The October 12 AI briefing rounds up the past 24 hours: Google DeepMind's AlphaProof Nexus independently solved nine long-open Erdős problems and passed peer review, while OpenAI raised GPT-6.1 Sol's default context from 272k to 500k. Also inside: Alibaba Qwen's training framework for long-horizon agents, reports of Amazon talks to acquire Decart, U.S. junior developer employment data, and Sakana AI's self-improvement research. Infinigence's Series C round, StepFun's Step 5 Preview on WorkBuddy, and Arena's Week 41 leaderboard were already published as standalone stories today and are not repeated here.
AlphaProof Nexus solves 9 Erdős problems, 2 of them open for 56 years
On October 12, The Decoder reported on a new study published in Science: Google DeepMind's AlphaProof Nexus independently solved 9 of 353 open Erdős problems, 2 of them unsolved for 56 years, and the team posted the proofs online ahead of the paper. It also proved 44 conjectures from the OEIS integer sequence database, solved an open question in algebraic geometry, and improved a known bound in min-max optimization. AlphaProof Nexus is built around Gemini 3.1 Pro, which drafts and revises proofs in the formal language Lean; Lean checks every edit and its error messages guide the next attempt, while harder sub-problems can call on AlphaProof, a specialized proving tool. Notably, the work went through peer review instead of dumping a large batch of AI proofs into a public repository — at a moment when academia is debating how AI proofs should be published, this path is far easier for the mathematics community to accept.
GPT-6.1 Sol default context rises from 272k to 500k, with a 1M maximum
On the evening of October 11, OpenAI's Thibault Sottiaux (Tibo) announced that the default context length for GPT-6.1 Sol in Codex and ChatGPT Work has been raised from 272k to 500k, an increase of about 80%. Users can still override the setting up or down, and the model's maximum context is 1M. He said the value was thoroughly tested as a sweet spot between experience and performance, and that it should noticeably reduce how often context gets compacted during long tasks. Alongside the change, subscription usage deductions will no longer charge a long-context fee for GPT-6.1 Sol. For people running long coding and document tasks every day, this means fewer lost details and less re-explaining background mid-task.
Alibaba Qwen releases QwenGyre, built to train hours-long agent tasks
On October 12, Alibaba's Qwen team released QwenGyre, an online reinforcement learning framework for long-horizon agent tasks. A single run of such a task can last hours, involve hundreds of interdependent interactions, and process millions of tokens, which made it hard to turn into training data. QwenGyre uses elastic scheduling to split GPUs into compute units: when a batch of tasks finishes, the freed units immediately join training, while tasks still running keep their progress, state, and cache intact. It also organizes trajectories into a tree by shared prefix, expands only the selected paths into training samples, and gives partial-score feedback for timed-out but partially completed tasks. In the official experiments on NL2RepoBench, training on tasks of roughly 700k tokens per rollout for 48 steps raised the model's pass rate from 52.5% to 58.5%, with about 1.78x end-to-end speedup over an asynchronous implementation under the same GPU budget.
Amazon reportedly in talks to acquire Decart at around a $7 billion valuation
On October 12, The Wall Street Journal reported that Amazon is in talks to acquire AI startup Decart at a potential valuation of about $7 billion. The San Francisco-based startup builds software that lets developers switch seamlessly between different AI chips — an adaptation layer that becomes valuable when everyone is competing for compute but nobody wants to be locked into a single chip vendor. People familiar with the matter said a deal could come together quickly if no new bidder emerges. PitchBook data shows Decart had previously raised more than $400 million, and a funding round in May valued it at close to $4 billion. Bloomberg previously reported that Anthropic had negotiated an acquisition of Decart at around $6 billion before walking away in September; Amazon stepping in now shows the fight over chip-agnostic tooling is still heating up. The talks are not final, and both the price and whether a deal closes remain uncertain.
U.S. junior developer employment falls about 20% while senior hiring grows
On October 12, data reported by CNBC split AI's effect on programmer jobs into two halves. Indeed Hiring Lab data shows U.S. software developer job postings have grown nearly 15% since Claude Code launched in February 2025, with 71% of that growth coming from senior roles and 37% from roles with AI in the title. But Stanford's 2026 AI Index found that since the rise of generative AI, employment among 22-to-25-year-olds in AI-exposed fields has fallen 16%, with early-career software developers down about 20%. Indeed experts warn that if employers stop building entry-level pipelines now, they may face a senior talent gap in ten years; the U.S. Bureau of Labor Statistics still projects software developer employment to grow 10% between 2025 and 2035. For newcomers the signal is clear: roles that only require basic coding are shrinking, while experience supervising AI output, integrating systems, and communicating is worth more and more.
Sakana AI proposes MASS, using multi-agent self-supervision for recursive self-improvement
On the evening of October 11, Sakana AI published MASS, a study in which the model being improved also serves as its own optimizer and evaluator: collective output from multi-agent collaboration is distilled into a self-supervision signal used to repeatedly optimize multi-agent workflows. The team comes from Sakana AI and UC Berkeley, and the official page reports improved score per output token on research benchmarks. The approach targets a practical problem: frontier answers are being generated faster than experts and evidence can review them, so a model may have to teach itself from its own collective work. Reliable autonomous self-improvement is still a long way off, but the setup — the optimizee is also the evaluator — is worth tracking for what it does next on non-verifiable tasks.
