这个 TasteVal 把研究品味换算成省了多少算力来打分,Opus 5.5 用 1/30 的成本跑赢人类专家,数字挺有意思的。
OpenAI publishes 4,000 AI-solved math problems as Mistral previews 1T-parameter Large 4
2 Sep 2026. The day's biggest news is OpenAI's release of 722 new mathematical research results produced by an internal frontier model — roughly 4,000 problems solved at about three hours of ChatGPT Pro compute per problem, published via the openai/math GitHub repo after consultation with an independent advisory group at Princeton's Institute for Advanced Study. Mistral previewed Mistral Large 4, a 1T-parameter, 49B-activated multimodal open-weights model it calls frontier-level for open models, with final weights due by month end. Capital and compute dominated industry news: DeepSeek is reportedly raising at least $12 billion at a possible $71 billion valuation, earmarked for at least 160,000 Huawei Ascend 950DT accelerators in Inner Mongolia, while SEC filings revealed ~80% of Anthropic's $518 billion compute commitments — including a $161.2 billion Broadcom TPU lease — are fixed payments regardless of usage. On benchmarks, DeepSeek's V4.1 Flash scored 72.9% on ARC-AGI-2 and 94.5% on ARC-AGI-1, beating V4 Flash at ~250% higher per-task cost, and the TasteVal benchmark found Opus 5.5 exceeding a 24-expert human baseline. Compute allocation is shifting decisively toward post-training: frontier labs' pre-training share will fall from 67% in Q1 2024 to 7% by Q4 2026, with reinforcement learning rising to 55% and inference to 38%.
Section Synthesis & Key Takeaways: Today's briefing curates 369 high-signal developments across 134 global and regional sources.
Models (5 picks): 「TasteVal 基准发布:量化 AI 的实验研究品味,Opus 5.5 超过人类专家」(这个 TasteVal 把研究品味换算成省了多少算力来打分,Opus 5.5 用 1/30 的成本跑赢人类专家,数字挺有意思的); 「Mistral 发布 Mistral Large 4 预览版:1T 参数、49B 激活的多模态开源模型」(Mistral 放出了 1T 总参、49B 激活的多模态模型预览版,说是欧美最强开源权重模型,月底连权重一起放出来,做开源部署的可以盯着).
Products (5 picks): 「OpenAI Decisions API for developers」(The API enables more efficient workflow automation by making decisions based on visual and textual inputs. Developers can use it to categorize large datasets, flag risky tool calls, and determine actions from screenshots, potentially reducing manual oversight); 「Anthropic开放Claude Mythos 5.1网络安全访问」(Anthropic给不同信任级别的安全团队分级开放Claude Mythos 5.1,GLM-5.3已被证明能生成接近Mythos的漏洞).
Industry (5 picks): 「Bloomberg:DeepSeek 拟融资至少 120 亿美元,估值或达 710 亿美元」(DeepSeek 这轮可能融到 120 亿美元以上,估值 710 亿,还要买 16 万块华为芯片建数据中心,腾讯宁德时代都跟投了); 「Anthropic 5180 亿美元算力承诺中约 80% 属于闲置也须支付」(Anthropic 5180 亿美元算力账单细节曝光,八成闲置也要付,Broadcom 独占 1612 亿,创始人还设计了 IPO 后不受股东制约的投票结构).
Papers (5 picks): 「OPPD Distillation Method Raises Single-Generation Reasoning Accuracy」(The non-obvious result is that inference-time candidate scoring can be absorbed into the weights, so the sharpening gain survives a single forward pass at deployment. Teams doing RLHF-style training on math or code can stack OPPD on top of GRPO for up to 9.3 extra points, and unlike GRPO it needs no verified reference answers, which matters where reward data is scarce. The loss coefficient gives direct control over the absorbed sharpening exponent); 「研究:特定起始 token 提示可让 base 模型接近 RL 训练后的推理表现」(不用 RL 训练,只在开头加个 "Okay" 就能让 Olmo-3-7B 数学成绩翻近一倍,这篇论文把原因挖得很透,做模型训练的值得看看).
How-to (3 picks): 「Agent research efficiency analysis」(The research introduces a practical framework for optimizing agent architecture by prioritizing verifiability, which could significantly reduce computational waste. Developers and researchers can apply this litmus test to design more efficient agents before scaling, potentially saving substantial resources in production environments); 「用 AgentCore 和 OpenClaw 搭建有记忆的 AI 助手」(AWS 出的手把手教程,教你用 OpenClaw 加 AgentCore 的 memory 功能做一个能记住你、越用越懂你的个人助手,适合动手党).
Models
5 itemsMistral 放出了 1T 总参、49B 激活的多模态模型预览版,说是欧美最强开源权重模型,月底连权重一起放出来,做开源部署的可以盯着。
DeepSeek 的 V4.1 Flash 跑分出来了,ARC-AGI-2 拿下 72.9%,比上一代高了 11.5 分,不过每任务成本贵了两倍多,看你怎么权衡。
OpenAI 把未发布模型解的约 4000 道数学题解法开了仓库,每题平均烧三小时 Pro 级思考算力,可以去看它的解题过程。
Products
5 itemsThe API enables more efficient workflow automation by making decisions based on visual and textual inputs. Developers can use it to categorize large datasets, flag risky tool calls, and determine actions from screenshots, potentially reducing manual oversight.
Anthropic给不同信任级别的安全团队分级开放Claude Mythos 5.1,GLM-5.3已被证明能生成接近Mythos的漏洞。
OpenAI 把搜索直接内置进 API 了,Artificial Analysis 实测便宜过 Perplexity,但多跳检索还打不过 Octen,做联网应用前先看看这份数据。
The pitch reframes enterprise agent reliability as an integration-layer problem: customers customize their CRMs and ERPs beyond recognition, so one broken write can kill a pilot or renewal. Teams shipping customer-facing agents in regulated sales, finance or HR stacks can use Ampersand's config-driven mappings instead of hand-maintaining per-customer connectors. No Southeast Asian availability or pricing was announced, and direct comparisons would be against building on raw vendor APIs.
Google 的 Nano Banana 2.1 上线了,1K 图约 3 美分一张,4K 也才 7 美分,做图的成本又降了。
Industry
5 itemsDeepSeek 这轮可能融到 120 亿美元以上,估值 710 亿,还要买 16 万块华为芯片建数据中心,腾讯宁德时代都跟投了。
Anthropic 5180 亿美元算力账单细节曝光,八成闲置也要付,Broadcom 独占 1612 亿,创始人还设计了 IPO 后不受股东制约的投票结构。
OpenAI 把内部模型算出来的新数学结果开源到 GitHub 了,还找了高等研究院顾问组把关发布方式,可以直接去看仓库内容。
This shift signals a fundamental change in AI resource allocation, moving from raw model creation to optimization and deployment. For developers and businesses in ASEAN, this means more accessible, refined AI models are on the horizon, potentially reducing costs for implementing advanced AI solutions compared to training new models from scratch.
Nvidia 把租借担保改成云收入分成了,The Information 报的,360 亿美元的盘子怎么玩,这篇讲得挺清楚。
Papers
5 itemsThe non-obvious result is that inference-time candidate scoring can be absorbed into the weights, so the sharpening gain survives a single forward pass at deployment. Teams doing RLHF-style training on math or code can stack OPPD on top of GRPO for up to 9.3 extra points, and unlike GRPO it needs no verified reference answers, which matters where reward data is scarce. The loss coefficient gives direct control over the absorbed sharpening exponent.
不用 RL 训练,只在开头加个 "Okay" 就能让 Olmo-3-7B 数学成绩翻近一倍,这篇论文把原因挖得很透,做模型训练的值得看看。
The key insight is that near-fixed-point recurrent states make much of the compute spent on full backpropagation and full KV caches redundant, cutting costs across training, decoding and RL at once. Practitioners building efficient inference stacks or RL pipelines for recurrent models can apply KV sharing and distilled prefill directly. Compared with Huginn's fixed prior, the learned prior plus orthogonal injection improves perplexity at every scale tested up to 1.6B parameters.
LeCun 系 JEPA 路线的分层版本来了,AntMaze 成功率从 18% 干到 73%,做机器人规划和世界模型的值得看看这篇论文。
开源网页智能体训练时奖励信号太稀疏?这篇 CLIFT 用保形自验证把裁判反馈变成可复用信号,测试时还能不调 API 就做轨迹筛选,WebArena Infinity 上开源里最強。
How-to
3 itemsThe research introduces a practical framework for optimizing agent architecture by prioritizing verifiability, which could significantly reduce computational waste. Developers and researchers can apply this litmus test to design more efficient agents before scaling, potentially saving substantial resources in production environments.
AWS 出的手把手教程,教你用 OpenClaw 加 AgentCore 的 memory 功能做一个能记住你、越用越懂你的个人助手,适合动手党。
The deck compresses hardware-level tradeoffs that usually take weeks of trial and error into one reusable resource, and attribution-based reuse makes it usable directly for internal team training or meetup talks. Anyone deploying local inference on constrained hardware can use it to pick between quantization levels or memory setups. Pair it with a llama.cpp install to follow the examples hands-on.