王虹与三维挂谷猜想：深度解析基础数学的AI能力边界

TL;DR - A commentary piece using Wang Hong and Joshua Zahl's 127-page proof of the three-dimensional Kakeya (挂谷) conjecture as a case study for where current AI hits its ceiling in frontier mathematics. It argues AI is an "outstanding student" — great at delegated technical subtasks — but not yet a "master" capable of paradigm-creating research.

• The 3D Kakeya problem is framed as qualitatively harder than the 2D case: rigid rather than flexible geometry, infinitely nested multi-scale/chaotic structure, "sticky" Kakeya configurations, and the limits of existing Fourier/harmonic analysis tooling.
• Wang and Zahl's contribution is described as system-level innovation — abandoning prior research paths and building new machinery — rather than incremental application of known techniques.
• Four stated AI limitations: reuse of existing paradigms instead of inventing new frameworks; no global research strategy judgment; inability to sustain very long, multi-layer, cross-domain logical chains (a 127-page proof spanning four branches of math); and no mathematical intuition for structures with no precedent or template.
• Conclusion positions AI as an accelerator for verification, gap-hunting, symbolic simplification, and pruning dead-end approaches, dominant on competition-style bounded problems, while century-old conjectures remain human territory.
• Note: the piece is opinion/analysis reposted via a WeChat account (数据派THU), not a technical report; no benchmarks or AI experiments are presented.

TL;DR - A WeChat post from 专知 (Zhuanzhi) advertising a curated library of 100+ "AI + military" intelligent-defense reports, doctrine documents, theses, and surveys hosted on its knowledge platform. It is essentially a link-farm/index post, so the value is the reading list itself rather than any new technical result.

• Heavy emphasis on US DoD programs and doctrine: Project Maven algorithmic-warfare architecture, DoD AI acceleration strategy memos, CCA (Collaborative Combat Aircraft), Project Convergence/JADC2, F-35 Block 4 delays, counter-sUAS strategy, and Mosaic/decision-centric warfare reports.
• Large cluster on autonomy and multi-agent systems: UAV swarm tactics via reinforcement learning, manned-unmanned teaming (SEAD mission planning), USV swarm path planning, multi-agent task allocation, and improved DDPG for UAV routing.
• LLM-specific defense angles: offline/on-prem LLMs for intelligence work, USMC lessons on LLM-assisted operational planning, a "military large-model evaluation white paper," BERT + knowledge-graph QA for weapons systems, and prompt-defect taxonomies.
• Also bundles general AI surveys unrelated to defense (GNN foundation models, LLM unlearning, implicit reasoning, multimodal RAG, agent memory, embodied-intelligence theses), indicating the list is a broad content-marketing aggregation; no results, benchmarks, or original analysis are presented.

TL;DR — MLS-Bench is a new benchmark from a multi-university team (Berkeley, Princeton, Tsinghua, CMU, and others) that tests whether LLM agents can discover genuinely new, generalizable ML methods rather than just tune existing ones; five frontier models all failed to beat the strongest human methods. It matters because it reframes "AI for research" evaluation away from saturated coding leaderboards toward real methodological discovery.

• Design: 140 real research tasks across 12 ML areas (LM pre/post-training, vision generation, RL, robotics, ML systems, AI4Science, optimization, causal inference, time series, active learning, trustworthy ML). Each task ships a real codebase, a specific component to improve, ≥3 test conditions for transfer, and ≥3 reproduced strong human methods (including the domain SOTA) run in the identical training/scoring pipeline. Scores are anchored per-metric (weakest reproduced baseline = 0, strongest = 50, theoretical bound = 100) and aggregated across metric → condition → task.
• Headline result: Even when given full implementations of the best human methods and multiple rounds of agentic experimentation, all 5 evaluated frontier models failed to surpass the strongest human method overall. Multi-round iteration improved individual submissions but did not change the conclusion.
• Failure mode: Prompting models to "optimize/debug" outperformed prompting them to "discover a new method." Expert code review found models mostly recombine losses, modules, and tricks from the provided baselines. Removing the parameter-count guard let several models beat human methods purely by scaling capacity — showing unisolated variables can masquerade as automated discovery. Evolutionary search and test-time training overfit visible conditions and hurt held-out ones.
• Deeper bottleneck: In a flexible-compute pretraining experiment (two of three 345M runs converted into a free budget), models generally did worse with more freedom — they failed to allocate compute to high-information experiments or revise plans from intermediate results. Web search and supplied papers/derivations helped only marginally. Cost note: full 140 tasks ≈ 700 H100-hours per candidate; a 30-task subset (~100 H100-hours) covers all 12 domains and has been adopted in official frontier-model release evals.

TL;DR — insitro founder/CEO Daphne Koller argues in a widely-discussed essay that AI's real bottleneck in drug development is not molecule generation but identifying correct disease mechanisms, and she critiques four popular "AI magic wand" narratives (superintelligence, molecular design tools, virtual cells, autonomous AI labs).

• Drug R&D splits into three stages (disease→mechanism, mechanism→drug, drug→patient); most AI work targets stage 2 post-AlphaFold, yet >90% of clinical trials fail mostly because the targeted mechanism is wrong, not the molecule. Historic wins on "undruggable" targets (e.g., KRAS) came from decades of structural biology and new modalities (biologics, siRNA/ASO, gene editing), not AI design.
• Misallocation is visible: 38 targets each carry 50+ programs (e.g., GLP-1 variants), while industry's new targets per year fell from ~100 in 2015 to ~30 in 2024.
• LLM-over-literature reasoning assumes the needed human-biology data already exists; Koller counters that existing cell atlases (hundreds of millions of cells) cover a tiny fraction of perturbation space, are limited to few cell lines, and lack causal perturbation data — and system-level, human-specific diseases (Alzheimer's, ALS) don't transfer from mouse/NHP models.
• Agentic self-driving labs work only with fast, cheap, objective feedback (like compilers for code); the sole ground truth in medicine is human clinical trials — years, millions of dollars, ethically constrained — so accelerating stage 3 without correct mechanisms just fails faster. The productive path is mechanism-derived clinical biomarkers for patient selection, target engagement, and early efficacy signals.

TL;DR - A WeChat write-up of Terence Tao's ICM 2026 lecture "Mathematics in the Age of AI," where he warns that AI can now mass-produce research-level proofs cheaply, creating "proof indigestion" because human verification, exposition, and assimilation cannot keep pace. It matters because it reframes the bottleneck of AI-accelerated science from generation to human understanding.

• Tao splits a proof's lifecycle into six stages — generation, verification, exposition, publication, digestion, canonization — arguing AI plus formal tools (e.g., Lean) accelerate only the first two while the remaining four congest.
• Cited evidence: the First Proof project reported 7 of 10 novel research-level problems solved to publishable quality by at least one team, at $10–$1,000 per problem; erdosproblems.com is reportedly accumulating AI-submitted proofs without enough qualified volunteer verifiers.
• He invokes Goodhart's law (optimizing "solve more problems" diverges from understanding/teaching), notes AI-polished text erases the "friction" that signals hard steps, and observes long correct proofs are easier to generate than short elegant ones.
• Three prescriptions: normalize disclosure of AI assistance; reward refereeing, surveys, and translation of machine proofs over priority races; and a "talk test" — if authors cannot give a clear, expert-level, correctly attributed talk on a result, it should not be published.