[AINews] OpenAI 称借助 Astra-next、约 1 万个 Agent 与 1300 亿 Token 在 88 小时内发现 Navier-Stokes 奇点；或将角逐史上第二个千禧年大奖

今天是个很难发布任何消息的新闻周期；我们通常承诺报道任何新的十角兽融资，因此，Cognition 的 480 亿美元融资和 Mistral 的 240 亿美元融资原本通常都会入选；我们热爱图像生成，所以 GPT Image 2.5本应拥有自己的头条；我们曾密切报道Dreamer 的故事，因此其以 Meta 的 Muse Agent重新亮相本来也应当入选；但是……你知道的……如今的门槛更高了。

[我们正在分享 Navier-Stokes 千禧年大奖难题的一个解答——这是数学前沿最深刻的问题之一。这份证明由一组 Agent 产出，所使用的是一款明显强于 GPT-6 Astra 的 OpenAI 下一代模型。该问题……下午 5:20 · 2026 年 9 月 8 日 · 14.5 万次浏览

---

239 条回复 · 603 次转发 · 2720 个赞](https://x.com/openai/status/2097374640582668336?s=12)

下文摘要涵盖实质性事实；我们建议不要过度深究署名争议，因为 OpenAI 和作者基本已给出足够细节，可以得出 OpenAI 的成就是真实的这一结论，尽管过程仍存在一些争议。

2026 年 9 月 7 日至 9 月 8 日的 AI 新闻。我们查看了 12 个 subreddit、544 个 Twitter 账号，没有再查看其他 Discord。通过 AINews 网站可以搜索往期全部内容。提醒一下，AINews 现已成为 Latent Space 的一个栏目。你可以选择订阅或退订不同的邮件推送频率！

---

OpenAI 关联账号称，一项 AI 辅助工作得出了 Navier–Stokes 结果，而其引发的即时反应在技术兴趣、怀疑态度与元话题争议之间迅速分化。

• 这组推文中最具体的公开说法来自 Ethan Knight。他表示，“Navier Stokes 解答是约 1 万个 Agent 共同协作的结果”，并补充称 OpenAI 在“过去一年”中一直使用“多智能体 RL”训练模型进行协作；他还称，难题可能会屈服于“海量非结构化并行测试时计算”，由模型自行决定如何组织协作 @eknight。

• 多名旁观者将此理解为 OpenAI 宣称生成了一份与 Navier–Stokes 千禧年难题相关的 AI 证明，具体涉及有限时间奇点／爆破；一则讽刺性转述将其表述为 OpenAI 称平滑流体能够“爆破成一个奇点”，并称使用了“1 万个 Agent”和“88 小时”，同时明确指出数学界的接受仍只是一个“微小的手续” @LearnOpenCV。

• 更广泛的评论将这一事件视为一种潜在压力测试，检验“前沿 AI 无法从事严肃研究或编程级技术工作”这一信念。Theo Jensen 称，这是科学界的“‘AI 实际上并不能写代码’崩溃时刻” @theo。

• Hrishikesh / hrishioa 将该公告视为“高算力时代”的证据，认为观察者应当“据此调整你们的计划” @hrishioa。

• 这一公告还引发了附带的运营猜测：有发帖者开玩笑地将看到 ChatGPT 延迟警告，与 OpenAI 可能将大规模算力重新调配至 Navier–Stokes 运行联系起来；但这纯属猜测，并非证据 @teortaxesTex。

从推文中可以确认的事实

• 有一项由 OpenAI 关联方提出的说法流传开来：一项 Navier–Stokes“解答”涉及约 1 万个 Agent 协同工作 @eknight。

• 同一来源称，这些系统通过 多智能体强化学习 进行了约 一年 的训练 @eknight。

• 所述高层方法强调 并行测试时计算 以及模型的自组织，而非一次由单一长链条完成的证明尝试 @eknight。

• 公众读者将该说法理解为涉及 Navier–Stokes 存在性／奇点问题——即 千禧年大奖难题之一；不过，推文中未提供精确的定理陈述和证明范围 @LearnOpenCV。

• 在讨论发生时，数学界是否接受该结论显然仍未确定；即使是那则玩笑帖也强调，其正确性尚未得到该领域验证 @LearnOpenCV。

推文尚未确立的内容

• 所提供的推文中没有出现定理陈述、预印本、证明纲要、形式化验证产物、基准测试报告，或独立审稿人的评论。

• 被频繁重复的 “88 小时” 细节，在这组材料中只出现在一则讽刺帖子里，并未出现在更直接的 OpenAI 相关表述中；因此，仅根据这些证据，不应将其视为已获确认的信息 @LearnOpenCV。

• 人类与模型各自的确切角色尚未说明：“约 1 万个 Agent 的协作”并不能告诉我们，是否由人类拆解了搜索任务、筛选了引理、验证了步骤，还是仅仅启动了基础设施 @eknight。

• “解答”一词含义模糊。在数学中，它可能指完整证明、证明策略、候选反例、形式化推导，或研究线索。推文并未厘清这一点。

• 这里没有披露该结果是否涉及标准的三维不可压 Navier–Stokes 方程在 (\mathbb{R}^3) 或环面上的整体正则性问题，还是某种变体／辅助命题。

为何这种模糊性至关重要

• Navier–Stokes 千禧年难题有极其具体的标准表述。若声称有限时间奇点“可以发生”，将具有爆炸性意义，因为这意味着在相关表述下，整体正则性问题的答案是否定的；这类说法需要非同寻常的精确性与审查。

• 在前沿模型讨论中，“AI 解决了 X”往往压缩了多层含义：猜想生成、搜索、证明起草、证明检查，以及社区验证。推文只给出了系统层面的描述，而没有说明数学结论的认识论状态。

所披露的技术图景，与其说关乎流体力学，不如说关乎一种研究系统架构。

• 规模：约 1 万个 Agent 协同运行 @eknight。

• 训练方法：历时约 一年 的 多智能体 RL @eknight。

• 推理理念：投入海量 非结构化并行测试时计算，由 Agent 自主决定如何分工与协作 @eknight。

• **隐含的研究论点：**对于困难的推理任务，扩大推理阶段的 协调 + 搜索，可能与单纯扩大一个单体模型同等重要，甚至更重要。

• **社会技术层面的含义：**这是对许多实验室一直暗示的一项趋势的具体阐释——从“更大的单一模型”叙事，转向 Agent 化集群、并行搜索 和 测试时计算扩展。

• **运营层面的含义：**如果属实，这一结果表明实验室愿意为一次性的科学目标投入大量推理算力，而不只是将其用于产品或基准测试。

这在技术上意味着什么

• 一个 1 万 Agent 的系统意味着需要大量基础设施来支持：

  • 任务拆解，

  • Agent 间通信，

  • 记忆／状态持久化，

  • 搜索树管理，

  • 奖励设计或代理评分，

  • 候选证明路径的聚合／筛选。

• “让它们决定如何协作”这一表述，暗示了一种部分涌现的协调策略，而不是完全由人工脚本编排的编排机制 @eknight。

• 如果这项工作确实触及了一个高难度数学问题，关键创新可能不只是“LLM 写出证明”，而是 采用学习型协作策略的分布式定理搜索。

技术上缺失的内容

• 没有提及：

  • 定理证明器集成，

  • 形式化验证，

  • 证明助手技术栈，

  • 符号代数系统，

  • 流体模拟组件，

  • 检索语料库，

  • 模型规模，

  • 算力预算，

  • pass@k 风格指标，

  • 相对于单 Agent 基线的消融实验，

  • 错误率或证明检查成功率。

这种缺失是核心问题：公开讨论的推进速度，已经超过了所披露技术基础的进展。

作为事实提出的事实／主张

• 约 1 万个 Agent 参与其中 @eknight。

• OpenAI 此前通过 多智能体 RL 训练协作型 Agent 已约 一年 @eknight。

• 该系统使用了大规模 并行测试时计算 @eknight。

• 该结果在公开讨论中被表述为一项 Navier–Stokes 解答／证明主张 @LearnOpenCV。

观点／解读

• 使用海量非结构化并行测试时计算和自组织 Agent，是“解决难题最有效的方法之一”——这是一个强烈的战略性解读，而非仅凭该推文证据就已得到普遍证明的结论 @eknight。

• “科学界正在经历他们的‘AI 实际上并不能写代码’崩溃时刻”是关于社区心理的评论，而非可验证的评估 @theo。

• “我们确实已进入高算力时代”是对行业走向的宏观框架 @hrishioa。

• “88 小时”“领导力课程”以及“委派 1 万个 AI Agent”等表述属于讽刺，不应被视为纪实细节 @LearnOpenCV。

• 声称 ChatGPT 变慢是由这项实验造成的说法，属于没有支持性证据的猜测 @teortaxesTex。

支持性／看涨观点

• 最有力的支持性观点是：这证明了一条新的扩展定律——不仅是模型规模和训练算力，大规模并行、可自组织的推理时协作也能够在前沿研究难题上解锁性质不同的新能力 @eknight。

• Theo 的反应体现了另一种看涨解读：如果 AI 能够对顶尖数学问题作出实质贡献，那么关于 AI 无法从事严肃技术工作的轻视就更难维持 @theo。

• Hrishioa 的“高算力时代”框架意味着实验室和初创公司的战略后果：那些低估推理时算力编排重要性的人，可能是在针对错误的前沿制定计划 @hrishioa。

怀疑性／谨慎观点

• 隐含的怀疑立场来自数学本身：在定理陈述、完整证明和专家审查出现之前，称其为“解答”仍为时过早。那则玩笑帖本身也通过强调全领域接受尚待完成，承认了这一点 @LearnOpenCV。

• 另一项怀疑的对象是叙事压缩：“1 万个 Agent 解决了 Navier–Stokes”可能掩盖了人类在问题框定、筛选或验证中所作的贡献。推文没有披露署名贡献比例。

• 此外还存在可复现性担忧：没有公开产物，独立研究者无法判断这一突破是否稳健、是否经过挑选，或只是一次性事件。

中性／分析性观点

• 一种中性解读是，即使证明最终失败，这仍然值得关注。如果一个系统能够针对如此级别的问题生成数学上非平凡的候选路径，这本身就是一项有意义的能力里程碑。

• 另一种中性观点是，应将 科学真理 与 系统创新 分开。即使

Today was a tough news cycle to launch anything; we ordinarily promise to cover any new decacorn fundraises so Cognition’s $48B round and Mistral’s $24B round would normally have made it; we love imagegen so GPT Image 2.5 would have been its own headline; we covered the Dreamer story closely so their relaunch as Meta’s Muse agent should have made it; but.. yknow… the bar is higher these days.

[We’re sharing a solution to the Navier-Stokes Millennium Prize Problem, one of the deepest problems at the frontier of mathematics. The proof was produced by a group of agents, using an OpenAI next-generation model significantly more capable than GPT-6 Astra. The problem …5:20 PM · Sep 8, 2026 · 145K Views

---

239 Replies · 603 Reposts · 2.72K Likes](https://x.com/openai/status/2097374640582668336?s=12)

The summaries below capture the substantive facts; we recommend not looking too deep into the authorship drama as OpenAI and the authors have pretty much laid out enough detail to conclude that OpenAI’s achievement is real though the process is in some despute.

AI News for 9/7/2026-9/8/2026. We checked 12 subreddits, 544 Twitters and no further Discords. AINews’ website lets you search all past issues. As a reminder, AINews is now a section of Latent Space. You can opt in/out of email frequencies!

---

OpenAI-affiliated accounts said an AI-assisted effort produced a Navier–Stokes result, and the reaction immediately split between technical interest, skepticism, and meta-drama.

• The most concrete public claim in the tweet set came from Ethan Knight, who said “The Navier Stokes solution was the result of a collaboration of ~10,000 agents working together,” adding that OpenAI had spent “the past year” training models to collaborate via “multiagent RL,” and that hard problems may yield to “huge amounts of unstructured parallel test-time compute” with models deciding how to organize themselves @eknight.

• Multiple onlookers interpreted this as OpenAI claiming an AI-generated proof related to the Navier–Stokes Millennium Problem, specifically around finite-time singularity / blow-up; one satirical paraphrase framed it as OpenAI saying a smooth fluid can “blow up into a singularity,” claiming “10,000 agents” and “88 hours” were used, while explicitly noting that mathematical acceptance remained a “minor formality” @LearnOpenCV.

• Broader commentary treated the event as a possible stress test for the belief that frontier AI cannot do serious research or coding-level technical work; Theo Jensen called it the science world’s “‘AI can’t ACTUALLY code’ crash out moment” @theo.

• Hrishikesh / hrishioa framed the announcement as evidence of a “high compute regime,” arguing observers should “adjust your plans accordingly” @hrishioa.

• The announcement also triggered incidental operational speculation: one poster jokingly linked seeing ChatGPT latency warnings to OpenAI potentially redirecting large-scale compute toward the Navier–Stokes run, though this was pure conjecture and not evidence @teortaxesTex.

What is factual from the tweets

• An OpenAI-linked claim circulated that a Navier–Stokes “solution” involved about 10,000 agents working collaboratively @eknight.

• The same source said these systems were trained over roughly a year using multi-agent reinforcement learning @eknight.

• The stated high-level method emphasized parallel test-time compute and model self-organization rather than a single long-chain proof attempt @eknight.

• Public readers understood the claim as concerning the Navier–Stokes existence/singularity problem , one of the Millennium Prize Problems , though the exact theorem statement and proof scope are not supplied in the tweet set @LearnOpenCV.

• Acceptance by the math community was clearly unresolved at the time of discussion; even the joke-post emphasized that correctness remained unverified by the field @LearnOpenCV.

What is not established by the tweets

• No theorem statement, preprint, proof sketch, formal verification artifact, benchmark report, or independent referee commentary appears in the provided tweets.

• The frequently repeated “88 hours” detail appears only in a satirical post in this set, not in the more direct OpenAI-adjacent statement, so it should not be treated as confirmed from this evidence alone @LearnOpenCV.

• The exact role of humans versus models is unspecified: “collaboration of ~10,000 agents” does not tell us whether humans decomposed the search, curated lemmas, verified steps, or merely launched infrastructure @eknight.

• “Solution” is ambiguous. In mathematics it could mean a complete proof, a proof strategy, a candidate counterexample, a formalized derivation, or a research lead. The tweets do not disambiguate this.

• There is no disclosed information here on whether the result addresses the standard 3D incompressible Navier–Stokes global regularity problem on (\mathbb{R}^3) or torus, or some variant/auxiliary statement.

Why the ambiguity matters

• The Navier–Stokes Millennium Problem has a very specific standard framing. Claims that a finite-time singularity “can occur” would be explosive because they imply a negative answer to global regularity in the relevant formulation; such claims require extraordinary precision and scrutiny.

• In frontier-model discourse, “AI solved X” often compresses multiple layers: conjecture generation, search, proof drafting, proof checking, and community validation. The tweets give only a systems-level description, not the epistemic status of the math.

The disclosed technical picture is less about fluid mechanics than about a research system architecture.

• Scale: approximately 10,000 agents operating together @eknight.

• Training approach: multi-agent RL over the course of ~1 year @eknight.

• Inference philosophy: large amounts of unstructured parallel test-time compute , with agents autonomously deciding how to divide work and collaborate @eknight.

• Implied research thesis: for difficult reasoning tasks, scaling coordination + search at inference time may be as important as, or more important than, simply scaling a monolithic model.

• Sociotechnical implication: this is a concrete articulation of a trend many labs have hinted at—shifting from “bigger single model” narratives toward agentic ensembles , parallel search , and test-time compute scaling.

• Operational implication: if true, the result is evidence that labs are willing to spend substantial inference compute on one-shot scientific targets, not just products or benchmarks.

What this suggests technically

• A 10,000-agent setup implies substantial infrastructure for:

  • task decomposition,

  • inter-agent communication,

  • memory/state persistence,

  • search-tree management,

  • reward design or proxy scoring,

  • aggregation / selection of candidate proof paths.

• The phrase “let them decide how to work together” suggests a partially emergent coordination policy rather than entirely hand-scripted orchestration @eknight.

• If the work genuinely touched a hard math problem, the key novelty may be less “LLM writes a proof” and more distributed theorem search with learned collaboration policies.

What is missing technically

• No mention of:

  • theorem prover integration,

  • formal verification,

  • proof assistant stack,

  • symbolic algebra systems,

  • fluid simulation components,

  • retrieval corpora,

  • model size,

  • compute budget,

  • pass@k style metrics,

  • ablations against single-agent baselines,

  • error rates or proof-check success rates.

That absence is central: the public conversation ran ahead of the disclosed technical substrate.

Facts/claims presented as facts

• About 10,000 agents were involved @eknight.

• OpenAI had been training collaborative agents via multiagent RL for about a year @eknight.

• The system used extensive parallel test-time compute @eknight.

• The result was publicly discussed as a Navier–Stokes solution/proof claim @LearnOpenCV.

Opinions / interpretations

• “One of the most effective ways to solve hard problems” is to use huge unstructured parallel test-time compute and self-organizing agents — this is a strong strategic interpretation, not yet demonstrated generally by the evidence in the tweet alone @eknight.

• “Science world is having their ‘AI can’t ACTUALLY code’ crash out moment” is commentary about community psychology, not a verifiable assessment @theo.

• “We truly are in a high compute regime” is a macro framing of industry direction @hrishioa.

• The “88 hours,” “leadership lesson,” and “delegate 10,000 AI agents” framing is satire and should not be read as documentary detail @LearnOpenCV.

• The claim that ChatGPT slowdowns were caused by this experiment is speculation without supporting evidence @teortaxesTex.

Supportive / bullish perspectives

• The strongest supportive perspective is that this is evidence for a new scaling law: not just model size and training compute, but massively parallel, self-organizing inference-time collaboration can unlock qualitatively new capabilities on frontier research problems @eknight.

• Theo’s reaction captures another bullish reading: if AI can materially contribute to a top-tier mathematical problem, then dismissals of AI’s ability to do serious technical work become harder to sustain @theo.

• Hrishioa’s “high compute regime” framing suggests strategic consequences for labs and startups: those who underweight inference-time compute orchestration may be planning against the wrong frontier @hrishioa.

Skeptical / cautionary perspectives

• The implicit skeptical position is mathematical: until a theorem statement, full proof, and expert vetting exist, calling this a “solution” is premature. The joke-post itself acknowledges this by stressing that field-wide acceptance remains pending @LearnOpenCV.

• Another skepticism target is narrative compression: “10,000 agents solved Navier–Stokes” can obscure how much was due to human framing, filtering, or verification. The tweets do not disclose authorship proportions.

• There is also a reproducibility concern: without artifacts, independent researchers cannot judge whether the breakthrough was robust, cherry-picked, or a one-off.

Neutral / analytic perspectives

• A neutral reading is that this is notable even if the proof fails. If a system can generate mathematically nontrivial candidate pathways on a problem of this stature, that alone is a meaningful capability milestone.

• Another neutral view is to separate scientific truth from systems innovation. Even if the theorem claim does not hold, the multi-agent RL + parallel test-time compute architecture may still represent an important advance in AI research methodology.

• The conversation also reveals a shift in what people now count as “capability.” The debate is moving from benchmark scores to real-world cognitive labor decomposition at scale.

This sits at the intersection of three ongoing shifts in frontier AI.

• From static models to agent systems: The central disclosed ingredient is not a single chatbot-like model but a large collaborative population of agents @eknight.

• From training-time scaling to inference-time scaling: The emphasis on “unstructured parallel test-time compute” directly aligns with a broader industry pivot toward spending compute at solve time, not just pretraining time @eknight.

• From benchmark theater to domain claims: Navier–Stokes is socially legible in a way benchmark deltas are not. A claim touching a Millennium Problem instantly broadens the audience and raises epistemic stakes.

Why Navier–Stokes specifically is symbolic

• The Millennium Problems function as cultural shorthand for the hardest kinds of formal intellectual work.

• Progress here would suggest AI systems are not just speeding up known workflows but entering domains where correctness is brittle and prestige filters are extremely strict.

• That said, mathematics is unusually unforgiving: unlike many product tasks, there is no room for “mostly right.” This is why external validation dominates the discourse.

Implications if the claim is substantiated

• Strong evidence for distributed theorem search as a serious research paradigm.

• New pressure on formal methods tooling to absorb model-generated proof candidates.

• A likely acceleration in AI-for-math investment, especially around orchestration, verifier coupling, and scalable search.

• A broader update on the usefulness of test-time compute and multi-agent RL beyond coding agents and office automation.

Implications even if the claim does not fully hold

• It still publicizes OpenAI’s internal strategic direction: large-scale agent collaboration as a core capability area.

• It changes expectations about where compute is being spent and what kinds of demonstrations labs will use to signal frontier progress.

• It may spur competitors to disclose similar systems or rush out rival “AI did science” claims.

A secondary thread of the discussion was about whether details were being indirectly revealed, who was authorized to reveal them, and how much people should infer from fragments.

• A tweet saying “Roon seems like the kind of person who would honor his NDA tbh.” points to a social layer around the story: some observers expected better-known insiders or adjacent figures to stay quiet, while details were instead being pieced together from others @jd_pressman.

• Theo’s “AI can’t ACTUALLY code crash out moment” post also functioned as social provocation, framing critics as emotionally reacting to a capabilities update rather than engaging first with proof standards @theo.

• The two tweets about an “OpenAI movie” image and guessing who appears in it are not about the Navier–Stokes claim directly, but they reflect a parallel tendency to map internal OpenAI narratives onto named personalities like Greg Brockman, Ilya Sutskever, Jared Kaplan, Dario Amodei, and Paul Christiano, even when evidence is thin @willdepue, @jachiam0. In the context of the Navier–Stokes discussion, that tendency matters because people quickly personalize technical claims into author-credit and insider-drama questions.

• The joke and speculation posts show a familiar pattern in frontier AI launches: sparse official detail creates a vacuum that gets filled by memes, leaked-sounding fragments, extrapolation, and overclaiming @LearnOpenCV, @teortaxesTex.

Why the authorship/drama issue matters technically

• For a mathematics claim, provenance is not just gossip. It affects:

  • who framed the conjecture,

  • who selected candidate lemmas,

  • whether the proof was machine-generated or machine-assisted,

  • what credit assignment looks like,

  • how much trust experts place in the artifact.

• In AI research, “multi-agent solved X” also muddies standard notions of contribution. If thousands of agents searched in parallel, then:

  • what is the “author” of the proof,

  • what is the role of the orchestration team,

  • and what exactly should be cited or reproduced?

• NDA and disclosure norms become especially salient when a claim is large enough to move public beliefs before a paper or proof is available.

Meta’s Muse Launch and the Personal-Agent Security Architecture

• Meta launched Muse , a consumer-facing “personal AI agent” positioned as always-on, app-connected, browser-capable, and goal-oriented, with strong distribution through Meta properties and integrations @finkd, @alexandr_wang, @MetaNewsroom. Product details repeatedly surfaced: persistent isolated Linux VMs , browser use, WhatsApp/app interfaces, and connectors to services like Gmail, Calendar, Outlook, Plaid, OpenTable, Docs, Spotify, Peloton, plus unique Meta-native connectors for Instagram, Messenger, Facebook, and Marketplace @alexandr_wang.

• Security architecture is the differentiator being pushed hardest. Meta’s team said each Muse runs in its own secure VM , actions are mediated by a separate Sentinel , secrets are never directly exposed to the agent, sensitive actions require approval, and there is a public bug bounty up to $300k @shengjia_zhao, @alexandr_wang. There’s also explicit commerce infrastructure: Stripe Link for payments with an agentic payment protection / refund guarantee , plus incoming Shop Pay integration @alexandr_wang.

• Early reception from practitioners was notably positive , especially on permissioning, secrets management, and consumer utility. Commentary from @matthuang, @signulll, and @lilyjclifford suggests Muse may be one of the first broadly legible personal-agent products where context and access , not raw model IQ, are the bottleneck. Meta also said usage exceeded internal projections by 10x on day one @alexandr_wang.

• Model and ecosystem placement: Meta’s Muse Spark 1.3 was quickly exposed in third-party tooling like Cursor @cursor_ai, while arena-style benchmarking positioned Muse Spark 1.3 Max as price/perf competitive in web-dev coding workloads @arena.

OpenAI’s Image 2.5 Release and Astra Rollout

• OpenAI also shipped ChatGPT Images 2.5 , though it was partially overshadowed. The release emphasizes up to 50% lower latency vs Images 2.0 , better realism, stronger edit consistency across repeated edits, comment-based localized changes, transparent backgrounds, and a new Sketch tool for guided generation @OpenAI, @ChatGPT, @sama.

• Two API variants were introduced : GPT-Image-2.5 Flare for speed/quality and Sunburst for higher-precision detailed work @reach_vb. Arena results claimed #1 and #2 positions across text-to-image, image-edit, and multi-image-edit leaderboards, with especially large gains in multi-image editing @arena. Integrations landed quickly on fal , Higgsfield , Manus , and Hermes Agent @fal, @higgsfield, @ManusAI, @Teknium.

• Astra availability widened materially. OpenAI said GPT-6 Astra is now fully rolled out to Plus, Pro, Business, and Enterprise users in Codex and ChatGPT Work @OpenAI. Community demos showed strong practical computer-use performance: @theo reported Astra compiling and running Super Smash Bros. Melee on macOS at 120 FPS after a roughly 6-hour loop, while Vals reported Astra nearly saturating an unreleased computer-use eval by building a Minecraft Nether portal in under 3 hours with no specialized harness @ValsAI.

Agent Harnesses, Post-Training, and Serving Infrastructure

• Harvey + Baseten’s M &A diligence work is one of the clearest model-harness co-optimization case studies. Their recursive language model (RLM) harness uses a root agent to search a data room, delegate to sub-agents for document review, and aggregate findings over corpora up to 80M tokens. On the synthetic LAB Diligence benchmark, moving from a standard tool loop to the RLM harness raised mean rubric pass rate from 23% to 62% across models @harvey, @nikogrupen.

• Post-training inside the harness mattered at least as much as the harness itself. Harvey reports self-distilled SFT on GLM-5.2 improved pass rate 46% → 60% , while GRPO on Qwen3.5-122B-A10B lifted pass rate 30% → 63% on held-out rooms and improved document coverage 62% → 96% @harvey. The broader implication, echoed by others, is that agent benchmarks increasingly need to treat orchestration and post-training as part of the model system , not external glue.

• LangChain/deepagents shipped quality-of-life primitives for harness design , including subagent forking that passes supervisor context down to subagents, plus managed connections to abstract OAuth/token/consent flows for either agent-owned or user-owned identities @colifran_, @hwchase17, @caspar_br. This is a useful sign of the stack maturing around long-horizon agent workloads.

Inference and Systems: Sparse Attention, Agentic Serving, and Decode Megakernels

• vLLM’s long-context serving work is notable. The project described Hybrid HiSparse for sparse-MLA models: KV stays on GPU while possible, then cold KV pages are offloaded to host memory , while a hot buffer serves the indexer. On GLM 5.3 with 1M context on an 8×H200 node, configured concurrency 32 , plain offloading sustained 5–6 requests while Hybrid HiSparse sustained 19–25 @vllm_project. This matters directly for RL rollouts and long-context concurrency , where VRAM-bound decode otherwise kills throughput.

• vLLM also published a full-stack optimization pass for real-world agent traffic , benchmarked on AgentX. Key takeaways: pipeline parallelism helps cold long prompts but loses on warm short turns; decode context parallelism depends strongly on the model’s attention stack; and session-sticky routing can beat naive load balancing because warm KV caches matter more than even queue distribution in fast-turn agent settings @vllm_project.

• Cohere introduced an open-source serving stack built around a “decode megakernel,” claiming up to 1.58× faster performance than vLLM on North Mini Code and 1.25×–1.41× end-to-end gains at higher batch sizes @cohere. Combined with Baseten’s note that frontier RL rollouts now get new policy weights live in under 40 seconds globally with only a 6-second pause @baseten, the clear trend is toward infra specialized for continuous post-training and rollout refresh , not static model serving.

Top Tweets (by engagement)

• Anthropic resignation / safety warning : Jacob Hilton resigned from Anthropic, arguing both Anthropic and OpenAI are racing toward self-improving superintelligence irresponsibly and that insiders privately treat extinction risk as real @hilbertspaess, with follow-up claims that current systems could soon hack infrastructure and transform fields rapidly @hilbertspaess.

• OpenAI’s user-data clarification : OpenAI’s formal statement that no specific user data was accessed for Navier–Stokes, alongside the caveat about possible de-identified derivative improvement, became a major flashpoint @OpenAI.

• Cognition financing : Cognition announced a raise of $2B+ at a $48B valuation , saying run-rate revenue grew from $492M to nearly $900M since May @cognition.

• Meta Muse launch : Mark Zuckerberg’s launch post for Muse was among the highest-engagement product tweets of the day @finkd.

---

• Qwen/Qwen-Drive-1.0-4B · Hugging Face (Activity: 549): Qwen releasedQwen/Qwen-Drive-1.0-4B, an open-weight autonomous-driving VLM derived from an unchanged Qwen3.5 4B VLM, with a full BF16 checkpoint around9Band extraplanner-sft,planner-rl, andperceptionmodules. Per the linkedtechnical report, Qwen-Drive-1.0 adds an external BEV perception head for 3D object detection, semantic occupancy prediction, and BEV map segmentation, plus a Planning Expert for future ego-trajectory generation, trained via staged mixtures of driving supervision and general VLM data. The release reports competitive performance across WOD-E2E, NAVSIM, driving VQA, and open-/pseudo-closed-/closed-loop planning evaluations while largely preserving general multimodal capability.

• DeepSeek Flash 4.1 is already being tested via API and rolling out. (Activity: 528): DeepSeek V4.1 Flash is reportedly in internal beta via API: keep the existingbase_urland call modeldeepseek-v4.1-flash-expires-on-0910, with pricing unchanged fromdeepseek-v4-flashand a20**concurrent request/account limit (source). The translated announcement claims a “new model architecture” with native multimodal support, stronger capability, faster throughput, and lower cost; commenters report roughly **2.24×speedup and up to~30%better token efficiency in benchmarks, though one edit speculates the observed speed gain may be partly due to lower beta concurrency rather than architecture alone. Comment sentiment is strongly positive toward DeepSeek/open-weight progress, but the only substantive debate is whether the claimed performance improvement reflects a genuinely new architecture or simply lighter API load during beta testing.

  • Users report that DeepSeek Flash 4.1 appears to be around 2.24x faster via API testing, with some speculation that the observed speedup may come from lower concurrent load rather than a fundamentally new architecture. Other comments suggest it may be multimodal , though this is not yet confirmed in the thread.

  • One technically relevant claim is that some users are seeing up to 30%better token efficiency in benchmarks , which could explain DeepSeek’s reported “lower costs” messaging if fewer tokens are needed for comparable outputs. The comment frames this as benchmark-dependent and not yet independently validated.

  • There is some discussion of release cadence and migration complexity: users mention not having fully moved from the 0731 model to the newer vision variant before another release appears imminent. This highlights a practical API-integration issue where fast model iteration can outpace downstream evaluation, regression testing, and deployment workflows.