a replayable build and an independent checker is the kind of receipt agent-produced work will increasingly be judged against.
— Anthropic (official research post) · GitHub · Techstrong AI
🔗 Anthropic: Formalizing Fermat's Last Theorem · GitHub: anthropics/fermats-last-theorem · Techstrong AI: Claude agents formalize FLT in 11 days
GPT-6 Astra spreads across Devin, Copilot, Codex and OpenCode — with the economics now measured by outsiders
Two days after launch, GPT-6 Astra stopped being an OpenAI-only event. Cognition put Astra into Devin Desktop, Devin CLI and the Devin Cloud model mixture; GitHub Copilot marked it generally available for all users on September 4; Codex CLI releases 0.153.3 and 0.153.4 added the model to the Amazon Bedrock picker and fixed a bug hiding it from the bundled picker; OpenCode shipped two patches so gpt-6-astra resolves correctly for OpenAI subscription users. What makes this different from a normal rollout is how fast independent benchmarks arrived to frame the economics. Cognition reports Astra scores 64.5 on its FrontierCode 1.1 benchmark — above Claude Fable 5.1 (63.6), within 0.4 points of Claude Fable 5, and roughly 64% cheaper per rollout. Perplexity's WANDR research-agent benchmark puts Astra at 0.682 for $11.98 per task, the highest score it has recorded, 13.5% above Fable 5.1 at lower cost.
The picture is more mixed once you look past vendor numbers. Artificial Analysis' Coding Agent Index gives an Astra run in Codex a 67 at max effort — level with Fable 5 and Opus 5 in their own harnesses — but on its Intelligence Index Astra ties its predecessor at 61, five points behind Fable 5.1, and the unit price is 2.5x GPT-5.6 Sol. Halving of the hallucination rate (to 51%) is offset by a higher price per task at the top end. Benchmark methodology also matters: Simon Willison flagged that Astra's headline 99.9% ARC-AGI-3 score came from a stateful provider-adapter harness estimated at $19,000 per run, versus 62.7% with the default harness. The takeaway for teams is practical: model choice is becoming table stakes across every major agent, so the deciding factors are shifting to measured cost-per-task, harness quality and the ability to swap models as benchmarks settle.
— Cognition (Devin blog) · OpenAI · Artificial Analysis
🔗 Cognition: GPT-6 Astra is now available in Devin · OpenAI: GPT-6 Astra · AI Primer: third-party benchmark roundup
GitHub's HydraFusion stops choosing one model and starts orchestrating several per request
GitHub introduced Project HydraFusion on September 4 as a research preview inside GitHub Copilot CLI, available on all Copilot plans behind the /experimental flag. Instead of routing a prompt to a single model, HydraFusion builds an execution plan per request and picks one of three patterns: Single, where one model solves the task; Cascade, where a cheaper model drafts and a quality gate decides whether to escalate; or Critique, where a draft from one family is reviewed by an independent read-only critic from a different model family before a single revision. There is no separate HydraFusion fee — usage is billed at each underlying model's standard token rate. GitHub's five operating principles (complete accounting, bounded execution, isolated review, fail-safe application, validated routing) are the unglamorous core that makes multi-model turns safe on a repository: critics run in tool-less contexts and cannot modify code, and no patch is applied if a run is cancelled or fails validation.
The published numbers tell a cost story more than a quality story. In controlled offline evaluations against a Claude Opus 5 baseline, HydraFusion improved verified task quality by 4.9 points on TerminalBench 2.1 at 67% lower estimated cost, but trailed Opus 5 by 1.5 points on DeepSWE while cutting cost 36%, and was essentially flat on CheckpointBench at 65% lower cost. Nadella promoted the feature the same day it shipped, framing it as the shift from picking one model to orchestrating several. The preview currently handles only first-turn, single-prompt tasks. The strategic signal matters more than the numbers: if routers like this become the norm, the frontier model becomes a specialist called only for the hard tail of requests, and pricing power migrates from model vendors to whoever controls the routing layer.
— GitHub (official blog) · GitHub Community · TechQuire
🔗 GitHub Blog: Project HydraFusion · GitHub Community discussion · TechQuire: how multi-model routing works
Shopify's Slack-based agent River cut the vulnerability backlog by 70% in 11 days
Shopify's engineering blog detailed how River, its AI agent that lives in company Slack, now drives vulnerability remediation to closure rather than stopping at patch generation. River works from the root of Shopify's monorepo, World, reusing the same reproducible development environments and written-down engineering conventions as human developers. Its key move is skepticism about its own ledger: every open finding is re-checked against live repository, PR and vulnerability-tracker state before any code is touched, which matters because roughly a third of the apparent "open" backlog turned out to be already fixed or obsolete. For real issues it updates the patch when safe, pulls in the right engineers for product-behavior decisions, follows the work after handoff, and only calls a vulnerability done after verifying the repository head and resulting dependency graph post-merge.
The reported results are concrete: in the first 11 days of the dependency workflow, the open backlog fell about 70% — roughly two-thirds direct merges, the rest closed with evidence as outdated — and security merges through the freshness-gated merge queue went from about 10% to 80%. Shopify has stressed the asymmetry behind the design: attackers can tolerate failed attempts, but defenders must preserve production behavior with every fix, so verification and merging matter more than patch volume. River is one of the most heavily used internal coding agents in the world — it now co-authors roughly one in eight merged PRs at Shopify. The transferable lesson for other security teams is not the agent itself but the operational loop: treat agent memory as claims to verify, confirm a vulnerability still exists before fixing it, and measure remediation by merges, not by patches written.
— Shopify Engineering (official) · AGI Hunt
🔗 Shopify Engineering: How River takes security work from a fix to merge · AGI Hunt: results summary
Figure books up to 100,000 NVIDIA Vera Rubin GPUs through Nscale — humanoid AI is now a compute problem
Figure announced on September 3 a strategic partnership with UK-based AI cloud provider Nscale to deploy the NVIDIA Vera Rubin platform across up to 100,000 GPUs, with an initial compute commitment of $3.5 billion and intent to scale beyond $6 billion. Initial deployment is targeted for the second half of 2027 at Nscale's facility in Barstow, Texas. Nscale is also making a strategic investment in Figure, and the two will explore using humanoids in Nscale's supply chain. Figure's rationale is direct: "We are entering a phase where we are largely bound by data and compute needed to train Helix." Its crowdsourced training-data initiative Index is generating roughly 35 minutes of real-world video every second, and the company says data alone cannot solve physical intelligence — it needs the compute to convert that firehose into motor policies.
NVIDIA's Jensen Huang framed the deal as the activation of a "physical AI flywheel": training Figure's models on Vera Rubin through Nscale's cloud, validating them in NVIDIA Isaac Sim, and deploying them on NVIDIA GPUs inside Figure's robots. The context explains the scale. Figure 03 production passed 350 units and ramped from one robot per day to one per hour earlier this year, and the company publicly split from OpenAI over the belief that its internal AI team had outpaced the lab. Vera Rubin, unveiled in January, pairs the Vera CPU with the Rubin GPU and claims up to 10x lower inference token cost and a 4x reduction in GPUs needed to train mixture-of-experts models versus Blackwell. Robotics has reached the point where the constraint is no longer actuators — it is data centers, power and the compute runway to train the next generation of embodied models.
— Figure (official) · NVIDIA (Jensen Huang) · Unite.AI
🔗 Figure: partnership with Nscale for up to 100,000 GPUs · Unite.AI: deal analysis · Interesting Engineering: 100,000 GPUs for humanoid training
Gimlet Labs raises $300M at a $3B valuation for a cloud that splits inference across every kind of chip
Gimlet Labs announced a $300 million Series B on September 4, led by Andreessen Horowitz, with new strategic investors Arm and Microsoft's M12 joining Sapphire, Menlo, 645 Ventures, Factory, Hudson River Trading, Samsung Ventures, Tiger Global and others. The round values the company at $3 billion roughly six months after its $80 million Series A. Gimlet's pitch is that inference is no longer one job: prefill is compute-heavy, decode is memory-bound, and agent workloads stack dozens of model calls, tool runs and CPU work into latency-sensitive loops. Its software traces a model, decomposes it into phases (prefill/decode, speculative decode, attention-FFN), and schedules each slice onto whichever accelerator fits the SLA — GPUs, near-memory SRAM compute, dataflow chips or CPUs — dynamically rebalancing when capacity fills. The company claims 3-10x faster frontier workloads or 5-10x more throughput within the same power envelope, and says monthly token generation has grown 6x in 12 months while models have reached 3-10 trillion parameters.
The strategic logic is about power as much as silicon. Gimlet says AI data centers consumed roughly 18 GW in 2025 and could triple by 2030, and a16z's own framing is a watt shortage — five US hyperscalers are expected to spend about $1 trillion on capex next year while plants and fabs still lag demand. The company reports billions in contracted revenue since March, a gigawatt-scale data center pipeline, and one of the top three frontier labs plus one of the top three hyperscalers as customers. The round is a bet that the next bottleneck in AI infrastructure is orchestration rather than another homogeneous GPU hall: Arm and M12 are effectively paying for a software layer that makes silicon choice a routing decision instead of a lock-in, and Gimlet joined MLCommons in June to push for vendor-agnostic agentic inference benchmarks. The open question is execution — mixing chip architectures, cooling profiles and networks is exactly where outages start.
— Gimlet Labs (official) · NewDecoded · AI2Work
🔗 Gimlet Labs: Announcing our Series B · NewDecoded: $300M for multi-silicon inference cloud · AI2Work: round analysis and context
South Korea's chip exports jumped 209% in August — the AI memory boom is now visible in national trade data
South Korea's August exports rose 68.7% year-on-year to $98.25 billion, staying above $90 billion for a third consecutive month, according to the Ministry of Trade, Industry and Resources. Semiconductors were the engine: chip exports surged 209.0% to $46.65 billion — an all-time monthly record, the third straight month above $40 billion, and 47.5% of total exports — supported by sustained AI infrastructure capex from hyperscalers such as Google and Amazon. The IT complex compounded the effect: computer exports rose 419.5% to $6.24 billion on climbing NAND prices for enterprise SSDs, and wireless-communication-device exports gained 21.2% on Galaxy S26 and Z Fold 8 shipments. The trade surplus reached $34.75 billion, the third month above $30 billion, and the January-August surplus hit $202.5 billion, up $162.2 billion from a year earlier.
The divergence elsewhere is as striking as the headline. Automobile exports fell 29.8% on summer-holiday base effects and partial strikes, and ship exports dropped 45.9% on delivery timing — the ministry itself describes the pattern as "K-shaped." By destination, exports to China jumped 119.3% to $24.1 billion and to the US 89.3% to $16.5 billion, both driven by chips and computers. The data reinforces what the memory vendors have been saying: SK hynix's CEO said last week that AI-driven demand will keep memory supply-constrained through the end of the decade, revising his earlier timeline two years later than Samsung's July forecast. For anyone tracking the AI buildout, national trade figures have become a useful second-order signal of how far the training and inference boom has spread beyond the model labs.
— Korea MOTIE (official) · Korea JoongAng Daily · Macrostream
🔗 MOTIE: August exports exceed USD 90 billion for third consecutive month · Korea JoongAng Daily: exports surge as AI chip demand soars · Macrostream: macro analysis
Next digest: September 7, 2026
Top comments (0)