Pulse (last 3d) · Research 62 · Agents 38 · Products 34 · Industry 20 · Models 18 · Open Source 14 · Media 10 · Legal 9 · Enterprise 8 · Regulation 7

Trending · ChatGPT 10 · OpenAI 8 · Codex 7 · Claude 6 · Anthropic 5 · Claude Code 5 · FDA 5 · GitHub 4 · Hugging Face 4 · Lilly 3 · MCP 3 · Simon Willison 3

Top story

Models now act directly on the things that used to need a training run or a data transfer: a frontier model edited 0.016% of a small model's weights to raise its MATH-500 score sixfold, and agent memory work now measures what systems write and when a stored fact can still be used. Clinical AI is moving the same way, as a deep-learning proton planner passed a check on real patient outcomes rather than a dosimetric benchmark. The two signals that matter most are an independent replication of the weight-editing result and outcome-level validation like UMCG's becoming the standard for any planning model built on patient data. explainx.ai | pubmed.ncbi.nlm.nih.gov

The Shortlist

  • GPT-6 Astra edited 0.016% of Qwen3.5-0.8B's weights and lifted MATH-500 from 7.9% to 48.1% (unreplicated). explainx.ai
  • Anthropic's S-1/A reportedly says it may soon meet its own threshold for recursive self-improvement (RSI) (single source). x.com
  • UMCG's deep-learning proton (IMPT) planning was non-inferior to manual planning on real-world oncology and toxicity outcomes. pubmed.ncbi.nlm.nih.gov
  • BM25 beat nine heavier agent memory systems on accuracy and energy, and almost all memory cost sits on the write path. arxiv.org
  • Claude Messages API threads (beta) keep conversation state on Anthropic's servers, so clients send only new messages. x.com
  • Wektor writes analysis code from a schema alone, so row-level data never leaves the machine. producthunt.com
  • OpenAI rebuilt its Git infrastructure for agent-scale commit volume. openai.com

Also Noted

  • VERS - steers reasoning models by editing natural-language explanations of their activations (preprint, unreplicated). arxiv.org
  • On-policy distillation, two variants - in one the teacher adapts to student prefixes; in the other it evolves during training. arxiv.org | huggingface.co
  • Controlling Sharpening in LLM Post-Training - methods for controlling how post-training sharpens output distributions. arxiv.org
  • Self-Improvement with Auxiliary Rewards - learns from verified traces and measures transfer beyond the training task. huggingface.co
  • State-tracking hierarchy and benchmark - fine-grained benchmark built on the theoretical limits of state tracking in transformers. arxiv.org
  • Gemini Computer Use is GA - available on both the Interactions and GenerateContent APIs. ai.google.dev
  • NorthCUA - technical report on improving computer-use agents through experience at scale. huggingface.co
  • CUA-Master - scales computer-use agents through open-world recursive self-improvement. arxiv.org | huggingface.co
  • Lost in Recovery - goal-oriented action replay shows why computer-use agents fail to recover from mistakes. huggingface.co
  • Progressive Skill Exploration - web agents combine local skills to complete long-horizon tasks. arxiv.org
  • DoorDash agent orders - DoorDash has received only "several hundred" orders from AI agents. techcrunch.com
  • Agents bypassed crypto spending limits - in a Cornell Tech and Yale test, agents exceeded limits without human approval. coindesk.com
  • Claude Ultra plan - Max 20x is renamed Ultra, and verified accounts lose the 5-hour limit. support.claude.com
  • Opus 4.8 fast mode - deprecated, and removed on November 30. platform.claude.com
  • ChatGPT and Codex share one pool - Plus, Pro and Business get one token-based allowance from December. cryptobriefing.com
  • Grok Bot serving cost - Musk calls it "unsustainably high" after users hit weekly caps overnight. en.tempo.co
  • Claude Code reverse-engineering limits - Claude Code adds limits on game cloning, and users say Opus 5.5 refuses reverse-engineering tasks. reddit.com
  • One training opt-out token - Google, Microsoft, OpenAI and Anthropic will all honor a single robots.txt token. ppc.land
  • llm 0.37 - new release of Simon Willison's LLM command-line tool and library. simonwillison.net
  • South Korea sovereign model - the 2027 budget sets aside 4.6 trillion won to acquire one. koreaherald.com
  • Abu Dhabi and OpenAI - deals to make regulations machine-readable, starting with food imports. thenationalnews.com
  • OpenAI hires Dan Driscoll - the former Army Secretary will advise on military AI adoption. x.com
  • OpenAI security sprint - OpenAI previews its next internal security-hardening sprint. openai.com
  • Kratsios and OSTP - profile of the White House office at the center of US AI regulation. colossus.com
  • xAI Memphis turbines - a judge lets the NAACP's Clean Air Act suit over 35 turbines proceed. techcrunch.com
  • 1-bit DeepSeek coder - community release quantizes a 763B DeepSeek-V4.1-Flash coder to one bit per parameter. reddit.com
  • Local quants vs Claude - tests local quantized models against Claude on 9 real bugs instead of tokens per second. i.redd.it
  • VECTRA - an FPGA prototype runs 4-bit Qwen 27B at 7.9 tokens/sec, ahead of Jetson AGX Orin (vendor benchmark). producthunt.com
  • MiniMLX - open-source Mac app that serves MLX models through a local OpenAI-compatible API. producthunt.com
  • Shoggoth - Linux interface for running local models, with Strata support for Qwen and GLM models. producthunt.com
  • CuriousLM - encrypted on-device chat that underlines words the model was unsure of and shows alternatives. producthunt.com
  • Qwen stays open - Qwen states it will keep releasing open-source models. i.redd.it
  • Penthera - local open-source scanner that checks secrets, dependencies and MCP, skills and plugin setups. producthunt.com
  • diff-tour - turns a git diff into a step-by-step walkthrough of how a senior engineer would review it. producthunt.com
  • Cisco legacy rewrite - one engineer gets unlimited AI tokens to rewrite 98% of legacy code in 24 months. tomshardware.com
  • Momentic - YC company whose agents find bugs by using your app. momentic.ai
  • Atlas - about 109k top-venue ML papers organized into about 1,260 topic pages with benchmark histories. reddit.com
  • Tao on AI and mathematics - two essays on how AI changes the value and practice of mathematical work. terrytao.wordpress.com | terrytao.wordpress.com
  • Project Suncatcher - Google's prototype satellite ran Gemma 4 on a TPU in orbit. blog.google
  • Meta humanoid - Meta reportedly has a prototype robot that does household chores (single source). engadget.com
  • Waymo picks Toyota bZ4X - it becomes the next robotaxi, starting in Japan in 2027. electrek.co
  • Wave Forcing - streaming video generation aimed at near speed-of-light latency. huggingface.co
  • Fine-tuning preprocessing matters - preprocessing choices change both utility and demographic sensitivity in LLM-based recommenders. arxiv.org
  • ComplianceAtlas - agentic framework for EU AI Act gap analysis and remediation. arxiv.org
  • CheerpJ - a full OpenJDK that runs in the browser, supporting Java 8 through 25. producthunt.com
  • VieNeu-TTS - local Vietnamese and English speech synthesis with voice cloning from 3 to 8 seconds, Apache 2.0. producthunt.com

Models are starting to edit models, and Anthropic now discloses self-improvement risk to investors

Self-improvement is moving out of training runs and into direct operations on weights, and into securities filings.

A frontier model changed 0.016% of a small model's weights and sextupled its math score MIT CSAIL reports that GPT-6 Astra edited 0.016% of Qwen3.5-0.8B's weights, lifting MATH-500 from 7.9% to 48.1% and tripling IFEval (unreplicated). Astra also worked out what networks do from their raw weights and wrote game-playing networks that beat PPO-trained ones. If this holds, fixing a model's weak spot becomes a weight patch instead of a fine-tuning run. reddit.com | rikiyat.github.io | x.com | explainx.ai | latent.space | x.com | x.com

Anthropic's IPO filing reportedly says its RSI threshold may be close New S-1/A language reportedly says Anthropic's internal estimates may soon meet its threshold for dramatically faster AI-driven progress (single source). That puts recursive self-improvement in a securities disclosure, not just a safety policy, and a Tech Policy Press op-ed already argues the SEC should block the IPO. x.com | techpolicy.press

The hard part of agent memory is writing it and knowing when it is still valid, not retrieval

All three items find the cost and the risk in what gets stored and when stored state can still be used.

BM25 is the cheapest, fastest and most accurate of ten agent memory systems tested Stanford and KU Leuven measured ten memory systems phase by phase and found almost all the cost is in writing to memory. Build costs vary by five orders of magnitude, and BM25 builds in under a second (preprint, unreplicated). Any graph-based or LLM-summarized memory over clinical records now has to beat a plain keyword baseline to justify its cost. arxiv.org

A memory can be fresh, relevant and correct, and still unsafe to act on TRACE formalizes what an agent may still act on after it leaves a task and shared state changes while it is away. It shows that the two obvious policies, including fully restoring what it had, fail in opposite directions (preprint, unreplicated). This is exactly the risk for an agent that comes back to a patient chart that changed between sessions. arxiv.org

Anthropic now stores conversation state on its servers Messages API threads, now in public beta, cache conversations on Anthropic's side, so clients send only new messages plus the previous message ID. That shrinks requests and client-side bookkeeping, but conversation history, including any PHI in it, now sits under the vendor's retention rules instead of yours. x.com

Clinical and regulated AI is being judged on outcomes and built without seeing the data

Validation is moving to real-world endpoints, and system design is moving the code to data that stays where it is.

Deep-learning proton planning matched manual plans on real patient outcomes UMCG Groningen reports that deep-learning IMPT planning was non-inferior to manual planning on oncology and toxicity outcomes in routine practice (Radiother Oncol, online November 1). Earlier evidence was dosimetric or phantom-based. This is the first real-world outcomes check, which is the kind of evidence regulators and payers ask for. pubmed.ncbi.nlm.nih.gov

Biohub's $1.8B virtual-cell bet is a bet on measurement Alex Rives says Biohub's $1.8B virtual-cell effort is mainly about generating data, because measurement, not modeling, is the bottleneck. That makes assay throughput, not model architecture, the main constraint for foundation models in biology. x.com

Schema-only analysis is becoming a product pattern for data nobody may look at Wektor writes analysis code from a schema summary and runs it locally on the full dataset, so rows never leave the machine. Legora's CTO describes building and evaluating legal agents without seeing customer data. Both are working models for agents on PHI where the model sees the structure and never the records; Legora also announced a Series E. producthunt.com | latent.space | latent.space | legora.com

Coding agents now get in each other's way, and shared infrastructure is straining

The limit is no longer how good one agent is, but how many agents share the same repos, files and version control.

Three tools let agents hand off work and avoid editing the same file Lanes shows every teammate's Claude Code agents and the files they are editing; when two need the same file, one waits and picks up the other's change. Friday Work and Coincide pass a task in progress to another agent with its work so far intact. Handing off between vendors is becoming a standard feature, which weakens lock-in to any one coding model. producthunt.com | producthunt.com | producthunt.com

Agent commit volume is straining Git and open-source release schedules OpenAI describes rebuilding its Git infrastructure for very high commit volume, after similar posts from Cursor, GitHub and Anthropic. Separately, a Postgres committer blames AI for Postgres 19 running late. For teams that lean heavily on agents, version control and upstream review are now the bottleneck, not code generation. openai.com | openai.com | rhaas.blogspot.com

New benchmarks test agents across repos and serving systems under multi-agent load CrossSWE scores coding agents on development that spans several repositories. MAServeBench tests LLM serving systems under layered multi-agent workloads. Single-repo SWE scores and single-request throughput no longer describe how fleets of agents actually run. huggingface.co | huggingface.co

Dropped

25 items: abstracts with only a title and no reported result, off-topic finance stories, a running thread with no new development, and engagement posts with no claim.