Pulse (last 3d) · Research 107 · Agents 67 · Products 50 · Models 47 · Industry 42 · Open Source 29 · Enterprise 24 · Releases 20 · Legal 20 · Inference 14

Trending · Anthropic 20 · OpenAI 16 · Claude 11 · Nvidia 10 · Claude Code 8 · ChatGPT 6 · Fable 5.2 6 · Jensen Huang 6 · Codex 5 · Google 5 · DeepSeek V4.1 Flash 4 · FDA 4


Near-frontier capability got much cheaper today: one point off the frontier at Flash prices, 594B open weights under Apache 2.0, and a 145B model in 29GB. The hard parts of clinical AI are now evaluation, audit trails and wet-lab validation, and AlphaGenome 2 shows the same pattern in biology. Expect fewer decisions about which model to use and more about which outputs to trust, and note that Anthropic and Nvidia both now answer to outside scrutiny. deepmind.google | edge0.ai

  • AlphaGenome 2 predicts variant effects per cell type and was trained on 107M single-cell profiles. deepmind.google
  • Gemini 4 Hydrogen scores one point below Sonnet 5.5 on Artificial Analysis at far lower cost. x.com
  • Inkling 2 Small ships as 594B-total, 17B-active open weights under Apache 2.0 with 1M context. x.com
  • Gemini Decision 1 scores up to 1,000 classification questions per call at $0.10/M input. reddit.com
  • Anthropic's IPO filing says it has met its RSP threshold for AI-accelerated research. reddit.com
  • A Claude Code audit log shows approvals that the user says they never gave. reddit.com
  • Jensen Huang has resigned from Nvidia amid affair and China-spy allegations. bbc.com
  • Grok Bot 1.0 - Grok bots join team workspaces and take on work unprompted. producthunt.com
  • Foundation Models framework - Apple gives eligible small developers free model access on Private Cloud Compute. developer.apple.com
  • smolmaxxing - runs 16B-parameter models on devices with under 1GB of usable memory. github.com
  • Gemini 4 Carbon, Antigravity /swarm, Gemini Robotics 2.5 - Carbon pricing plus Google's agent-swarm and robotics releases. x.com
  • Automated red teaming at Anthropic - automated attacks find strong jailbreaks before safeguards ship. x.com
  • Export-controlled GPU fingerprints - claims checkpoints reveal Chinese labs used restricted Nvidia GPUs; commenters dispute the evidence. unemploymentlabs.com
  • Claude Meetings - Claude joins Google Meet calls and takes notes for Team and Enterprise customers. reddit.com
  • United Airlines bans AI agents from booking flights - the first major airline to block agent bookings for travelers. cbsnews.com | finance.yahoo.com
  • Brave Search API - now accepts payments from agents over the x402 protocol. brave.com
  • AI chatbots recommend fake retail websites - a shopping test caught assistants citing risky fake stores. theverge.com
  • The AI built an accurate map, then answered without it - the model ignores an internal spatial map that would score 80% versus 73% (single source). reddit.com
  • Benchmark Router - hosted public-benchmark runs on any model or harness at a claimed tenth of the cost. x.com
  • ElevenAgents Campaign Manager - outbound voice campaigns with call policies and outcome tracking. producthunt.com
  • Mitra MCP - agents place calls from your personal number and hand off to you when needed. producthunt.com
  • Mode - an open-source iPhone agent that drives your apps on your own model key. producthunt.com
  • Sunday - an MIT-licensed Mac agent for browser, files, terminal and email, with markdown memory. producthunt.com
  • Adobe AI Assistant - conversational multi-step tasks across Photoshop, Premiere, Illustrator and Firefly. producthunt.com
  • CanyonOS - an AGPL platform that traces calls between agents to autoscale bottlenecks and attribute cost. producthunt.com
  • WTFM - a Claude Code and Codex plugin that writes user docs with screenshots from source. producthunt.com
  • Fizzlayer - an Apache 2.0 layer that feeds physical sensor readings to agents over MCP. producthunt.com
  • AgentRun - an open-source managed-agent platform that claims to improve itself. github.com
  • Amplitude with OpenAI - Amplitude's Wave product agent, built on OpenAI models and Codex, ships fixes in days. openai.com
  • ArtCraft - cloned Photoshop and now aims to rebuild the wider creative stack. latent.space
  • Apple app review slowdown - AI-generated submissions are clogging App Store review for synth developers. synthanatomy.com
  • Game developers adopt generative AI - a GDC survey shows adoption rising alongside layoffs. theverge.com
  • No AI-discovered Maya site - the claim that AI agents found an unknown Maya site is false. theverge.com
  • AI and children's safety - OpenAI backs a research declaration on children's safety. openai.com
  • China data center support - a UN survey finds 80% support construction and two-thirds use AI daily. cbsnews.com
  • Do robots still need their own models? - argues VLAs, tactile encoders and world models still beat frontier multimodal models on real robots. reddit.com
  • Desk rejections at NeurIPS/ICML/ICLR - since 2013, the main reasons moved from formatting to reviewer and author conduct. reddit.com
  • Legora-style conflict detector - a tutorial that builds a legal conflict check from NLI and encoder models only. i.redd.it
  • Look Here, Then Decide - small decision models in a harness find memory-safety bugs on commodity GPUs. arxiv.org
  • Singularity-Bench - tests whether agents can build other agents. arxiv.org
  • Reducing Batch Makespan - coordinates the inference engine with execution environments so agent batches finish sooner. arxiv.org
  • Folio - recoverable multi-resolution context folding for long-horizon agents. arxiv.org
  • Actor, Judge, or Teacher? - compares the roles a model can play in long-horizon agents. huggingface.co
  • Evidence-Guided Recovery - agent recovery when tool outcomes are ambiguous or only partly executed. huggingface.co
  • Agent-R^2 - scales productivity agents with training data synthesized from real sources. huggingface.co
  • MobileProbe - trajectory-aware injection attacks for testing mobile GUI agents. arxiv.org
  • AstroMemBench - benchmarks agent memory on spacecraft operational history. arxiv.org
  • When Expert Disagreement Hurts - prestige and selectivity cues sway how LLMs revise decisions. arxiv.org
  • LLMs as Digital Twins - a distribution-aware questionnaire test of LLMs simulating people. arxiv.org
  • System One models - adaptive computation turns LLMs into fast models, and SO-Arena benchmarks their reliability. huggingface.co | huggingface.co
  • Factored One-sided Adam-Moment - cuts the memory cost of the SOAP optimizer so it scales. huggingface.co
  • Safety as a Budget - calibrates the control channel of constrained GRPO for VLM safety training. arxiv.org
  • Humanized but Still True? - audits whether AI-text humanizers keep facts intact and evade detection. huggingface.co
  • Manifold Attracting Flow Matching - keeps generative sampling on physically valid manifolds. arxiv.org
  • Natural Gradient in Predictive Coding - lateral circuits can implement natural-gradient learning. arxiv.org
  • Dataset Distillation via Drifting - a drift-based method for distilling datasets. arxiv.org
  • SPARC - representative vectors make photonic matrix multiplication more accurate. arxiv.org
  • Decomposed JEPAs - asymmetric multimodal learning with split joint-embedding predictors. arxiv.org
  • UniSpark - hybrid attention for unified multimodal understanding and generation. huggingface.co
  • G-STAR - group-scaled token capacity for deeply compressed autoregressive generation. huggingface.co
  • MuSeToM - tests LLM belief revision using textual cues across six senses. arxiv.org
  • JRDB-Social transfer - in VLMs, person-action understanding helps with group intent only in one direction. arxiv.org
  • Over-Realization in Text-to-Image - models draw elements the prompt mentions but says should be absent. arxiv.org
  • Continuity-Aware Action Priors - smoother action priors improve planning with world models. arxiv.org
  • WorldDraft - video models render worlds that code specifies. huggingface.co
  • PhysLoop - LLM physics checks drive a self-improvement loop for video generation. huggingface.co
  • Duo - a reciprocal world model for full-duplex interaction. huggingface.co
  • BirdDreams - generates perspective and bird's-eye-view video together for driving world models. huggingface.co
  • Ropedia Xperience 2 - a reference resource for scaling physical-AI environments and behaviors. huggingface.co

In each case below, computational prediction gets sharper while the experiment that confirms it stays the bottleneck.

AlphaGenome 2 predicts variant effects per cell type DeepMind's AlphaGenome 2 makes its variant-effect predictions conditional on cell type and was trained on 107 million single-cell profiles (press release). The new piece is cellular context: a noncoding variant can be scored in the specific cells where a phenotype arises, which is the question tissue-specific VUS interpretation depends on. deepmind.google | deepmind.google

The labs' own data says AI sped up the desk, not the bench The AI labs' science programs show AI speeding up literature, analysis and hypothesis work while wet-lab throughput stayed flat. The labs are now moving into the lab itself, so assay data and lab automation become the contested resource in AI for science. rundatarun.io

TRIBE v2 recovers one known brain map out of three An independent test of Meta's TRIBE v2 fMRI-prediction model finds motion selectivity in MT+ but no retinotopic or tonotopic maps (single source). Strong prediction scores did not mean the model learned known biology, the same gap that separates a good variant-effect benchmark from a validated mechanism. reddit.com

Google's pricing and three open-weight releases close most of the gap between hosted frontier models and what a team can run on its own hardware.

Gemini 4 Hydrogen lands one point off the frontier at Flash prices Artificial Analysis scores Gemini 4 Hydrogen at 54 on its Intelligence Index, one point behind Claude Sonnet 5.5 and GPT-6.1 Sol, at much lower cost than Sonnet. It replaces 3.8 Flash in Search and the Gemini app, and it ships with an Orchestrator that makes cheap-tier routing a first-party feature. x.com | blog.google | antigravity.google | theverge.com

Open weights reach 594B parameters, and a 145B model fits in 29GB Thinking Machines released Inkling 2 Small under Apache 2.0, a 594B-total, 17B-active multimodal MoE with 1M context. Edge0-145B shares matrices across expert groups to run in about 29GB (vendor benchmark), and Ernie 5.5 Flash Preview performs strongly at 14B active (single source), so keeping PHI on-premises no longer costs much capability. x.com | edge0.ai | i.redd.it

Gemini Decision 1 makes classification a scoring call Google's Decision 1 API returns calibrated probabilities for up to 1,000 yes/no or choice questions per call, with no output tokens, at $0.10 per million input tokens (single source). Phenotyping or trial-eligibility screening over clinical notes becomes one scored call instead of a generation pipeline, as long as the calibration holds on clinical data. reddit.com

The new agent tools assume the agent will make mistakes and build review, simulation and label-free evaluation around it.

Claude Code's log shows approvals a user says never happened The new detail in the Claude Code Windows report is that the log records approvals for nine PowerShell commands the user says were never given; in manual mode, those commands moved about 11GB to the recycle bin (single source). If confirmed, the approval log cannot serve as evidence of human sign-off in a regulated workflow. reddit.com

Shared agent memory gets branches and merge review Omnigraph is an open-source graph database on your own S3-compatible storage that puts every agent write on a Git-like branch for review before merge. Harness Memory shares pluggable long-term memory across Claude Code, Codex and Hermes with fine-grained access control, so agent memory gets governed like any other data. producthunt.com | producthunt.com

Pre-production simulation is becoming its own product category Plurai builds sandboxed digital twins of an agent's tools, documents and users, and Breken runs apps against simulated users, services and time, with reproduction evidence for each bug. three.dev mines live traffic for silent failures ranked by severity, built on the same premise that agent errors only show up under realistic conditions. producthunt.com | producthunt.com | producthunt.com

AGORA weighs LLM judges without human labels AGORA uses item response theory to learn each LLM judge's reliability across all items at once, and it matches or beats Dawid-Skene when human validation sets shrink to a few hundred items (preprint, unreplicated). That is the normal situation in clinical evaluation, where expert labels are scarce and expensive. arxiv.org

Anthropic's IPO filing and Nvidia's CEO exit both move decisions about frontier models and chips into public, disclosed processes.

Anthropic's IPO filing says AI now roughly doubles its research pace Anthropic's amended IPO filing, for an offering reportedly raising about $100B, says it has met its RSP threshold for AI-accelerated AI research. Under the RSP, that threshold marks where AI roughly doubles the pace of AI progress, and a safety trigger is now a statement in a securities filing that release cadence and safeguards will answer to. ipo.anthropic.com | reddit.com

Huang leaves Nvidia with the China question still open Jensen Huang resigned as Nvidia CEO after affair allegations tied to an alleged Chinese spy, and reporting on chip smuggling continues. Internal emails call it an "extraordinarily difficult period," Nvidia's roadmap and export posture now sit with new leadership, and Hugging Face's CEO has pledged to keep fighting for open-source AI. bbc.com | variety.com | theverge.com | theverge.com | theverge.com | scientificamerican.com | techcrunch.com

8 items: opinion pieces, retrospectives, a model-identity question and an internal sandbox log, none with a usable claim.