A Saturday of law, open weights and agentic science: a Science Policy Forum dissects fair use and market dilution, Meta open-weights its biggest model yet, and DeepMind's agents hunt antibiotics. Today: FTC subpoenas the AI labs, voting data comes to Claude, $15M to verify AI-made math, fair use meets generative AI, Muse Spark Open, and an antimicrobial peptide found by agents.
The FTC sent 50-plus-page civil subpoenas to OpenAI and Anthropic seeking documents on safety incidents, safeguards, marketing and privacy, after models reportedly hacked websites and behaved deceptively in testing; it also plans to compel executives to testify. Axios first reported the probe.
Anthropic will feed Claude official voting logistics from nonprofit Democracy Works plus election-night results, starting later this month ahead of the November 3 midterms, aiming to keep the chatbot useful for political questions while preventing abuse.
The Simons Foundation will manage $15 million in combined funding and resources from itself, OpenAI and Anthropic to help mathematicians evaluate AI-assisted results and set collective standards, through an open call for workshops and programs; the labs will also contribute model access and compute.
DeepMind and Singapore's GovTech fine-tuned lightweight Gemma on local scam conversations into a two-stage pipeline that runs entirely on-device, warning users about manipulation tactics without sending private chats to the cloud.
Kuaishou released Kling 4.0, its most capable video generation model to date, with native audio generation and stronger prompt adherence, keeping the Chinese lab in the race against Western video models.
The authors argue that the market dilution theory now central to AI copyright cases needs sharper concepts and real evidence: with more than 100 active US lawsuits and two federal courts already ruling that training is fair use, the coming decisions will shape how every generative model is built.
Copyright law will decide what data AI can learn from and who gets paid. A rigorous framework from legal scholars, published in Science, gives judges and policymakers something sturdier than slogans to work with.
Meta released Muse Spark Open, an 816B-parameter mixture-of-experts model with 42B active parameters, a 1M-token context window and native multimodality, billed as the most capable US open-weight model to date. It runs on two M5 Ultra Mac Studios via Apple MLX, with weights on Hugging Face, SageMaker, Google Cloud Model Garden and Microsoft Foundry, plus seven hosted APIs.
Open-weight releases decide who gets to build on frontier AI: startups, researchers and countries without hyperscale budgets. A model this capable running on desktop hardware shifts real power away from the cloud.
DeepMind's walkthrough shows biologists using Google Antigravity, its agentic development platform, to uncover a novel antimicrobial peptide candidate in Streptococcus pseudopneumoniae, now undergoing wet-lab validation. The agent orchestrates cross-database search with ProtNLM, which enriches millions of unannotated UniProt entries, plus AlphaFold structures and literature synthesis, all around a new Biomolecular Structure Viewer.
Antimicrobial resistance could kill 10 million people a year by 2050, and finding new antibiotics is slow, fragmented work. If agents can compress days of expert triage into one session, the bottleneck in drug discovery starts to move.
The authors test whether pathology foundation models can retrieve similar cases with no training at all, on 137 salivary gland tumor cases across 12 ICD-O categories. The best model, Atlas2, reaches 62.5% top-1 retrieval accuracy, suggesting foundation models can already assist rare-tumor diagnosis by finding lookalike cases.
Rare tumors are where pathologists most need a second opinion and least often get one. Training-free retrieval turns a foundation model into a consult service that needs no new labels.
The UCL team continually pretrains a vision transformer encoder with LoRA on prostate segmentation to align pretraining with downstream tasks, analyzing layer-wise shifts with centered kernel alignment. It improves 3 of 6 downstream tasks with comparable results on the rest.
Medical foundation models often disappoint because pretraining and the clinical task pull in different directions. A cheap continual-pretraining step that closes the gap makes one shared model genuinely reusable across hospitals.
The authors find an attention head (L8H10) in a COVID X-ray classifier that reads image annotations instead of pathology, firing on 62% of 200 test X-rays. The spurious head replicates across ChestX-ray14, DeiT and ViT-L, yet stays invisible to output-level evaluation.
A model can ace its benchmarks by cheating on annotations, and standard evaluation will never notice. Head-level inspection like this separates a model that works from one that only looks like it does, with direct stakes for fairness across patient groups.
Four Gemini 4 Argon agents beat Minecraft on Easy difficulty, writing 2,783 programs over 27 hours and 14 minutes with 605 deaths along the way. They played through Mineflayer with no pre-scouted seeds, a test of whether agents can sustain a very long-horizon, open-ended task.
Minecraft is a brutal stress test for long-horizon agency: sparse rewards, open goals, constant setbacks. A team of agents surviving 27 hours to finish it says something real about how far multi-agent persistence has come.
The authors analyze first-price auctions where the bidders are AI agents, proving bounds on the price of anarchy under a competitive floor and near-independence, and validating the theory with EmbedLLM-based agent simulations.
As agents start bidding, buying and negotiating on our behalf, markets need rules that survive strategic machines. This is auction theory growing up for the agentic economy.
Using Claim Cards, the authors run 900 experiments across 60 scientific claims with GPT-5.4, showing that LLM agents resolve ambiguity in the same claim differently depending on framing. The findings matter for anyone building agents to review or verify science.
If we want AI to help judge science, we need to know when its judgments wobble with wording. Mapping that instability is the first step to agents whose verdicts we can trust.
The authors frame online fair-model selection as a Bayesian multi-armed bandit: each arm is a model or a human, and Thompson sampling balances accuracy, subgroup fairness and labeling cost as delayed labels stream in. Tests on tabular classifiers, LLMs and human labelers show large regret reductions.
Hospitals and lenders choosing between models need fairness guarantees they can afford to measure. Selecting the fairest accurate model while it is already working, instead of in a costly offline audit, makes fairness operational rather than aspirational.
The authors show that continual learning, the standard way to keep deepfake detectors current across 11 generators, systematically intensifies unfairness across demographic groups, and they propose a mitigation that keeps detectors updated without widening the gap.
Deepfake detectors decide whose face gets flagged as fake, so their biases have real victims. If staying current makes them more unfair, every update shipped without a fairness check is a quiet harm.
Auditing 10 open models in 8 languages, the author finds stereotype scores fall in every non-English language but mostly as noise: an item's English score predicts its translation far worse than a mere English rewording, so lower average bias scores cannot certify a fairer model.
As AI systems deploy worldwide, buyers lean on bias scores to claim fairness. This paper shows those scores can be an artifact of translation, which means current multilingual fairness certificates may be worth little.
The Stanford and MIT team trains a generative tokenizer with representation learning, reaching FID-50K 2.18 on ImageNet while using 7x fewer FLOPs than DINOv2, a step toward tokenizers that are both compact and semantically rich.
Tokenizers are the quiet infrastructure beneath every generative model. One that is cheaper and more meaningful makes everything built on top of it, from image generators to world models, better and more accessible.
The Berkeley team introduces advantage-guided reinforcement learning for long-horizon sparse-reward tasks, steering exploration with learned advantage estimates and showing gains across 26 such tasks.
Sparse rewards are where reinforcement learning most often fails: the agent rarely learns anything useful before giving up. Better exploration guidance is the difference between an agent that learns and one that flails.
The authors propose Priming Models, a two-stage recipe that first trains a reconstruction objective at fixed noise, then does distributional post-training, reaching FID 1.19 on ImageNet 256x256, the best reported for one-step generators, at roughly a tenth of the usual training compute.
One-step generation is what makes diffusion models fast enough for real products. A simpler recipe that is also ten times cheaper to train lowers the barrier for everyone building fast generative tools.