Thursday, October 22, 2026

Pulse (last 3d) · Research 146 · Agents 100 · Products 59 · Models 44 · Open Source 38 · Industry 32 · Legal 20 · Media 20 · Enterprise 18 · Releases 16

Trending · Anthropic 20 · Claude Code 19 · Codex 16 · OpenAI 14 · Claude 11 · FDA 9 · Nvidia 8 · Cursor 7 · Gemini 7 · Google 5 · ChatGPT 4 · Hugging Face 4

---

• Claude designed scorpion toxin minibinders over a 64-hour autonomous campaign; 17 of 211 wet-lab tested designs bound. anthropic.com
• OpenAI is acquiring Orchestra Bio to run agentic drug programs inside Rosalind, making the vertical bio stack operational. openai.com
• OpenAI models land on Google Cloud's Model Garden, with GCP spend commitments drawn down. openai.com | x.com
• Agent staleness is trainable: GRPO against a fact-currency reward beats better retrieval, and the gap persists at frontier scale. arxiv.org
• LiteLLM 1.107.0 on PyPI shipped a credential stealer and backdoor; check your gateway pins. safedep.io
• SemiAnalysis puts 2027 at 50GW of new data center capacity, bottlenecked on chips and memory rather than power. newsletter.semianalysis.com
• Anthropic's draft safety bill would let a judge block a release over 50 potential deaths or $1B in damages. techcrunch.com

• NVIDIA and UW IPD release RFD4-Proteina - open protein design model now on Hugging Face. x.com
• OpenAI runs thousands of agents over biological data - fan-out agent search for new leads. openai.com
• FrontierMemory - benchmark scoring agent memory by downstream task impact, not recall. x.com
• Meta's Prometheus supercluster - Ohio site scaling toward a million GPUs across buildings 100km apart. engineering.fb.com
• TorchTPU joins TorchNeuron - PyTorch moves off Nvidia with a config change. techcrunch.com
• Anthropic in talks for a second $45B Nscale deal - compute commitments compounding (single source). techcrunch.com
• Microsoft Guardian Agents - risk-score, throttle or kill other agents; detects 60 agent types. geekwire.com
• Claude Security re-scans repos continuously - public beta for Enterprise as code changes. x.com
• Claude Duet - background proactive agent for inbox and calendar, now a three-lab race. reddit.com | x.com
• Polysemanticity reread - Cammarata argues those neurons were interpretable all along. technologyreview.com
• DeepMind agents beat Minecraft in 27 hours - 2,783 self-written programs, 605 deaths, $520 in tokens. reddit.com
• SWE-DIVE - multi-turn coding agent eval with dynamic user simulation. arxiv.org
• Asymmetric on-policy distillation - teaches in-context exploration rather than imitation. arxiv.org
• Automating ML for science - benchmark on whether frontier agents climb real scientific hills. huggingface.co
• Agents discover linear attention architectures - multi-agent research teams running architecture search. huggingface.co
• LLM confidence scores are unreliable - asking the model for its own certainty does not calibrate. rumple.github.io
• Decision models for evals - structuring eval design around the decision being made. hamel.dev
• Qwen-Audio-3.1-ASR technical report - Alibaba's latest speech recognition stack. huggingface.co
• Cloudflare ties discovery to remediation - vulnerability-to-defense loop built on OpenAI models. openai.com
• Eight compromised TLS certificates - internet-wide alert on certificate trust. arstechnica.com
• PHP security release - hardening plans after the runtime exploit chain. simonwillison.net | reddit.com
• Databricks AI Runtime GA - serverless on-demand GPUs under Unity Catalog governance. x.com
• NVIDIA GTC Berlin - Kumo, Cosmos 3.5, Alpamayo 2 Edge inside a near-$10B Europe commitment. blogs.nvidia.com
• Antirion - Rust LLM gateway adding 0.5ms at 5,000 RPS with PII redaction and per-team budgets (vendor benchmark). producthunt.com
• ActOnce Spend Gateway - purchase limits and approval queues for MCP agents with spending power. producthunt.com
• Aient - correlates traces, logs and metrics into one cause, then opens the fix PR. producthunt.com
• developer-os and Anja - MIT and Apache harnesses adding persistent memory and review discipline to Claude Code. producthunt.com | producthunt.com
• video2ctx - makes video queryable by agents with second-level citations. producthunt.com
• Omnara - open-source alternative to Claude Managed Agents. github.com
• 2-bit draft in VRAM, exact residual in RAM - speculative decoding for models larger than the GPU. reddit.com
• Inference-time abliteration - llama.cpp fork strips refusals live, no fine-tune or extra files. reddit.com
• LongZero - long-context reasoning gains without external training data. arxiv.org
• Circuits of LLM classifiers - interpretability of models used as classifiers. arxiv.org
• Adversarially hardened P vs NP-complete pairs - tests whether rewrites actually raise difficulty. arxiv.org
• Reasoning models read "some" and "or" literally - thinking suppresses pragmatic inference. huggingface.co
• Looped uniform diffusion LMs and pure self-conditioning - two architecture probes beyond autoregression. arxiv.org | arxiv.org
• Quant-Arena - benchmark for LLM agents doing quantitative trading research. arxiv.org
• ChatGPT ad stack and 26-country expansion - over 100,000 advertisers on the platform. openai.com | openai.com
• Canadian publishers sue Anthropic - Postmedia, Torstar, Globe and Mail, Canadian Press file in Ontario. cbc.ca
• Christopher Stubbs joins Anthropic's Long-Term Benefit Trust - physicist added to the governance body. anthropic.com
• NYC targets chatbot doom loops - proposed fines for trapping users in customer-service loops. reddit.com
• Data centers may erase decades of air quality gains - emissions report on buildout externalities. theverge.com
• Tesla and SpaceX form committees on a $3.5T AI merger - board-level exploration only (single source). reddit.com
• Zvi's "Misaligned!" - alignment roundup and argument. thezvi.substack.com
• 30 months on the exponential - progress charted rather than asserted. oneusefulthing.org
• WorldTUI - 360M model trained to emulate a terminal as a world model. simedw.com
• Odyssey drops a fruit fly brain into a generated world - biological controller inside a world model. x.com
• SKY-001 - world model generating CAD designs inside Autodesk Fusion. x.com
• PhysiCalWorld - in-context calibration of physical parameters in world models. arxiv.org
• Computer-use environment synthesis - agent world model generating its own training environments. huggingface.co
• DeLoop - repeated actions as self-supervision for long-horizon agents. huggingface.co
• DONIC and SP-HLM - bimanual tracking control and humanoid loco-manipulation from sparse poses. huggingface.co | arxiv.org
• LIBERO-10 frontier benchmark - reasoning models scored on robot manipulation. arxiv.org
• Embodied agents unified survey - one capability frame from computer-use to robot-use. arxiv.org
• PrismaCoder - open multimodal code model trained on decomposed rubric synthesis. huggingface.co
• TraceForge - distillation using the model's own thinking traces. huggingface.co
• Ming-3.0-TTS-0.5B - half-billion-parameter open text-to-speech release. huggingface.co
• sparqlite - SPARQL 1.1 inside SQLite, written in Rust. simonwillison.net
• Few-read fault localization in analog crossbars - sparse-recovery decoder beats full-cell scans in simulation (preprint, unreplicated). reddit.com
• OpenAI on consistent access standards - policy framing for broad model access. openai.com

Bio went from model demo to multi-day campaign with experimental confirmation, and the platform companies are acquiring the workflow around it.

Claude's 64-hour design campaign produced 17 validated binders out of 211 tested Claude ran an autonomous protein design campaign against scorpion neurotoxins and 17 of 211 lab-tested minibinders met binding criteria, roughly an 8% hit rate. The novelty is not the designs but the duration and closure: the agent held a multi-day research loop and the output went through wet-lab confirmation, which is the first credible template for agent-run assay cycles on your own targets. anthropic.com | reddit.com

OpenAI is buying drug-program orchestration, not models OpenAI is acquiring Orchestra Bio to fold agentic drug-program orchestration into Rosalind. The target is the coordination layer above the models, which means the frontier labs are now competing for the same vertical workflow surface precision-medicine teams build in-house. openai.com | openai.com

Two results reframe staleness as a scheduling and reward problem rather than a retrieval or context-length problem.

Supersede trains the staleness gap down with GRPO against a fact-currency reward Supersede shows the agent staleness bottleneck is memory maintenance rather than comprehension: it survives at frontier scale and does not shrink as memory grows, and GRPO against a fact-currency reward reduces it (preprint, unreplicated). Scaling context or retrieval quality does not fix an agent that confidently serves a superseded fact, which is the exact failure mode for anything reading a changing clinical record. arxiv.org

ChurnBench shows refresh scheduling, not cache age, governs staleness ChurnBench generates a four-source enterprise data fabric as a timeline with an append-only ground-truth ledger, computing gold answers at two timestamps so a once-correct answer scores as wrong later. It isolates refresh policy as the controlling variable, giving a way to test whether your pipeline would still answer correctly after the underlying records change. arxiv.org

Offense, defense and supply chain all produced concrete incidents and shipped controls in the same day.

A backdoored LiteLLM release reached PyPI LiteLLM 1.107.0 on PyPI shipped a credential stealer and backdoor, hitting a package that commonly sits in front of every model call and holds every provider key. Gateway and proxy dependencies are now the highest-value target in an AI stack, so pinned versions and key rotation matter more than model-level controls here. safedep.io

Agent-scale vulnerability discovery now needs a disclosure protocol OpenAI is pushing coordinated global response as its agents find critical vulnerabilities faster than ecosystems can patch, and published a cyber-defender blueprint alongside it. The constraint has shifted from finding bugs to sequencing fixes across every affected operator at once. openai.com | openai.com

Routing and small specialist models are being priced against frontier calls directly.

Mouse-1 sells the small decisions inside an agent loop Mouse-1 drops into an existing agent harness, takes agent state plus available options and returns a probability, claiming 163x cheaper and 65x faster than GPT-6 Sol on those micro-decisions (vendor benchmark). It splits the harness into a cheap policy layer and a frontier reasoning layer, which is the first productized version of a split most agent teams currently hand-roll. producthunt.com

Notion routes per task and claims frontier parity at a tenth the cost Notion's rebuilt agent harness runs tasks past 24 hours with a computer, a browser and a Review for me approval mode, and its Auto router claims parity with Opus 5.5 and Astra at about 90% lower cost (vendor benchmark). The pattern converging across both is per-decision model selection inside long-running harnesses, replacing one-model-per-app. producthunt.com

Eight items dropped: duplicate postings of the same announcement, pure theory with no applied hook, and link-only aggregator posts carrying no claim.