Pulse (last 3d) · Agents 84 · Research 83 · Products 63 · Industry 37 · Models 36 · Enterprise 30 · Legal 25 · Releases 19 · Infra 19 · Open Source 18

Trending · Anthropic 22 · OpenAI 18 · Claude 14 · Google 14 · ChatGPT 13 · Claude Code 11 · Codex 11 · The Verge 8 · Cursor 7 · Hugging Face 7 · Nvidia 7 · Apple 5


Trust is moving out of the model and into whatever checks it. Two Riemann-class proof claims landed on the same day, and one shipped as a 1.3GB Lean artifact. Most of the day's agent launches sell sandboxes, undo, approval and audit, while open models at 27B-class quality now fit in 16GB, so the question to ask of any proof, agent or model touching clinical data is what verifies it independently and whether it can run inside the perimeter. alpo.ge | producthunt.com

  • Levent Alpöge claims the Generalized Riemann Hypothesis, released as a 1.3GB Lean formalization. alpo.ge
  • An Anthropic mathematician claims a proof of the Riemann hypothesis made with Claude. theverge.com
  • Inkling 2 Mini, an open-weight 26.4B MoE, matches Qwen3.8-27B and runs in 16GB RAM. x.com
  • Docker now ships agent sandboxes and AI governance, which makes agent isolation a default part of the stack. producthunt.com
  • OpenAI now publicly treats adversarial distillation of its models as an attack. openai.com
  • Anthropic fell about 21% from its IPO open on day two as the Nasdaq erased its 2026 gains. theguardian.com
  • Commerce approved 879,000 of the 3.4 million Nvidia and AMD chips requested for Chinese buyers. theregister.com
  • Anthropic IPO - $1.8T debut and full greenshoe, then down 10% premarket on day two. cnbc.com | anthropic.com | x.com
  • Anthropic staff holdings - about 5,300 employees hold roughly $232B at the offer price, and 400-plus hold over $100M. thenextweb.com
  • Language Models Can Rewrite Their Own Weights - models generate their own weight edits, but the gains were mostly shorter reasoning traces (preprint, unreplicated). huggingface.co | arxiv.org
  • A CPU Inside an LLM - compiles exact circuits into pretrained weights with no gradient training (preprint, unreplicated). huggingface.co
  • RL Post-Training in Diffusion Transformers Is Activation Steering - argues RL post-training of DiTs mostly steers activations. huggingface.co
  • Project Suncatcher - Google booked Blue Origin's New Glenn to launch TPU satellites in Q1 2029, but the long-term launch problem remains. thenextweb.com | arstechnica.com
  • Palo Alto Networks to buy Applied Compute - about $4.5B in stock (single source). aiweekly.co
  • Alibaba earnings - net income tripled to 64.4B yuan, and AI cloud and compute revenue rose 53%. thenextweb.com
  • Trump data center tax - threatens a "tremendous tax" on companies that move AI data centers outside the US. theverge.com
  • Apple Mac agent - reportedly covers device control, apps, browsing and coding, and a job listing is the only public evidence. techcrunch.com | bloomberg.com
  • Cursor CLI bundles Grok Build 2.0.1 - one build takes Cursor or Grok sign-in and lists all 38 models, which suggests a merger (single source). x.com
  • Perplexity Teammates - Perplexity adds Teammates and multiagent features. x.com
  • Signal blocks AI screen reading - Gemini and Siri can no longer read or screenshot Signal's mobile app. theverge.com
  • Kepler by RubixKube - a local on-call agent that shows the commands it ran and asks before changing infrastructure. producthunt.com
  • tensorViz - runs one forward pass so coding agents can read measured PyTorch tensor shapes (MIT license). producthunt.com
  • Manufact - publishing checks, session traces and cross-client release tests for MCP servers. producthunt.com
  • AgentDrive - passes the goal, touched files and recent turns from one coding agent to another. producthunt.com
  • Sub - standing "AI employees" with their own email, phone and spending limits, from $1,000 a month. producthunt.com
  • Cerebellum Cloud - per-token hosting for full weights or LoRA fine-tunes, with no charge while idle. producthunt.com
  • APKSonar - Android package analyzer that detects 83 families of leaked secrets, mapped to MASVS and CVSS. producthunt.com
  • Ankhet - replaces coding tests with a 45-minute simulated workday and timestamped ratings. producthunt.com
  • Pictowalk - turns one photo into a 20-second world you can steer, built on Google's Genie 4, for $9 each. producthunt.com
  • Pluto - open-source app for Apple Silicon that links meetings, notes and tasks, with on-device transcription. producthunt.com
  • Reprise - narrates articles on device, with voice cloning, on Apple's Neural Engine. producthunt.com
  • launchroom - open-source self-hosted deploy platform with pull-request previews and S3 backups. producthunt.com
  • Vibecoding an inference chip - an inference chip designed largely through AI-assisted coding. ankitg.me
  • In Situ Training of Optical Neural Networks by Coding Agents - coding agents train photonic networks directly on the hardware. huggingface.co
  • System 1 agents - proposes a design for local, reliable, fast agents. huggingface.co
  • Phantom Success, Silent Failure - simulated manipulation benchmarks report false successes and miss real failures. huggingface.co
  • Squoosh Bench - scores UX-1 against 318 historical website A/B tests. huggingface.co
  • Time Series Structure Learning with Local Causal Parent Discovery - learns causal structure in time series through local parent discovery. huggingface.co
  • The Hungry Butler Effect - argues that concealed commercial incentives erode the epistemic rigor of LLM answers. researchgate.net
  • DeepMind on what comes after AGI - maps four routes to superintelligence, including recursive self-improvement. institute.deepmind.com
  • ML peer review crackdown - ICLR deletes suspected AI reviews and desk-rejects papers whose co-authors reviewed negligently. reddit.com | reddit.com | reddit.com
  • Teaching Vision-Language-Action Models to Stop Safely - trains VLA robot policies to stop safely. huggingface.co
  • SpecWM - turns verification into refinement to speed up streaming world models. huggingface.co
  • Direct3D-V3 - a latent-free diffusion transformer that generates 3D directly in voxel space. huggingface.co
  • AURA - generates streaming avatars driven by audio. huggingface.co
  • PAD - values autonomous-driving training data by its effect on the policy. huggingface.co
  • Lang-SVG - hierarchical image vectorization guided by language priors. huggingface.co
  • Online Mistake Detection in Instruction-Conditioned Procedural Video - flags procedural errors in video while they happen. huggingface.co
  • Dettato audio-to-MIDI - claimed to be 60x more efficient, trained only on public-domain recordings. kobi.music
  • Daniel Ek backs a fusion startup - the Spotify founder funds a fusion power company founded by an AI researcher. techcrunch.com
  • ClawDOS - an AI coding agent that runs on DOS. ricapar.net

Three Millennium-class claims in a week, two tied to Anthropic, now depend on formal or expert verification to be believed.

A Riemann claim now credits a frontier model An Anthropic mathematician claims a proof of the Riemann hypothesis made with Claude (claim, unrefereed). The real question is no longer whether models can contribute to open research problems but whether this argument survives expert reading. x.com | x.com | theverge.com | latent.space

The rival human proof arrives machine-checkable Levent Alpöge independently claims the stronger Generalized Riemann Hypothesis and released it as a roughly 1.3GB Lean formalization (claim, unrefereed). A kernel-checked proof moves review from reading the argument to auditing the theorem statement and any added axioms, and that is a much faster path to a verdict. alpo.ge | techcrunch.com

Navier-Stokes gets a second AI-assisted claimant An NYU mathematician and an Anthropic researcher say they solved the unforced Navier-Stokes problem with AI help (single source). The claim follows an earlier, disputed OpenAI claim on the same problem. x.com

Small open models are closing on mid-size quality fast enough that OpenAI now treats extracting its outputs as an attack.

Two open models put 27B-class reasoning on a laptop Thinking Machines' Inkling 2 Mini, a 26.4B looped-transformer MoE, matches Qwen3.8-27B on the Artificial Analysis Intelligence Index in 16GB of RAM (vendor benchmark). OVHcloud's Apache 2.0 Olala-v1.0-mini-7A1B, a Mamba-3/attention hybrid with about 1B active parameters, runs near 72 tokens per second on a 16GB MacBook Air, which makes PHI-local inference a laptop-class deployment (vendor benchmark). x.com | x.com

Local agent stacks now cut tokens, not just weights Dial, an open-source coding agent, re-sends only changed lines, summarizes large files, batches reads, runs the tests before it calls a task done, and serves models through MLX, llama.cpp, Ollama or LM Studio. LOMI runs GGUF models offline on Android, and a widely read post argues for compression over buying RAM. producthunt.com | producthunt.com | curiousbits.nl

OpenAI names adversarial distillation as a threat OpenAI has published its approach to combating adversarial distillation of its models. Tighter output-use terms and detection on API traffic would hit any team that trains internal models on labels generated by frontier models. openai.com

Most of the day's agent launches sell a control on the agent, not a new capability.

Docker now treats agents as a security subject Docker repositioned around agents as actors rather than workloads. It shipped a trusted build supply chain, Docker Sandboxes as an isolated runtime, and Docker AI Governance across teams (press release). Now that the default container vendor isolates agents, sandboxed execution becomes the baseline auditors expect for agents that touch regulated data. producthunt.com

Undo is cheaper than pre-approval UndoGuard is a rollback proxy and MCP server that snapshots and AES-256-encrypts state before every mutating API call, so a team can revert one call or a whole agent session. Reversibility at the API boundary maps directly onto audit-trail requirements and avoids gating every write. producthunt.com

Draft-only permissions are becoming a design pattern Authentify checks every agent action, sends the consequential ones to a human, and keeps a tamper-evident log that regulators can verify (press release). Busymail applies the same split to email: an MCP token can read, search, archive and draft but cannot send, and approval needs a signed-in session. producthunt.com | producthunt.com

Agent inventory tooling is filling in keypeek is a read-only offline CLI that shows which API keys coding agents can read through env files, chat transcripts and shell history. 3rd Eye and Timely Agent Tracking show what each agent can reach, whom it calls, and spend by model and project, which is the inventory a security review asks for first. producthunt.com | producthunt.com | producthunt.com

Labs are building drug-discovery evaluations, and competition results show where current tools stall.

OpenAI launches a medicinal-chemistry benchmark OpenAI introduced MedChemBench to test whether models can help scientists identify promising drug candidates (vendor benchmark). A benchmark built by the lab whose models it scores sets the bar those models clear, so replication on proprietary assay data carries more weight than any headline number. openai.com

Binder design stalls on hard targets Proteus Bio and FinalDose compared three AI architectures in the Adaptyv x Anthropic binder competition. They found that multi-constraint targets such as TNF need reasoning beyond current state-of-the-art tools (single source), so hard targets in design pipelines still need humans to handle the constraints. x.com

18 items: off-topic content, link roundups, posts and podcasts with no stated content, and sentiment posts with no claim.