Pulse (last 3d) · Research 88 · Agents 73 · Products 50 · Industry 48 · Models 36 · Enterprise 21 · Legal 19 · Open Source 17 · Releases 14 · Media 14
Trending · Nvidia 16 · Anthropic 14 · Jensen Huang 12 · Claude 11 · Codex 11 · Google 10 · The Verge 10 · Claude Code 9 · ChatGPT 8 · Cursor 7 · OpenAI 7 · MCP 6
Top story
Agents can now spend money, issue their own API keys, rent machines and sit in meetings, and two papers this week show their memory can be poisoned in ways that content and lineage filters provably cannot catch. Under the infrastructure, Nvidia's CEO change takes away the industry's main GPU backer just as CoreWeave, SoftBank and a Fed governor all point to strain in compute financing. Any agent that touches clinical data needs provenance controls on its memory writes and its spending, and compute contracts signed this quarter need exit terms. arxiv.org | nvidianews.nvidia.com
The Shortlist
- A machine-checked TLA+ proof shows that content- and lineage-based defenses for agent memory fail under laundering. arxiv.org
- MemoryDojo publishes the first poisoning numbers that compare five agent memory architectures. arxiv.org
- hEVO trains a genomic foundation model on 4.78B haplotypes from 63,809 whole genomes. pubmed.ncbi.nlm.nih.gov
- Isomorphic's "first AI-designed drug in Phase 2" claim has no ClinicalTrials.gov registrations behind it. clinicaltrials.gov
- ChatGPT Wallet lets OpenAI's Dots agents check out on a user's behalf. openai.com
- Anthropic finds groups do worse when each person has their own agent than when they share one. alignment.anthropic.com
- Nvidia announces a CEO transition, ending Jensen Huang's 33 years in charge. nvidianews.nvidia.com
Also Noted
- Gemini 4 Hydrogen - Google's low-cost agentic coding model, with a new Orchestrator option in Antigravity. blog.google | antigravity.google | producthunt.com
- Qwen4-Max-Preview - preview of Alibaba's flagship for reasoning, coding and multimodal work. producthunt.com
- A frontier machine intelligence for human neuroimaging - a neurosurgery lab's foundation model for brain imaging (preprint, unreplicated). huggingface.co
- A DeepSeek Moment for Biotech - Bloomberg on Chinese drugmakers such as Akeso closing the gap. bloomberg.com
- Linear Regression with Self-Selection Bias in Polynomial Time - an efficient estimator for data where subjects select themselves into the sample. arxiv.org
- Scientists mostly use Claude to write - an Anthropic report finds writing, not coding, is the main scientific use. x.com
- Fable 5.2 robotics jump - Anthropic reportedly saw a sudden gain in robotics from a general model (single source). x.com
- JustRUA - a controlled study of GPT-6 acting as a robot-use agent. huggingface.co
- Perplexity Decider v1.3 - training and eval details for Perplexity's fast decision model. perplexity.ai
- ggml v1.0.0 - the tensor library under llama.cpp reaches 1.0. github.com
- Arca by Brevis - pays users weekly for coding-agent session logs, scrubbed on the user's machine first. producthunt.com
- Blendra - turns recorded Chrome flows into Playwright tests without an LLM, so output is deterministic. producthunt.com
- Claude Duet hands-on - The Verge finds Anthropic's agent anxious in manner but useful. theverge.com
- LangChain - the agent framework company is now valued at $3B. techcrunch.com
- Saturated research evals - four frontier models scored 97.8 to 100 on a 7-part materials-science test. reddit.com
- When Both Sides Adapt - why evaluating interactive LLM systems breaks down when users adapt too. arxiv.org
- DeepStress - a framework for stress-testing deep search agents. huggingface.co
- LiveGeo - a benchmark for geographic agents in worlds that keep changing. arxiv.org
- AgentSSL - tests whether ML-engineering agents make use of unlabeled data. arxiv.org
- Learning the Task or Learning the Verifier? - showing criteria to embodied agents may teach them the verifier instead of the task. huggingface.co
- Who Remembers What? - memory extraction at several levels of detail for multi-user, multi-agent settings. huggingface.co
- CoReMem - memory that accounts for its own cost across a series of coding tasks. arxiv.org
- Metal Sandbox - isolated execution and scheduling for GPU coding agents on Apple Silicon. arxiv.org
- Classical Energies Learn to Reason 1,000x Faster than Looped Models - energy-based models reportedly learn reasoning tasks much faster (preprint, unreplicated). arxiv.org
- Negative-Only Policy Optimization - RL for verifiable rewards that only signal failure. arxiv.org
- Experience-Handoff Reinforcement Learning - an RL scheme aimed at recursive self-improvement. huggingface.co
- The Value of Thinking - learning when hybrid models should switch on reasoning. huggingface.co
- Learning When Diffusion Language Models Should Regenerate - learned token regeneration for diffusion LMs. arxiv.org
- Goal Encoding in LLMs - examines how goals are represented inside models. arxiv.org
- Sketching Language Models for Context Attribution - a cheaper way to trace outputs back to context. arxiv.org
- Towards a Theory of Learnability via Debate - formal limits on what debate between agents can teach. arxiv.org
- Ballad - bandit-based LLM routing for automated heuristic discovery. arxiv.org
- ParaTrace v8 - an open-weight 1B detector at 99.2% TPR and 1% FPR, 5th on RAID. reddit.com
- JEPA world models - one practitioner's findings from 600 planning experiments. haddyalchaer.com
- Back to Pixels - moves latent video diffusion into pixel space. huggingface.co
- Unbundling Multimodal Vision Encoding - breaks vision encoding in multimodal models into its separate parts. huggingface.co | arxiv.org
- Open Vision Banana - a fully open recipe for generative visual perception. huggingface.co
- EgoX - cross-embodiment manipulation learned from first-person video. huggingface.co
- Consistent Video-to-Video Translation via Explicit Correspondences - better frame-to-frame consistency through explicit matching. arxiv.org
- TurtLES - a large benchmark for neural surrogates of turbulent 3D PDEs. huggingface.co
- Two mathematical models of knowledge distillation - theoretical accounts of how distillation works. arxiv.org
- Theoretical and Empirical Analysis of Sparse Convolution - theory and measurements for sparse convolution. arxiv.org
- Feedback Delays in Adversarial MDPs - optimal √S and √K regret dependencies. arxiv.org
- Linking Rapid Learning to Compositional Structure - additive analogies in meta-learned parameter spaces. arxiv.org
- ArchSpace - a system for recording and reviewing LLM architecture research. arxiv.org
- Self-Evolving Multimodal Agents - a survey across application scenarios. huggingface.co
- Programming Quantum Circuit Parameters with LLMs - LLMs set quantum circuit parameters. arxiv.org
- Scalable Conversational Assessment - a two-agent LLM pipeline that runs chat-based exams. arxiv.org
- The Narrative Gap - whether LLMs help readers follow narratives across languages. huggingface.co
- ICML 2027 changes - two-stage review, a cap of 20 submissions per author, and mandatory reviewing by authors. reddit.com
- Desk-rejection grounds - top-venue calls for papers since 2013 have shifted from formatting rules to reviewer conduct and LLM policy. reddit.com
- Slop Is Moving Up the Stack - an essay on low-quality AI output reaching more sophisticated work. gabeorlanski.github.io
- Incident.io's AI spend - a line-by-line look at one company's AI costs. blog.pragmaticengineer.com
- Gemini smartglasses - Google's glasses start at $395 and lead with privacy. arstechnica.com
- SpaceX supercompute - more than $74B for AI campuses in Texas and Arizona. reuters.com
- ByteDance timeline - Zhang Yiming reportedly sees no parity with US frontier labs before 2029 or 2030. bloomberg.com
- Federal training permit - a Democratic draft bill would require a permit to train advanced models. futurism.com
Agent memory is now the injection surface, and filtering it is provably not enough
A proof, a benchmark and an exfiltration attack all put the weak point in what agents write down and reuse.
Laundering defeats every content- or lineage-based memory defense TMA-NM (Louck, Ariel University) defines three laundering channels: an agent summarizing itself, a trusted tool echoing input back, and manufactured corroboration. A machine-checked TLA+ proof shows that defenses keyed on content or lineage are unsound against all three. Any agent that summarizes chart notes or tool output into long-term memory gives a poisoned input a route to trusted status. arxiv.org
MemoryDojo measures poisoning across five memory stacks Microsoft and RPI ran multi-session AgentDojo trajectories against A-MEM, Graphiti, Letta, Mem0 and Memora, the first comparison of how poisoned entries persist and spread across memory architectures (preprint, unreplicated). Its ranking of defenses independently confirms rankings reported earlier, so the choice of memory backend is now a security decision with comparative evidence behind it. arxiv.org
Synthesis smuggling hides inferred attributes in ordinary output In this attack the model infers personal attributes and hides them inside responses that look normal (preprint, unreplicated). Output-side PHI filters that look for literal identifiers will miss attributes the model inferred and then concealed. arxiv.org
Agents are becoming economic actors, and control is moving to budgets and gates
Payments, agent-issued API keys, per-second sandboxes and merge gates all reflect one change: agents act with money and machines, and people govern them through caps and checkpoints.
OpenAI gives agents a checkout button while Amazon builds the catalog ChatGPT Wallet lets OpenAI's Dots agents shop and complete purchases for users. Separately, Amazon plans a world product graph that indexes online shopping data for AI search. Agent commerce now has both a payment rail and an index, the two pieces earlier shopping-agent demos were missing. openai.com | techcrunch.com | theverge.com | theverge.com
Agents now provision their own infrastructure under prepaid caps On Cloudish, an agent creates its own API key with one call and deploys long-lived containers on prepaid credits billed per second. Nevermined Catalog lets agents find, quote and pay for 160 scored services within a stated budget (press release). The key's balance, not a human approval, is now the control point. producthunt.com | producthunt.com
Isolation and merge policy are where teams take back control Shelter41 and Offstage rent isolated cloud and macOS machines to coding agents by the second, and a 10-minute Offstage run costs about $0.18. Keelson Merge Gate makes a person approve any agent pull request that touches workflows, Dockerfiles, infrastructure or CODEOWNERS. producthunt.com | producthunt.com | producthunt.com
Shared agents beat personal ones, and Anthropic is shipping the shared one Anthropic's alignment team finds that groups perform worse when each member has a personal agent than when they share one. Claude Meetings, in beta on Team and Enterprise plans, adds one visible Claude to a Google Meet call to take notes and build artifacts that attendees can edit. Tumor boards and data-review meetings are the obvious clinical fit. alignment.anthropic.com | producthunt.com
Genomic models are getting more specific while claims about AI-designed drugs outrun the registry
The models are becoming more biologically precise just as public claims about AI-designed drugs become harder to check.
Variant-effect modeling moves down to the haplotype and the cell type hEVO is a haplotype-level genomic foundation model trained on 4.78B haplotypes from 63,809 whole genomes. DeepMind's AlphaGenome 2 predicts how single-letter DNA changes affect disease in specific cell types (press release). Together they make phased, cell-type-specific variant scoring practical, which puts a premium on phasing quality in internal WGS data. pubmed.ncbi.nlm.nih.gov | x.com
Isomorphic's Phase 2 headline has no registry entry behind it The claim that Isomorphic Labs has the first AI-designed drug in Phase 2 fails a primary-source check. A ClinicalTrials.gov query returns no registrations at all, not even for its first-in-human Phase 1 work. Any comparison of AI-discovered drugs by clinical milestone should list Isomorphic at zero registered trials. clinicaltrials.gov
Without Huang, Nvidia no longer backstops AI compute financing
A leadership change, demands for prepayment, a write-down and a Fed warning all point to compute capital getting more expensive.
Huang's 33-year tenure ends, and the "Nvidia put" may go with it Nvidia announced a CEO transition that ends Jensen Huang's tenure, amid reports of a personal scandal (single source). Founders and VCs say his exit leaves a hole in startup funding. Semafor argues it ends the implicit Nvidia backstop under the AI trade, and a separate report alleges Nvidia funded startups to create demand for its GPUs (single source). nvidianews.nvidia.com | apnews.com | latent.space | wsj.com | thezvi.substack.com | techcrunch.com | semafor.com | techpowerup.com | theverge.com
Compute financing is under strain from three directions CoreWeave wants $9B in customer prepayments and is moving to shorter contracts even as revenue rose 168% to $3.66B. SoftBank took a $3.7B charge after over-ordering data-center power equipment, and Fed Governor Bowman warned that AI debt and valuations could trigger a correction. October corporate AI spending set a record, but monthly growth slowed to 1.65%. techcrunch.com | semafor.com | semafor.com | the-decoder.com
Hugging Face's future and China export policy are both in play Hugging Face says Nvidia's $12.9B acquisition of it is still expected to close in 2027, and its CEO vows to keep fighting for open-source AI. Export-banned Nvidia chips power China's strongest models, and OpenAI is reportedly joining Anthropic in backing a bipartisan export-control bill aimed at passage before year-end (single source). techcrunch.com | techcrunch.com | x.com
Dropped
7 items: no content, promotional tutorials, opinion video, non-AI hobby projects, or a running thread with no new finding.