Sunday, November 1, 2026

Pulse (last 3d) · Research 94 · Agents 71 · Products 52 · Industry 40 · Models 34 · Legal 25 · Open Source 23 · Enterprise 22 · Media 21 · Infra 18

Trending · OpenAI 22 · Anthropic 19 · Claude Code 15 · Codex 13 · Hugging Face 13 · Google 12 · Cursor 10 · Nvidia 8 · The Verge 8 · Claude 7 · MCP 7 · FDA 5

---

• Most agent memory rot comes from misfiled updates, which answer-level evals cannot see. arxiv.org
• MGB/DFCI agentic irAE extraction scored F1 0.72 to 0.79 in prospective use and also won a randomized crossover. pubmed.ncbi.nlm.nih.gov
• With matched capabilities, agents behave differently over MCP than over a CLI. arxiv.org
• Engine speed flags measurably lower Qwen 3.8 Flash Next quality on real tasks. reddit.com
• The best coding agent writes mergeable PRs only 45.5% of the time on FrontierCode 2.0. x.com
• A chain of web flaws allowed ChatGPT account takeover; OpenAI fixed it and paid a $5,000 bounty. parameter.ai
• Anthropic is reportedly blocking users whose requests touch on training larger models (single source). twitter.com

• How we took over ChatGPT accounts - A chain of web flaws, including JSONP abuse, enabled takeover; OpenAI paid $5,000. parameter.ai | simonwillison.net
• Anthropic blocks users over larger-model training - Accounts reportedly blocked for requests touching on training larger models (single source). twitter.com
• OpenAI lowers funding target - OpenAI reportedly cut the target for its new round (single source). techcrunch.com
• Separately trained JEPAs converge - LeCun and Balestriero reportedly prove a shared representation and 42x fewer alignment samples (single source). x.com
• Ideogram Precise Edit - Edits images up to 8K at original pixels, with sharper small text and logos. producthunt.com
• ilse - A toolbar over a running React app that sends visual edits to an existing coding agent. producthunt.com
• NPU compiled inside Qwen 3.5-4B weights - The author who embedded a RISC-V CPU now runs stories260k inside the model. v.redd.it
• bootgpt - A working language model that fits in a boot sector. github.com
• Air-gap leakage via RAM radio signals (2024) - Modulated RAM emissions leak secrets from air-gapped machines. arxiv.org
• NAACP v. xAI proceeds - A judge let the data center pollution lawsuit move forward. techcrunch.com
• Apple on AI strategy - The CEO says Apple will not ship AI for AI's sake. techcrunch.com
• Alibaba Cloud Amsterdam region - Alibaba's first public cloud region in the Netherlands is open. x.com
• datasette-llm 0.1a9 - A new alpha of Simon Willison's LLM plugin for Datasette. simonwillison.net
• Why frontier labs want biology - Wet-lab ground truth closes their training loop, but biotech is unlikely to become core within 3 to 5 years. x.com
• Neonatal brain MRI pathology detection - A two-stage hierarchical deep learning model detects neonatal brain pathologies on MRI. frontiersin.org

Both memory papers move the failure, and the fix, from retrieval to the moment a fact is stored.

Stale memory comes from misfiled updates, not missed ones A Berkeley audit (arXiv 2610.10265) checks agent memory before generation. It finds that key assigners with 0.98 to 0.996 recall still file updates under the wrong key, silently deleting current values (preprint, unreplicated). Answer-level evals miss this, so an agent on patient records can score well while serving a superseded medication or variant call. arxiv.org

A small-model cascade can own the write decision Dual-Layer Agentic Memory (arXiv 2608.22215) sends each incoming item through a 1.7B/8B model cascade that labels it non-write, write-new or write-update before anything is stored (preprint, unreplicated). Read alongside the Berkeley audit, it makes write-time routing the control point to instrument and test. arxiv.org

How an agent is connected and how its output survives review now decide how it gets evaluated.

MCP and CLI produce different agent behavior at identical capability A controlled study (arXiv 2610.34298) gives language agents the same tool capabilities over MCP or a CLI and finds their behavior differs by interface (preprint, unreplicated). That makes the interface an experimental variable, so validation run over one interface does not carry over to the other. arxiv.org

Coding agents clear the merge bar less than half the time Cognition's FrontierCode 2.0 scores whether an agent's PRs would actually be merged, and the top system reaches 45.5% (vendor benchmark). Merge rate tracks production value more closely than pass@k, and the ceiling shows how much review load agents still create. x.com

MCP servers are starting to ship with citations and quality scores PartsPlan returns each datasheet value with its source document and page, and mcpscore rates any MCP server from 0 to 100 on protocol conformance, tool quality, security and auth (press release). Page-level provenance and auth scoring are what a regulated team would require before connecting an MCP server to clinical data. producthunt.com | producthunt.com

Local open-weight models are now good enough that throughput tricks, more than model size, decide output quality.

Engine speed flags measurably degrade MoE output The Basalt developer tested MoE routing-bias and tail-skip flags on Qwen 3.8 Flash Next across real tasks and measured quality losses from settings presented as free speedups (single source). Any on-prem deployment therefore needs evals run on its exact engine configuration, not just the model card. reddit.com

Bit-exact MoE inference runs on a 12GB consumer GPU Betelgeuse claims bit-exact Qwen 3.8 Flash Next inference at 22 tok/s sustained, peaking at 48, on an RTX 3060 12GB with 16GB of DDR4 (single source). Bit-exactness is what matters, because it gets the offload speed without the quality drift the Basalt tests measured. v.redd.it

27B Qwen models handle agentic coding on 64GB machines and in teams After ten days of agentic coding with a 27B Qwen on a 64GB M4 Max, one practitioner reports the memory tier is now usable (single source). Separately, Memdoor reports that a team of five Qwen3.8-27B models beats Fable 5 on LiveCodeBench Hard (vendor benchmark). reddit.com | github.com

The agent deployments worth copying measure themselves prospectively and build the human into the loop.

MGB/DFCI agentic irAE extraction holds up in silent deployment and a randomized trial An agentic LLM from Mass General Brigham and Dana-Farber that extracts immune-related adverse events scored F1 0.72 to 0.79 in prospective silent deployment. In a randomized crossover it also made abstraction faster and more accurate (Nature Cancer). Silent deployment plus a randomized crossover is now the evidence standard clinical extraction pipelines can be held to. pubmed.ncbi.nlm.nih.gov

Long-horizon agents work better when they prompt the human Latent Space profiles Arca's agents for long-horizon financial work, which are built so that agents prompt people for decisions instead of people prompting agents. The same pattern fits clinical review, where the agent runs the workflow and calls in the human at set checkpoints. latent.space

13 items: discussion threads, open questions, promotions, and posts with no substantive content.