Latest
OpenAI和Anthropic的顶尖科学家们警告AI自我进化可能导致人类边缘化或灭绝，Anthropic已有26%研发工作由AI完成。
Amazon AGI 这篇 AutoGym 论文挺有意思：一句话就能自动生成训练环境和验证器，做智能体强化学习的人可以省掉大量手工搭环境的活。
OpenAI 首席科学家 Pachocki 联合 Hinton 和 Bengio 写的论文，聊 AI 自己改进自己的风险，还喊话监管层要企业交底，观点很硬核。
Perplexity 和 Nvidia 搞了个沙盒 SPACE，让 9 个模型拿 root 权限试着越狱，108 次全没跑出去，搞智能体安全的可以看看。
The insight is that most reward variation in multi-turn RL comes from downstream noise, so credit assignment should target the pivotal call rather than the full trajectory. Researchers and engineers building tool-using agents can apply nested sampling and contextual-bandit updates to concentrate training on decision-critical states. The BFCL v4 result offers a concrete benchmark for comparing against trajectory-level reward approaches.
The implication is that agent workloads may force a rethink of the decades-old split between OLTP and OLAP stacks, with storage rather than compute as the unification point. Teams building data platforms or agent infrastructure can watch LTAP to see whether it reduces the need to run separate transactional and warehouse systems side by side. Compared with bolt-on sync pipelines between an operational database and a lakehouse, this approach aims to keep one representation per tier under both engines.
Chelsea Finn 团队写了篇博客，讨论机器人怎么走 LLM 后训练的老路，还给出 EXPO-FT 方法，抓鸡蛋这类任务做到了 30 次全成功，只花 19 分钟真机交互。
The takeaway is that hierarchy may be optional for large-scale agent coordination — a shared workspace and message channel can replace the orchestrator pattern most teams default to. Engineers building multi-agent coding systems can study the paper's shared-state design and emergent cooperation practices as an alternative architecture. Benchmarks with GPT-5.6-sol give concrete scaling curves to compare against hierarchical frameworks like LangGraph or CrewAI.
The gap between conceptual recall and experimental reasoning suggests frontier LLMs are still unreliable as lab-design collaborators even when domain quizzes look strong. Subfield-level scores tell bioengineering teams which applications, like diagnostics, are closer to production and which, like synthetic biology, still need human review. Anyone evaluating LLMs for biomedical R&D workflows can use the benchmark as a selection reference.
The book connects neuroevolution theory to the design philosophy behind one of Japan's most-watched AI labs, offering readers the intellectual grounding for Sakana AI's collective-intelligence research. Researchers and engineers interested in evolutionary computation can start with the free online edition before committing to print. With co-authors Risi and Miikkulainen, the coverage spans both classic and current methods in the field.
A more-than-doubling of accuracy on a widely used 8B open model suggests the training recipe, not just model scale, drives reasoning gains. Researchers fine-tuning smaller open models can test DCE + SRCL against OPSD as a cheaper distillation path. Qwen3-8B's popularity in ASEAN developer communities makes the recipe directly reproducible with local hardware.
The reframing matters because most robot foundation-model effort targets richer visual perception, while this argument suggests tactile and proprioceptive feedback loops deliver more robust grasping at lower cost. Robotics researchers and embodied-AI builders can use it to prioritise sensor fusion over camera-only stacks. It also echoes human evidence: blind people perform many manipulation tasks robots still fail at.
NVIDIA 把社区写的 3.4k 个 Agent Skills 直接变成 RL 训练环境，300 步就把 Qwen 27B 在 Terminal-Bench 上拉高近 5 个点，流水线和负结果都写得很细，做 agent 训练的值得看看仓库。
Test-time scaling, proven in language reasoning, may transfer to robot control without touching model weights — SAIL is an early sign of that shift. Robotics researchers and embodied-AI builders can apply the simulator-plus-VLM-feedback loop to their own stacks, trading compute for reliability. The work competes conceptually with end-to-end approaches like GPT-6 Astra's direct robot operation rather than replacing model training.
McKinsey 拿出了实际数据：AI 参与的候选药物，发现时间砍掉 15% 到 80%，做医药或关注 AI 落地的可以看看。
McKinsey 拿出了实际数据：用了 AI 的候选药物，发现阶段时间缩短 15% 到 80%，不是空谈。
Continual training on distilled skills outperforming a larger-context approach suggests skill libraries can be baked into weights rather than retrieved at inference time. Teams building terminal or coding agents can test whether the same recipe transfers to their own models and checkers. The result also matters for anyone benchmarking open-weight Qwen models against Claude Sonnet 4.6 and GPT-5.4 Mini on agentic tasks.
The cost gap, 62x cheaper than LLM judges, makes calibrated decision models a practical replacement for expensive evaluation pipelines. It also surfaces label errors in three existing benchmarks, which anyone relying on published alignment scores should note. Thresholds do not transfer between benchmarks, so teams adopting this should fit one on roughly 10 labelled items, which lifts median F1 from 0.706 to 0.793.
训练只喂短任务，推理却能搞定32倍长的任务，靠的是外层 harness 做分解，思路挺新颖
dair-ai 又更新一周论文榜了，这次有 7 篇，稀疏化和智能体自组织团队那几篇可以翻翻，找找自己方向的灵感。
The key insight is that budget allocation strategy, not raw model capability, drives research-agent performance — treating experiments as a resource-allocation problem rather than a fixed pipeline. Teams building autonomous research or discovery agents can apply this explore-then-concentrate policy to cut token costs while improving output quality. It compares directly against AutoResearch on FIRE-Bench, giving a concrete benchmark baseline for evaluating competing agentic research frameworks.
Removing the IMU and force-torque sensors cuts cost, weight and failure points in humanoid hardware, which could shape how low-cost robot platforms are designed. Robotics engineers working on balance and control can study the method as a baseline against proprioception-heavy pipelines. It comes out of German labs DFKI and TU Darmstadt rather than a commercial humanoid vendor.
Memory and self-improvement are usually bolted on as separate subsystems; JAZ shows both can just be code the agent writes inside its own loop, with hooks enforcing constraints and monitoring. Builders designing agent harnesses can use this to cut complexity and cost instead of adding memory layers. The paper is arXiv:2609.26891, with an explore guide at academy.dair.ai.
The shift from write-at-task-end to retrieve-then-curate changes how agent memory pipelines should be designed, favoring raw trace storage plus task-specific extraction. Builders working on long-lived agents can apply the approach to improve success rates on benchmarks like tau2-bench. Compared with traditional summarization-based memory, even an untrained curator already performs competitively.
The shift here is from LLMs as proof-checkers to LLMs as autonomous curators of what counts as a meaningful result, with interestingness turned into a measurable objective. Researchers in automated theorem proving and AI-for-math can borrow the quantitative framing to evaluate their own discovery pipelines. It contrasts with purely benchmark-driven evaluation, where correctness exists but relevance does not.
哥大商学院算了笔账：7700万块 GPU 要赚回 3.725 万亿，摊下来每 GPU 小时 5.5 美元，比现在租价还低，基建账能算平。
Most systems marketed as self-improving only automate parts of the improvement loop and cannot yet improve how they improve — a useful filter when evaluating agent claims. AI remains strong at knowledge and reasoning questions but weak at long, multi-step tasks with tools and changing environments, which is where the real bottleneck sits. Researchers and agent builders can use the 5-level framework to benchmark where current systems actually sit.
Epoch AI 新报告讨论了一个有意思的问题：大家都不卷模型了，API 价格还会降吗？关心成本走势的可以看看全文。
Occlusion breaks scene consistency in generated video and embodied simulations, and object permanence offers a cognitive-science-grounded fix. Researchers working on world models, video generation or embodied agents can use the method to reduce object flicker and disappearance during occlusion. Compared with purely data-driven fixes, this frames the problem as a learnable cognitive capacity.
The result suggests that data synthesis by self-play alone can reproduce scaling behaviour normally attributed to real corpora, which challenges assumptions about what pretraining data must contain. Researchers studying synthetic data and training dynamics will want to test whether the predictable loss curves hold at larger compute budgets and on harder tasks.
The design insight is that most memory overhead comes from calling the LLM during storage and retrieval, which a rule-based controller can replace without losing accuracy. Developers building retrieval-heavy agents can adopt the controller pattern to cut latency and memory-construction costs on their own stacks.
有人拿"模型会喊 oops 然后改错"这个训练行为，去质疑 LeCun 那套自回归不能纠错的老论点，争议点很具体。
CSET 出了份报告，说中国走机器人加无人机的具身智能路线冲 AGI，跟美国堆 LLM 的路子不一样，还建议政府怎么应对。
The wide-then-deep practice order, with a verifier gating what gets saved, is the lever — the 18-point gap suggests task sequencing matters more than adding training data. Builders deploying agents in novel apps can replicate the curriculum-actor-verifier loop without collecting new datasets. It compares directly against the costly collect-and-retrain route that most teams default to today.
The benchmark targets a capability gap that reasoning-token budgets do not close: knowing which branch to take before the evidence arrives. Teams building long-running coding or research agents can use Taste-Bench-style forks to evaluate direction-finding rather than only final task success, and the distillation result suggests a practical way to train that judgment into smaller models. It complements SWE-bench Pro instead of duplicating it.
The shift from search loops over small trajectory batches to holistic log analysis changes the economics of prompt optimization for anyone running agents in production. Teams can apply CASD without building validation sets or wiring up environment access, making it practical for real-world agent debugging. Its 22x cost advantage over validation-gated search makes it a concrete alternative to GEPA for budget-conscious projects.