Thursday, November 5, 2026

Pulse (last 3d) · Research 98 · Agents 92 · Products 75 · Models 51 · Industry 36 · Open Source 26 · Enterprise 23 · Legal 20 · Infra 20 · Inference 13

Trending · OpenAI 17 · Anthropic 15 · Claude 15 · ChatGPT 13 · MCP 11 · Codex 10 · Google 9 · Claude Code 8 · Apple 5 · Cursor 5 · Legora 5 · Astra 4

---

• Six biology reasoning systems pass their benchmarks with the biological input destroyed. arxiv.org
• Revolut made self-hosted models the default and cut coding inference cost about 90%. theregister.com
• Cryptographically provable inference now works, at 90-100x compute overhead. anthropic.com
• Best frontier model scores under 0.32 on drug-perturbation gene expression prediction; a 27B specialist hits 0.366. x.com
• Moonshot shipped Kimi K3.1 Pro weights at roughly 1.45TB. huggingface.co
• Opus 5 did mechinterp on ProteinMPNN and edited ~1% of weights to reproduce SolubleMPNN behavior. x.com
• Apple's 10-K now carries a rogue-AI-agent risk disclosure. reddit.com

• OpenAI Images 2.6 - new image model; ChatGPT now outputs animated GIFs. openai.com | theverge.com
• FLUX 3 Image - bounding-box layout control, 10 reference images, native 4K. producthunt.com
• OpenAI acquires Pierre Computer - developer-tools startup working on git and diffs. openai.com
• SWE-2 Flash - 276B total, 12B active; 3x faster at a fifth the price. reddit.com
• Ripple-65B preview - dense open-weight model pitched as US agentic-coding SOTA (vendor benchmark). ricursive.com
• TwIL-8B open-sourced - 8.8B claiming 93% of frontier coding performance (vendor benchmark). x.com
• MemForm - consolidation order alone poisons agent memory across six LLMs (preprint, unreplicated). arxiv.org
• wpgit - git-style version control for gradient-free weight edits by LLM agents. reddit.com
• LoopVQ - sub-1-bit KV cache quantization for looped transformers. huggingface.co
• Systolic Linear Attention - tile-grouped gated delta attention co-designed for TPUs. arxiv.org
• SpecCompact - slot-parallel speculative verification speeds up agent context compaction. arxiv.org
• Multi-turn RL audit - asks what multi-turn RL actually improves in agents. huggingface.co
• On-policy distillation limits - generalizes poorly when consolidating RL-trained models. huggingface.co
• Turn-level best-of-N - test-time scaling for agents at the turn granularity. arxiv.org
• ServerBench and CompressionBench - can coding agents design serving systems and compress models. arxiv.org | arxiv.org
• FrontierPhysics - end-to-end physics research benchmark for agents. huggingface.co
• BEAKER - expert-curated benchmark for embodied agents in self-driving chemistry labs. arxiv.org
• Context-answer metamodels - forecast activations before the first answer token. arxiv.org
• Persona perspectives - adopting a viewpoint reorganizes internal representations. arxiv.org
• Fine-tuned 2B classifier - frontier quality at 35ms and ~1% of cost in production. x.com
• Claude Code subagents - background status updates now ~99% cheaper, written by Haiku 5.5. x.com
• Render Sandboxes - per-agent Linux VMs, sub-second start, suspend on idle. producthunt.com
• Hindsight inference API - drop-in rerank/embed plus a 167ms typed-decision model. producthunt.com
• fileForensics - API returns tampering and AI-generation evidence for documents agents act on. producthunt.com
• QualityLayer - seven-stage gated pipeline where a fresh agent reviews the author agent's code. producthunt.com
• Zero Wire - Cloudflare vulnerability feed with agent-runnable read-only exposure checks. producthunt.com
• CrowdStrike AI kill switch - enterprise control to halt agent activity. crowdstrike.com
• OpenAI on privilege paths - internal work removing thousands of routes to root. openai.com
• AI vulnerability discovery - essay on paying down 40 years of software technical debt. spectrum.ieee.org
• Siri opens to developers - third-party access to Apple's assistant. developer.apple.com
• Qualcomm pivots to data center - inference silicon as phone sales decline. theverge.com
• Token cost analysis - what inference actually costs beneath list pricing. substack.hadley.nz
• Agentic commerce flop - DoorDash reports only hundreds of outside-agent orders. arstechnica.com
• Data centers lost the midterms - local opposition translated into election results. theverge.com
• OpenAI usage figures - ChatGPT search ~20% of Google's volume, 250M voice users. bigtechnology.com
• Left-of-center AI willingness collapsed - FT polling since Q1 2025 (single source). reddit.com
• DeepSeek super-agent design - proposal for agent systems that manage themselves. x.com
• Looped transformer HBM profiling - weight reuse measured on a B200. redd.it
• Visual token routing - learn to keep only the visual tokens that matter. arxiv.org
• MUON-RT - drops Muon's momentum buffer by reusing the teacher's trajectory. huggingface.co
• VA-TDM - unified view of few-step diffusion distillation via velocity adversarial training. huggingface.co
• Embedding quantization trade-off - hierarchical structure versus compression in text embeddings. arxiv.org
• ARC-AGI-2 calibration data - rules, traces and routing statistics released for open models. arxiv.org
• Time-sensitive safety - benchmark for assistants recognizing urgency. arxiv.org
• Multilingual safety templates - scaling alignment data generation across languages. huggingface.co
• Alignment myopia - argues current alignment work optimizes too short a horizon. arxiv.org
• Robot policy RL scaling - RL scales generalist policy improvement. arxiv.org
• Native VLA policies - pixels straight to actions without intermediate stages. huggingface.co
• PyRUA-DataFarm - LLM agent replaces human teleoperators for robot data collection. huggingface.co
• VideoMemBench - tests multimodal recall across video conversations. huggingface.co
• SleuthVLM - questions whether audio-in-video is needed for omni understanding. huggingface.co
• Turn-level credit for video RL - streaming video RL improved by per-turn attribution. huggingface.co
• SurrealDB and KiviDB - agent-oriented databases: scale-to-zero multi-protocol, and Rust Redis-compatible. producthunt.com | producthunt.com
• Fikra A4H - agents provision accounts and scoped credentials with auditable receipts. producthunt.com
• 13B for under $4K - prerelease claim of a cheap frontier-adjacent training run (single source). producthunt.com
• Mika - on-device Android agent that drives apps and the web, 6.8k stars. github.com
• SaaS growth inflection - essay dating the seat-based software decline to October. mackgrenfell.com
• Occupational exposure limits - working conditions bound what AI exposure scores predict. arxiv.org
• Explanation faithfulness metrics disagree - criteria conflict, so faithfulness claims depend on the metric chosen. arxiv.org
• GFlowQEC - generative flow networks sample quantum LDPC codes. arxiv.org
• Belief propagation network - neural generalization of BP for decoding linear codes. arxiv.org
• LongBanana - expert-verified benchmark for multi-reference image synthesis. arxiv.org

The systems that claim to reason over sequence data are mostly ignoring it, and specialist models are beating frontier models where the biology actually binds.

Shuffling the DNA input changes a disease-prediction score by 0.001 Four reliance tests across six LLM-plus-foundation-model biology systems found BioReason with Evo 2 scoring 0.848 with shuffled DNA against 0.847 intact, and BioReason-Pro with ESM3 following text over ESM embeddings in 99.7% of conflict cases (preprint, unreplicated). The benchmarks these systems report on are satisfiable from gene names and textual priors alone, which means a genomics pipeline built on one of them may be doing retrieval dressed as inference. arxiv.org

Frontier models score under 0.32 on drug-perturbation prediction; a 27B specialist hits 0.366 CellType released BioSimBench and BioSimGym, a benchmark for predicting gene-expression change under drug perturbation, and its own 27B model trained in the gym outscores frontier models on every task (vendor benchmark). The absolute numbers matter more than the ranking: at 0.366 nobody is close to usable, so perturbation prediction is still an open problem rather than a procurement decision. x.com | x.com | x.com

A long-horizon biology benchmark finds models cannot judge which inputs are significant BioResearchBench puts 50 long-horizon biological research tasks to frontier models and reports persistent failures in complex reasoning and in assessing the significance of biological inputs (single source). That second failure mode is the same defect the reliance tests measured directly, which places the weakness in the models rather than in any one benchmark's construction. x.com

Three separate lines of work treat model parameters as something to inspect, patch and prove rather than consume as a service.

An agent found and edited roughly 1% of ProteinMPNN's weights to change its behavior Opus 5 performed mechanistic interpretability on ProteinMPNN unprompted by special tooling and modified about 1% of the weights to produce SolubleMPNN-like behavior (single source). Targeted weight surgery on a published scientific model, done by an agent in a session, turns fine-tuning budgets into interpretability budgets for small domain models. x.com

Provable inference works at 90-100x overhead Cryptographic verification that a served response came from the weights you paid for now runs end to end, at roughly 90-100x compute cost, reported alongside a FlashProve scheme for proving model identity without exposing weights. The overhead rules it out for bulk traffic, but it makes spot-auditing a regulated or clinical inference contract technically possible rather than contractual only. anthropic.com | interestingengineering.com

Moonshot shipped a 1.45TB open-weight model Kimi K3.1 Pro's weights are public at roughly 1.45TB, which is open in license and closed in practice for anyone without a multi-node serving cluster. Set against the same day's 8.8B and 2B results, the open-weight tier is splitting into models you can actually host and models you can only inspect. huggingface.co

Cost reduction stopped being a research claim and showed up in a production migration and in per-task model sizing.

Revolut cut coding inference cost about 90% by defaulting to self-hosted models Revolut made self-hosted models the default for AI coding work in place of Claude and reports roughly a 10x cost reduction. This is the first large regulated-industry engineering org to publish the switch rather than pilot it, and it sets the number any internal build-versus-buy case will now be measured against. theregister.com

Apple's rogue-agent warning moved from blog to 10-K Apple's annual filing warns that AI agents may act without authorization, including concealing or manipulating their actions, and Apple changed macOS full disk access alongside it. A risk factor in a securities filing is a legally reviewed statement, which is why the same week brought a CrowdStrike kill switch and Cloudflare shipping agent-runnable exposure checks. reddit.com

Eleven items: sentiment posts, duplicate URLs for stories already covered, consumer gadget reviews, and link-roundup newsletters with no claim of their own.