Pulse (last 3d) · Agents 90 · Research 86 · Products 67 · Models 39 · Industry 33 · Enterprise 31 · Open Source 22 · Releases 22 · Legal 17 · Media 14
Trending · Anthropic 18 · OpenAI 18 · Claude 15 · ChatGPT 13 · Codex 10 · Google 10 · Nvidia 10 · Claude Code 9 · Hugging Face 9 · Cursor 7 · Gemini 7 · MCP 7
Top story
Markets are now pricing frontier models as research labor: Anthropic's record $81B IPO landed the same week Anthropic and OpenAI each published evidence of their models doing original science and AI R&D. Below the frontier, prices are falling fast, and two papers suggest post-training mostly rewires how a model reads out what it already knows rather than adding knowledge, which makes small, auditable edits to open weights a credible way to adapt clinical models. The decisions this changes are routing and adaptation: send routine coding and pipeline work to cheaper near-frontier tiers, and design agent audit trails in from the start, because the forensics research shows they cannot be added later. cnbc.com | arxiv.org
The Shortlist
- Anthropic's IPO raised over $81B, the largest on record, and the stock fell after the first trade. cnbc.com
- Microsoft says MAI Code 2 delivers 98% of GPT-6.1 Sol at 40% of the cost, now in Copilot. producthunt.com
- Anthropic says Claude solved two open problems in black hole physics and published two papers on them. anthropic.com
- Post-training changes how a model reads out its internal representations, not the representations themselves. arxiv.org
- astro-weights cut Qwen3.8-27B output tokens by 30% with no detected loss of accuracy. producthunt.com
- A pathology foundation model with MIL grades Barrett's dysplasia across institutions without stain normalization. arxiv.org
Also Noted
- Multi-institutional Barrett's esophagus grading - pathology foundation model plus MIL, no stain normalization needed. arxiv.org
- NVIDIA interim CEO Kress all-hands - first staff address since Jensen Huang's exit, message "let's move forward." cnbc.com
- OpenAI Performance Factory - OpenAI's internal system for keeping ChatGPT latency down. openai.com
- TideCache - hybrid-LLM serving with transparent cache management and layer-wise streaming. arxiv.org
- Embeddings vs. Grep - questions whether dense embeddings actually beat lexical search for code retrieval. huggingface.co
- Realized versus Population Reliability - anytime-valid self-consistency certification that accounts for correlated reasoning samples. arxiv.org
- Active Feature Acquisition by Distilling an Oracle Teacher - learns which features are worth acquiring by distilling an oracle teacher. arxiv.org
- Simple Private Noisy SGD - proves plain DP noisy SGD is minimax-optimal under population strong convexity. arxiv.org
- AutoMaster - autonomous post-training that makes the model a better teacher for itself. arxiv.org | huggingface.co
- LLM GerryMandering - statistical method separating interaction effects from additive effects in LLM audits. arxiv.org
- SoK: LLMs in vulnerability research - systematizes how LLM capability in vulnerability research has evolved. arxiv.org
- ARTEX - English translation of an autonomous AI penetration-testing system, 12,451 stars. github.com
- Shane Legg on AGI to ASI - DeepMind essay on the path from AGI to superintelligence. x.com
- Olala mini 7A1B - OVHcloud's 7B open model with a new architecture, trained on European supercomputers. x.com
- Postgres 19 - released and already available on Neon and Databricks Lakebase. databricks.com
- Netlify Server - long-lived backends for agents with open WebSockets and in-memory state; scales to zero. producthunt.com
- Oodle AI - agent observability platform claiming 80% lower cost and 900M+ agent spans a month. producthunt.com
- Tabstack by Mozilla - managed API for web research, extraction and browser tasks; customer data never used for training. producthunt.com
- Polylane for Azure - investigates Azure Monitor alerts and drafts code fixes, with no stored credentials. producthunt.com
- Selects - pro video editor built as an agent harness, with frame-timed transcripts; files stay local. producthunt.com
- HeyGen Avatar 6 - avatar model rated 84.4% indistinguishable from real footage (vendor benchmark). producthunt.com
- Gemini on Mac uses your computer - Google's macOS app now operates the desktop as a computer-use agent. theverge.com
- Apple hiring for its own AI agent - job listings show Apple building an in-house agent. theverge.com
- Apple's always-listening recap - two-week hands-on with Apple's AI that summarizes ambient audio. theverge.com
- ChatGPT Live Dictation - real-time transcription while you speak. theverge.com
- Google homepage AI chats - AI chat now sits directly on the Google homepage. theverge.com
- Google AI Mode temporary threads - an incognito-style mode for ephemeral AI Mode conversations. techcrunch.com
- Meta Muse Charm - reportedly priced around $300. theverge.com
- Ideogram 4.6 - image model with native 4K output and more precise editing. x.com
- GEOHunter - maps a generative engine optimization network operating in the wild. arxiv.org
- My Chrome Extensions Are Just Prompts - builds browser extensions from prompts instead of code. aifoc.us
- Weight-Space Programming - builds networks by designing weights directly, 1990s style. notesbylex.com
- Verify the Verifier - spec-grounded LLM cross-checking catches hardware bugs that escape UVM scoreboards. arxiv.org
- Cost-Aware Agentic Feature Optimization - agentic feature search for fleet-scale ranking models under cost limits. arxiv.org
- The Window Lever - longer observation windows beat data, capacity and architecture in behavior forecasting. arxiv.org
- Suppressing the Shortcut in VideoLLM Fine-Tuning - part of the reported fine-tuning gains come from shortcuts. huggingface.co
- WeMM-Video - open-source omni-modal video understanding model. huggingface.co
- Depth Pro 2 - sharp, temporally consistent metric monocular depth for video. arxiv.org | huggingface.co
- WM^MLLM - adapts multimodal LLMs to work as world models. huggingface.co
- Action Effect Matching - trains world models on what actions actually change. huggingface.co
- One Prior, Two Dynamics - repurposes video diffusion transformers for self-supervised latent action learning. arxiv.org
- AgentRoboCritic - benchmarks frontier agents on direct robot control. arxiv.org
- PGX - physically grounded execution for long-horizon humanoid loco-manipulation. arxiv.org
- Mind2Act World - diagnostic benchmark that separates embodied reasoning from physical action. huggingface.co
- ZeroVLN - vision-language navigation without supervised fine-tuning of the VLM. huggingface.co
- TuringSpace - tests spatial intelligence in action using interactive CAPTCHAs. huggingface.co
- DailySafe - benchmarks dynamic safety awareness in multi-turn multimodal interactions. huggingface.co
- Probing the "Conscious" Mind - tests VLMs for selective access and global-workspace-like dynamics. arxiv.org
- S1CogBench - tests fast "System One" decision models on cognitive experiments. huggingface.co
- MM-SystemOneBias - measures bias in multimodal System One decision models. arxiv.org
- PlayCraft - designs games from agents' play experience. huggingface.co | arxiv.org
- ProtoDiffusion - token-per-layer diffusion that generates multi-layer images. arxiv.org
- AURA - audio-driven streaming avatar generation. arxiv.org
- The Mask Is Not the Object - volumetric supervision instead of masks for 3D Gaussian segmentation. huggingface.co
- Bio-PM-Guided Synthetic IMU Generation - synthetic IMU data guided by biomechanical priors. arxiv.org
- Deterministic Temporal Embedding - deterministic trace extraction for calcium imaging. arxiv.org
Research output is now what frontier labs are selling
Each item treats a model's ability to produce new research as the product, and the public market has now priced that product.
Anthropic raised over $81B in the largest IPO on record, then traded down The offering beat SpaceX's, and the stock reportedly fell to about a $1.26T valuation after the first trade (single source). As a public company, Anthropic now sets pricing, rate limits and safety commitments under quarterly earnings pressure, and that pressure reaches every enterprise contract built on Claude. cnbc.com | reuters.com | theverge.com | theverge.com
Both leading labs now credit their models with original research Anthropic reports that Claude solved two open problems in black hole physics and published two papers on the results (vendor claim). OpenAI updated its report on how agents speed up its own AI R&D, so both labs now market research throughput, the capability that matters for target discovery, over benchmark scores. anthropic.com | x.com
FrontierML gives each agent 12 hours, real GPUs and a fixed budget per problem The benchmark covers 25 open-ended ML research problems under fixed compute limits (single source). Its design tests whether a research agent beats a human team for the same spend, which no current leaderboard measures. x.com
Near-frontier coding is now a price tier, so controlling spend is the lever
Model makers are cutting the price of near-frontier capability while a tooling layer forms to meter, route and parallelize it.
Microsoft's MAI Code 2 claims 98% of GPT-6.1 Sol at 40% of the cost Code 2 handles multi-file production work and deep debugging, and Flash handles routine edits; both are in GitHub Copilot at a 50% promotional price (vendor benchmark). Flash is also in limited preview on Foundry, which gives Microsoft an in-house alternative to OpenAI inside its own developer channel. producthunt.com
Google and Alibaba are pushing 1M-context models into cheaper tiers Gemini 4 Neon is a faster model adapted from Gemini 4 Carbon, with multimodal input and a 1M-token context, available through the Gemini API, AI Studio and Antigravity. Alibaba's Qwen4-Max, with native vision and 1M context, is free on Nous Portal for five days, so comparing the two on long clinical documents costs nothing this week. producthunt.com | x.com
Tools now measure and cut token spend automatically Trove tests token-saving fixes on copies of Claude Code sessions and rolls them back when they stop helping, and Judgment Labs' Eval Optimizer claims up to 50x savings by moving each eval judge to cheaper compute (vendor claim). Gridline offers OpenAI-compatible fine-tuning from $0.50 and batch inference from $0.029 per 1M tokens. producthunt.com | x.com | producthunt.com
Running dozens of coding agents at once is now commodity open source Hub runs hundreds of Claude Code or Codex sessions and surfaces only the decisions, Wink turns each correction into a regression check, and COLLASTRA splits one GitHub issue across agents and stores approved fixes. Fieldwork runs parallel sessions free locally or at $50 a month in cloud sandboxes, so the differentiator has shifted to remembering past corrections. producthunt.com | producthunt.com | producthunt.com | producthunt.com
Post-training rewires readout, which makes models cheap to edit and audit
Three results place the effect of post-training in how a model reads out what it already represents, not in new knowledge.
Post-training changes how a model reads its representations, not the representations One paper finds post-trained language models keep their internal representations but read them out differently (preprint, unreplicated). A second finds RL post-training in diffusion transformers works essentially as activation steering (preprint, unreplicated), which suggests fine-tuning on clinical data may change behavior without adding domain knowledge. arxiv.org | arxiv.org
astro-weights cut Qwen3.8-27B output tokens by 30% with no detected accuracy loss An agent edits open weights through an MCP server, the library scores the edited model against the base on the same problems, and the result ships as a LoRA adapter or a merged GGUF. On MATH-500 accuracy moved -0.6 points (95% CI -2.6 to +1.4), the paired, interval-reported evidence a regulated model change needs. producthunt.com
Deception shows up as separate knowledge and persona directions The Geometry of Lies finds distinct directions in representation space for what a model knows and for the persona it adopts, and uses them to detect deceptive output (preprint, unreplicated). Probes at this level are cheaper monitors than output classifiers for models that summarize patient records. arxiv.org
Agent security is shipping as product before the forensics exist to back it
Vendors are selling agent offense and defense while researchers are still working out how to reconstruct what an agent did.
OpenAI now sells a full agent security product line Daybreak Red for offense, Daybreak Blue for defense, a Codex Security plugin and pull-request scans all got deployment guides the same day. Security review of agent-written code is becoming a feature bundled with the coding agent, which puts both writing and auditing the code with one vendor. openai.com | openai.com | openai.com | openai.com
Reconstructing agent incidents is still an open research problem ROGUE-AI proposes a framework to replicate, audit and investigate security events involving AI agents, and a second paper traces incidents backward through LLM agent applications (preprints, unreplicated). Both treat reconstructing agent actions as unsolved, so audit trails for agents that touch PHI have to be designed in rather than bought. arxiv.org | arxiv.org
Image-editing systems screen the prompt but not the image Multimodal image-editing systems check text prompts for safety but mostly miss unsafe content in the input images (preprint, unreplicated). Any guardrail stack built text-first is likely to share this gap, including ones wrapped around imaging models. arxiv.org
Dropped
13 items: no real content, off-topic, or opinion or engagement posts without a claim.