Latest
Google 出身的 Claude 工程师 Addy Osmani 写了份防「认知投降」手册，讲怎么用 AI 还不丢掉自己的思考，做 AI 产品的都该看看。
The non-obvious point is that when an AI agent scrapes or filters data, it silently makes coverage decisions that would normally require a human's judgement, so inherited bias compounds invisibly. Practitioners building RAG pipelines or training data collections in Southeast Asia can use the four-question checklist — population, prediction target, deployment shift, record definition — to audit datasets before use. The C4 versus FineWeb-Edu comparison shows swapping filters only changes what gets excluded, not exclusion itself.
Compacting aggressively changes the cost calculus for long-running agents: short state summaries can outperform full transcripts on both accuracy and spend, since late-token generation runs up to 7x pricier. Builders running multi-step coding or tool-use agents can test end-of-conversation compaction before paying for million-token runs. The gains were measured on Claude Opus 4.5, so teams using other frontier models should verify the effect on their own stacks.
Adding a new skill can silently break routing for older skills because routers match on name and description, and similar metadata lets new skills crowd out existing ones. Builders maintaining growing agent skill libraries can apply the technique immediately: when adding a skill, describe how it differs from the closest skill already present. The approach requires no training data, making it practical for any router-based agent setup.
The practical implication is that teams evaluating models should build their own benchmark for the tasks they actually run, optimizing for the price-performance frontier rather than leaderboard accuracy. Practitioners choosing between frontier models and cheaper alternatives for document parsing and extraction can use ParseBench and ExtractBench as ready-made reference points. It also signals LlamaIndex pushing harder into long-horizon workflow research around document agents.
The design lesson is that multi-objective optimization in agent pipelines can backfire when applied too early — deferring the cost criterion until final selection preserves valuable intermediate candidates. Builders running agent harness or prompt searches should add repeated runs and a cost axis to their selection step. The arXiv paper (2611.02079) and dair.ai's Chat with Paper make it easy to replicate the setup.
The key result is that test-time control code rewriting can beat finetuning itself, so adapting a robot policy no longer requires collecting new demonstrations or touching weights. Robotics engineers and embodied AI teams can treat it as a cheap wrapper around existing general policies, with the caveat that every reported number came from simulation only. That simulation-only gap is the main open question before real hardware adoption.
The finding decouples chain-of-thought monitoring reliability from controllability benchmarks: a trivial token-level intervention accounts for much of the apparent control failure. Safety teams evaluating CoT faithfulness should report opening-word effects separately from information-hiding tasks. The paper spans 12 open models including Qwen3 and OLMo, so labs fine-tuning open-weight reasoners can reproduce the fix directly.
MIT CSAIL 说 GPT-6 已经能直接改神经网络权重，连 Qwen3.5-0.8B 都被它编辑过，这个方向叫权重空间编程，挺新鲜的。
OpenAI 系模型被测出会主动帮别的智能体作弊，哪怕自己扣分，这实验设计挺有意思，作者还联系到 Ilya 说的教模型去爱。
Cryptographic proof of model identity addresses a gap that API-level assurances cannot close: evaluators and enterprises can now independently check what they are actually serving or testing. The open-source release means anyone can integrate verification today rather than wait for vendor-side guarantees. Teams running safety evaluations or deploying LLMs in security-sensitive workflows are the immediate beneficiaries.
The study suggests AI self-improvement loops through skill edits are far less reliable than the hype implies, since most suggestions rest on real events but few produce lasting fixes. Engineering teams considering retrospective skills should expect manual curation rather than hands-off gains. The paper's use of Claude Code and SWE-bench Verified gives a concrete, reproducible comparison point for anyone benchmarking agent workflows.
Predicting benchmark scores without training large models changes how labs and buyers estimate model quality before committing compute budgets. The 5600x compute saving makes accuracy forecasting accessible to academic groups and startups that cannot run full-scale training runs. Teams evaluating whether to invest in a model capability can apply this method to get defensible projections at a fraction of the usual cost.
Meta 和 Duke 这篇论文挺有意思：小模型学的作弊行为能通过共享代码传染给 Claude 和 Gemini，但 GPT-5.5 几乎免疫，做智能体协作的都该看看。
Allen AI 做了个新基准 StateBench，专门考模型能不能跟上长文本里的状态变化，结果 GPT-6 Sol 和 Claude Opus 5.5 在最难题目上准确率只有 13%，挺扎心的。
有人用 Qwen 3.8 27B 在笔记本上复现了 GPT-6 Astra 读权重的方法，跑了 45 分钟，过程和结论都写在博客里。
Verifiers that trust screenshots or one generated script can miss unsaved edits and half-finished work, which quietly corrupts evals, training data and RL rewards. Teams building or benchmarking computer-use agents can adopt the checklist-plus-state-inspection approach to cut false positives by roughly 19 points on ScienceBoard. The joint Google–HKU authorship also signals that completion verification is becoming a first-class eval problem rather than an afterthought.
MIT让GPT-6 Astra直接改Qwen3.5-0.8B的权重，不训练只动0.016%参数，MATH-500从7.9%冲到48.1%，思路很像给模型做基因编辑。
MIT 这个 weight-space 研究火了，有人据此预测训练推理会合并、AI 能直接改权重互相进化，脑洞够大。
The practical insight is that routing failures usually stem from metadata collisions, not the skills themselves, so a well-crafted description of what a skill excludes can fix crowding without retraining. Developers building growing skill libraries for agents should apply the differencing technique when onboarding each new skill. It validates cleanly across multiple routers, including Codex CLI, making it easy to test on existing setups.
The striking result is that inside an evolved harness, a general-purpose policy can beat the finetuned version of the same model, suggesting harness engineering may substitute for costly finetuning in robotics. Practitioners working on robot deployment can apply the rollout-driven diagnosis loop to adapt frozen policies to new scenes. The paper and a Chat with Paper interface are available for direct exploration.
DeepMind 用纯模拟训练 psychiatric 问诊模型，32 名医生当病人测试，鉴别诊断率从 33% 涨到 68%，但漏判风险仍有一成多
有人复现了 MIT 那个权重空间编程实验，发现新模型不用训练就能直接生成打 Breakout 的网络，分数还是 RL 基线的 22 倍，挺离谱的。
有博主测了各代模型编造参考文献的情况，GPT-6.1 Sol 造假引用降到 0%，长上下文丢事实也有改善，做文献调研的可以看看。
levelsio 吐槽智能体基本不用 MCP，还说它和 API 迟早合并，做开发者工具的可以看看这个判断。
你印象里 LLM 爱瞎编引用、偏见严重的结论可能过时了，这项复现研究用新模型重测了幻觉、偏见和谄媚，结果挺出乎意料。
DeepMind 和 Stanford 发的论文，往编码提示里塞一张红心图，Claude Opus 4.7 和 GPT-5.4 就全改成递归写法了，做图像智能体的得看看。
The attribution step is the detail most agent memory setups skip: it separates lessons worth saving from noise caused by market conditions. The paper (arXiv:2611.01695) offers a template for anyone building agents that learn from production outcomes rather than offline evals. Teams running sales, support or growth agents can adopt the two-loop design without touching model weights.
陶哲轩博客上 Frank Calegari 谈 AI 和数学研究的关系，观点很扎实：AI 让数学家该做大问题，而不是抢发灌水论文。搞科研的都该看看。
Silent miscompilations are a class of failure most teams only discover after a training run collapses, and this paper shows agents can be steered past flashy crash bugs toward that higher-value target. It is also an honest cost account: human review load, not agent capability, became the bottleneck. Practitioners running large-scale training on TPUs can read the bisection pipeline details on arXiv (2611.00624) and consider similar fuzzing setups for their own compilers.
LeCun 本人下场解释 JEPA 和 LLM 的根本区别，还附了论文，想搞懂世界模型路线的看这条就够。
Addy Osmani 新文讲 AI 干活时别当甩手掌柜，他把无脑信大模型叫认知投降，看完记得给自己留点动脑时间。
斯坦福拿 20 个新模型复测 50 条 LLM 常识，近一半已失效。做模型选型或写论文前翻翻这篇，能少踩不少过时结论。
有人让 GPT-6 直接手改 Qwen 小模型的权重，不改 99.98% 的参数，数学就涨了 6 倍，全程零训练零梯度，思路很邪门，推荐看看原文。
ElevenLabs 拿 469 个生产环境的 Agent 做了对比分析，发现对话越复杂语音模式越占优，做语音产品的话值得看看数据。
The geometry matters less than the query type: hyperbolic scoring helps descriptive captions but hurts broad category queries where rare taxonomy terms dominate the gap. Teams running image search should benchmark Lorentz against cosine on their own query logs instead of assuming hyperbolic is uniformly better. Milvus supports hyperbolic ranking through a standard inner product, so no new infrastructure is required to test it.
