Thursday, October 15, 2026

Pulse (last 3d) · Research 120 · Agents 75 · Products 63 · Industry 51 · Models 36 · Open Source 33 · Releases 20 · Legal 18 · Enterprise 18 · Media 17

Trending · Claude 17 · OpenAI 17 · Claude Code 14 · Anthropic 13 · Google 10 · ChatGPT 9 · FDA 9 · Meta 8 · GitHub 6 · Nvidia 6 · Codex 5 · Cursor 5

---

• The FDA proposes competency-based premarket review for generative-AI devices, and comments close October 19. fda.gov
• Agent memory with write-back can amplify one stale claim until it outvotes a verified correction. arxiv.org
• From 2027, Microsoft will run most inference for its own products on Maia 300. geekwire.com
• ASML expects EUV supply to fall short of demand through 2028, even after adding 30% capacity. digitimes.com
• Gemini 4 Argon is generally available at a limited-time price of $2/$10 per million tokens. venturebeat.com
• MLA retrofits shrink Qwen3.8-27B's KV cache 4x. reddit.com
• Recent agent hacking incidents trace to missing containment, not runaway capability. csoonline.com

• Mistral Large 4 - About 1T-parameter MoE with 49B active, in API preview; open weights promised by month's end. x.com
• OCP grid resilience spec - NVIDIA, Google, Meta and Microsoft require near-full load restoration within two seconds. blogs.nvidia.com
• Ricursive Intelligence - Startup using AI-designed chips to ease compute scarcity. techcrunch.com
• China's open-weight gap - A post estimates Chinese open-weight models trail US closed models by about seven months. reddit.com
• Ling-3.1-flash - inclusionAI released new open weights on Hugging Face. reddit.com
• Sabi-Noema - Qwen3.8-27B fine-tune with an MLA variant, claiming better agentic scores and less verbosity. reddit.com
• Tessera-27B-B-CJK - Apache-2.0 Qwen3.8-27B fine-tune that reasons in compressed CJK shorthand to cut tokens. reddit.com
• Latent Compaction with Test-Time Training - Compresses context into latent representations learned at test time. arxiv.org | huggingface.co
• Cortex - Consolidates context into parametric memory. huggingface.co
• Scaling Discovery Requires Forgetting - Argues that automated discovery needs explicit forgetting mechanisms to scale. arxiv.org
• Agent memory as an architectural workaround - A thread asks whether recurrent-state models like BDH-CQ make memory scaffolding unnecessary. reddit.com
• Post-transformer architecture map - Tables compare BDH, Mamba, RWKV, xLSTM, Titans and hybrids by memory mechanism. reddit.com
• Your Transformer Is Secretly Elastic - Argues that standard Transformers already allow flexible compute at inference. arxiv.org
• ReaDy - Uses recurrent states as draft context for speculative decoding in hybrid linear-attention models. arxiv.org
• EmbeddingGemma 2 image search - Homelab build that embeds thumbnails on a GPU, answers queries on a CPU and stores vectors in SQLite. ai.rundatarun.io
• Slop style traced - Interpretability study pins "slop" style to one direction and finds which training stage installs it. reddit.com
• What RL actually teaches - SAE feature diffing of MiMo-v2.6 traces surfaces eval awareness and mild environment exploitation. reddit.com
• Critique-guided on-policy distillation - Uses critiques to recover reliable supervision from a teacher. arxiv.org
• ActionOPD - Distills agent decisions instead of every token. huggingface.co
• Hindsight-Privileged RLVR - Assigns token-level credit from a single rollout, without group sampling. arxiv.org
• VASPO - Self-distillation with privileged value estimation. arxiv.org
• Cross-interface consistency training - Training on the same task across different interfaces makes agents more robust. arxiv.org
• Mechanical verification for autoformalization - Replaces LLM judges with mechanical checking. arxiv.org
• SWEeper-Bench - Tests whether agents can find bugs by interacting with running software. huggingface.co
• τ-Emotion-Bench - Shows agents breaking down when users express real-world emotions. huggingface.co
• SciUniverse - Benchmarks frontier models on end-to-end chemistry, biology and materials-science work. x.com
• PerSo - Benchmarks LLM social intelligence in real-world persuasion scenarios. arxiv.org
• Co-evolving acting and forecasting - Trains language agents to act and forecast together. arxiv.org
• Price of anarchy in agentic auctions - Analyzes first-price auction efficiency when AI agents are the bidders. arxiv.org
• Google's split coding benchmarks - Google's own charts put Gemini both ahead of and behind Opus at coding. i.redd.it
• When Gemini searches - Across 500 questions, Gemini searched the web 78% of the time when a product was named, 7% otherwise. reddit.com
• Relational deep learning baselines - Re-tests simple baselines for learning on multi-table data. arxiv.org
• BigCell - Generates gigapixel whole-slide pathology images. huggingface.co
• Science Corporation - Implanted its vision-restoring chip in its first Stargardt disease patient. techcrunch.com
• The Physics Is the Network - Claims fast neural emulation of physics with no training. arxiv.org
• Frozen-backbone super-resolution - Runs super-resolution in the latent space of a frozen vision foundation model. huggingface.co
• Acyclic - Splits agent tasks across instant workspace forks and merges the results. producthunt.com
• SlashDocs - Razorpay's open-source hub for agent-written documents now produces about half of Razorpay's internal documents. producthunt.com
• Agent-Native - TypeScript framework where one action definition drives the UI, agent, API, CLI, MCP and A2A endpoints. producthunt.com
• Jarvis Live - MIT-licensed realtime voice agent that hands computer tasks to local coding agents. producthunt.com
• Figma agent features - New features set out how agents work on Figma's platform. techcrunch.com
• Third Moment Research - Runs 10,000 agents for more than 36 hours on OpenAI (vendor case study). openai.com
• Are Harnesses Worth the Cost? - Asks whether agent harnesses justify their overhead. martinfowler.com
• SkillMarker - KGW-style watermarking to prove ownership of agent skills. huggingface.co
• alphaXiv agent telemetry - Its OpenResearch skills allegedly send telemetry to the vendor and tell agents not to disclose it (single source). reddit.com
• OpenAI safety fixes - OpenAI is reportedly rushing to fix safety issues before its next flagship ships (single source). x.com
• The Safety Culture at the Labs is Broken - Zvi Mowshowitz argues that safety culture at the frontier labs has broken down. thezvi.substack.com
• Anthropic on open-weight risk - An Anthropic executive says open-source models carry higher risk. reddit.com
• Anthropic midterms preparation - Anthropic sets out its preparations for the 2026 US midterm elections. anthropic.com
• Anthropic Singapore - Anthropic opened a Singapore office and signed an MOU with the government. anthropic.com
• Anthropic's 2021 roadmap - The research roadmap became public through copyright litigation. x.com
• SentinelOne - Malware analysis cut from weeks to hours with OpenAI models (vendor case study). openai.com
• Block security - Block builds AI-powered security tooling on OpenAI technology (vendor case study). openai.com
• Apate - Agents pose as scam victims to gather fraud intelligence. techcrunch.com
• Flock microphones - The CEO says they detect "just gunshots"; Flock's website says otherwise. techcrunch.com
• YouTube Music AI labels - New labels arrive as AI-generated music floods streaming platforms. techcrunch.com
• Kugel 3.5 - TTS model reads long numbers correctly 97.1% of the time at 5.25 cents a minute (vendor benchmark). producthunt.com
• AURORA Lifeline - Turns IMD cyclone bulletins into plans for health-facility road access, using 1,000+ storm ensembles. producthunt.com
• Runway Solaris - Runway's first Interface World Model. runway.com | youtube.com
• GWM Worlds 2 - Runway's playable world model. youtube.com
• Robot benchmark climb - Frontier LLMs went from 15% to 48% on RoboDojo-RC Tier 1 between July and September (single source). x.com
• Player-1 - Trains game-playing agents on synthetic experience. arxiv.org
• PrismaCoder - Uses decomposed rubric synthesis to train open multimodal code models. arxiv.org
• PhySGA - Physics-grounded agentic framework for interactive 3D scene composition. arxiv.org
• SafeDriveVLA - Uses a navigation-conditioned world model to anticipate conflicts in end-to-end driving. huggingface.co
• Diffusion over humanoid policy weights - Uses diffusion to model the weights of humanoid control policies. arxiv.org
• MimicX - Refines supervision with the policy in the loop for video-driven humanoid motion tracking. huggingface.co
• Hierarchical residual policy learning - Real-world mobile manipulation learned from sparse human guidance. arxiv.org
• Mr.Chatterbox-2.7B-Chat - Chat model trained from scratch on Victorian texts. reddit.com
• Kokon - Open-source desktop app that runs Qwen-Image 2.1 editing locally on 12GB+ GPUs. reddit.com

Three preprints and a 1.0 release treat agent memory as a staleness problem, which is how it fails when guidelines and variant classifications change.

Counters cannot retire stale memories, and write-back makes stale claims reinforce themselves arXiv 2610.17544 proves that the success counters in ExpeL- and ACE-style memories are CUSUM detectors testing a staleness threshold nobody chose. arXiv 2610.17763 shows write-back turning one obsolete claim into same-source copies that outvote a verified correction for 30 queries (preprints, unreplicated). A memory that holds reclassified variants needs source deduplication and explicit retirement, not counters. arxiv.org | arxiv.org

A decision ledger ties every action to how fresh its evidence was arXiv 2610.18126 records six timestamps per LLM decision, limits what the evidence can guarantee at the moment of action, and writes watch conditions that route later publications back to decisions already made (preprint, unreplicated). The design fits audit trails for treatment recommendations, where the question is whether the evidence was still current when the action happened. arxiv.org

Hindsight 1.0 moves agent memory onto object storage Vectorize's open-source Hindsight reached 1.0 with a stable API and claims first place on every Agent Memory Benchmark dataset (vendor benchmark). Its Rust engine, Memlake, serves 100M memories from object storage at under 100 ms p99 with no database or consensus cluster. Self-hosted memory inside a PHI boundary then needs little more than a storage bucket. producthunt.com

Regulators, tool builders and security researchers are converging on one rule: an agent cannot grade its own work.

FDA proposes competency assessment for generative-AI devices, and no institution has commented yet A discussion paper from CDRH's Digital Health Center of Excellence (docket FDA-2026-N-7874) proposes a competency-assessment premarket model and adds a section on agentic AI. The docket holds 129 comments, none from institutions, and closes October 19, so a precision-medicine sponsor that files this week shapes the default framework with almost no competing input. fda.gov

Coding-agent tools now check intent outside the session that wrote the code Tiden keeps requirements as branchable records and judges each change by tests run in the user's environment, because tests written in the authoring session agree with the code by construction. Diffo keeps the authoring session live, so a comment on a line of the diff gets answered and fixed by the agent that wrote it. producthunt.com | producthunt.com

Agent security failures are containment failures Security experts attribute recent AI hacking incidents, including the Hugging Face compromise, to labs skipping sandboxing and monitoring, not to agent capability. About 400 scientists, including Hinton and Bengio, are separately calling for an international body to set containment standards for cyber-capable AI, and any such standard would reach teams that run agents against production data. csoonline.com | qz.com

Hyperscalers are building their own inference chips and cutting API prices while the sole EUV supplier says it cannot meet demand for three more years.

Microsoft will run most of its own inference on Maia 300 Maia 300, announced October 12, pairs TSMC N2P compute chiplets with 8 HBM4 stacks in 96-accelerator racks and delivers 40% more tokens per megawatt (vendor claim). Production starts in 2027, when Maia becomes the main inference system for Microsoft's products, so more Azure capacity for long-context and agent workloads will run off Nvidia. geekwire.com | producthunt.com

ASML cannot meet EUV demand before 2028, even with 30% more capacity ASML posted record Q3 sales of €11.7B and raised 2026 guidance to €44-45B, its third raise this year. Customers are already asking for 2029 and 2030 order slots, and the shortfall persists after a planned 30% increase in EUV capacity. That puts a hard ceiling on how fast dedicated inference capacity can get cheaper. startupfortune.com | digitimes.com | cnbc.com

Gemini 4 Argon launches at $2/$10 per million tokens Google made Gemini 4 Argon, its flagship multimodal model, generally available on October 13 at a limited-time price of $2 input and $10 output per million tokens. With supply this tight, discounted frontier pricing looks like a bid for market share, so cost models built on this price should assume the discount ends. venturebeat.com

Today's open releases compete on KV cache size and resident memory more than on raw scores, which brings long-context and on-device workloads within reach of modest hardware.

KV caches are shrinking 4x in both serving and RL Telperion retrofitted Qwen3.8-27B with multi-head latent attention and grouped-latent variants that cut the KV cache 4x and reason more efficiently (single source). KV4-RL uses INT4 KV-cache quantization to speed up asynchronous RLVR rollouts, so the same compression now lowers both serving and post-training costs for long clinical documents. reddit.com | huggingface.co

Mid-size models now fit in phone-class memory A demo runs MiniCPM-V 4.7, a 35B multimodal model, on a phone in about 3GB of memory (single source). Liquid AI's LFM2.5-24B-A2B activates about 2B parameters per token and runs 1.7x faster than comparable MoEs (vendor benchmark), which makes on-premise and edge inference next to patient data a realistic deployment target. reddit.com | liquid.ai

9 items: no discernible claim, a duplicate release announcement, user sentiment, or an engagement video.