Sunday, November 8, 2026

Pulse (last 3d) · Research 77 · Agents 74 · Products 50 · Models 35 · Industry 32 · Legal 24 · Infra 23 · Open Source 22 · Enterprise 20 · Releases 15

Trending · Anthropic 18 · OpenAI 14 · Claude 12 · Codex 11 · Claude Code 10 · The Verge 8 · FDA 5 · Microsoft 5 · Nvidia 5 · ChatGPT 4 · Cursor 4 · Git 4

---

• Type-blind context compaction cuts retained safety rules from 53% after one round to 10% after five. arxiv.org
• Claude Code's "Don't Ask Again" approvals put passwords and keys into 1,122 public GitHub files. apiiro.com
• PSI says a non-public Anthropic model proved the Aaronson–Ambainis conjecture, and it published a Lean formalization. reddit.com
• A METR report on Anthropic automating its own AI development is alarming researchers. theverge.com
• Geneva University Hospitals published a validation and accreditation playbook for clinical digital pathology tools. doi.org
• Meta's Muse avatar answers in about 870 ms with full-body motion and beats HeyGen in rater tests. producthunt.com
• Exa cut scheduled agent web monitoring from $15 to $2 per 1,000 runs. producthunt.com

• Run My Agents - free MIT tool warns a Claude Code session before it overwrites another session's recent edits. producthunt.com
• Maskink - self-hosted redaction swaps secrets, IPs and names for consistent placeholders before LLM calls. producthunt.com
• Professor loses lecture notes to AI - an AI tool deleted a year of a professor's notes while it ran unattended. dexerto.com
• LiveReview - free, self-hostable PR risk scoring that auto-approves the lowest-risk 30 to 50% of changes. producthunt.com
• MapX - local code graph lets LLMs query symbols and dependencies without re-reading files. producthunt.com
• Goodbye Claude, Hello Pi - one developer's case for moving from Claude Code to the customizable Pi harness. caleb.software
• Instagram Edits for Desktop - free Mac video editor that exposes MCP so AI assistants can assemble rough cuts. producthunt.com
• Tencent Hy4 - scores 35 on Artificial Analysis Intelligence Index v5, second in China behind Kimi K3.1 Pro (46). x.com
• FrontierPhysics - a new benchmark built by 70 physicists over three months. x.com
• Microsoft Research AI Frontiers shut down - the closure revives debate over big tech narrowing exploratory AI research. hunch.net
• Nvidia chips to an empty office - the US approved Nvidia sales to a vacant Chinese address, exposing gaps in export-control enforcement. theverge.com
• Megaspeed spy memo - a US intelligence memo says a Megaspeed executive was recruited to spy on Jensen Huang (single source). techcrunch.com
• Data centers after the midterms - post-election politics create new headwinds for Musk and AI data-center buildouts. techcrunch.com
• UK AI productivity doubts - the Bank of England doubts an AI productivity boom while UK banks keep hiring. theregister.com | hotminute.co.uk
• Consumer EEG headsets - wearable makers are harvesting brain data to train AI without telling users. science.org
• Rex (Beta) - a multiplayer terminal multiplexer with native and web apps, from the creators of Ghostty. producthunt.com
• Cue - free, open-source terminal autocomplete built on Fig's specs, currently macOS zsh only. producthunt.com
• HackerTube - a Karpathy-built index of 6,454 videos shared in 20 years of Hacker News comments. producthunt.com
• Geneziz - a local Markdown bookmark archive with semantic search and an MCP server, $99 once. producthunt.com

Compaction, memory consolidation and permission caches are all places where an agent's stored state quietly drifts from what its operator intended.

Summarized context forgets safety rules first Across 20 production agent configurations, Claude Code's /compact on Sonnet 4.6 kept 53% of safety rules after one compaction round and 10% after five (CIKM 2026). Type-aware retention, which handles stored constraints differently from task history, fixes the loss, and that matters for any long-running agent that carries PHI-handling rules in its context. arxiv.org

Anthropic's answer is letting Claude rewrite its own memory Anthropic's engineering blog describes "dreaming," a process in which Claude reorganizes its stored memory, corrects earlier mistakes and improves retrieval. Critics describe it as a model writing more markdown files, limited by the quality of its pretraining, and the compaction result shows that self-edited memory can also drop constraints. x.com | x.com

Approved commands are leaking credentials into public repos Apiiro found 1,122 public GitHub files where Claude Code's settings.local.json, filled in by "Don't Ask Again" approvals, held passwords or API keys. The permission cache works as a secret store that nobody designed as one, so it needs the same gitignore and secret-scanning coverage as .env files. apiiro.com

Each item turns on whether AI-produced work can be checked by something other than the people it is replacing.

A Lean-checked proof of an 18-year-old quantum conjecture PSI says its system, running a non-public Anthropic model, proved the Aaronson–Ambainis conjecture, and it published a Lean formalization (company claim). The formal proof can be checked without trusting the lab, and the same approach works for any AI output that can be written as a specification. reddit.com

METR's report on Anthropic's automated AI development is alarming researchers A METR report on how far Anthropic has automated its own AI development has alarmed outside researchers (single source). Faster automated R&D shortens the gap between model generations, and each new generation is one more a validated clinical pipeline has to absorb. theverge.com

Scientific models written as code make rival hypotheses comparable A guest post on Anthropic's Science Blog shows AI writing, fitting and revising scientific models as executable code to compare competing explanations. Models in code can be diffed and rerun, so they can be reviewed in a way a narrative hypothesis cannot. x.com

Geneva publishes an accreditation playbook for clinical digital pathology HUG (Geneva University Hospitals) published an audit-ready validation and accreditation playbook for digital pathology tools, based on about 230,000 slides a year of primary diagnosis. Templates written by the operators themselves are rare, and this one shows how to take pathology tooling through accreditation at clinical volume. doi.org

Three launches cut latency or price by large multiples, so these capabilities now cost less to run than to validate.

Meta's Muse avatar answers in 870 ms with its whole body Muse Realtime Avatar drives voice, face, hands and posture from a single speech stream, starts replying in about 870 ms and keeps the character consistent from turn to turn. In Meta's live-call tests, raters preferred it 78% of the time over Runway Characters and 88% over HeyGen LiveAvatar (vendor benchmark). producthunt.com

Gemini Omni 1.2 Flash is 30% cheaper and makes 15-second clips Gemini Omni 1.2 Flash is now generally available at $0.07 per second for 720p, 30% below Omni 1.1, with clips of up to 15 seconds instead of 10 (vendor pricing). Multi-turn editing without regenerating and upscaling to 1080p or 4K turn it from a demo tool into a production video component. producthunt.com

Exa cuts the price of scheduled web monitoring by 87% Exa Monitors v2 runs agent web searches on a schedule, remembers what it has already sent and calls a webhook only when new, relevant results appear, at $2 per 1,000 runs instead of $15 (vendor pricing). At that price, continuous literature and competitor monitoring costs less than the engineer who would triage the results. producthunt.com

19 items: hobby projects, opinion threads, off-topic posts and social posts with no content.