A Monday of open releases: Google lets agents rewrite their own harness, Tencent ships a model that decides what agents should remember, and NVIDIA opens a 2.8-trillion-token Indic corpus. Off the wires: Anthropic gives away 5% of itself and the web starts closing its doors to agents.
Anthropic said it will commit 5% of its shares to a new Public Benefit Fund, subject to stockholder approval, with the stake growing as the company's valuation rises. Combined with the founders' pledge to donate 80% of their equity, roughly 20% of Anthropic's shares would be dedicated to public benefit.
After legal experts warned that new commercial terms could let Meta scrape any publisher with a business account to train AI, Meta deleted the clause. The company says the terms only ever covered ads and business products and that its crawling practices have not changed.
The Verge reports a wave of consumer companies shutting out AI agents: Yelp admits only agents that pay for its data, Delta blocks agents from booking tickets, and JPMorgan Chase and Charles Schwab bar agents from customer accounts over security concerns.
Google Research open-sourced Harness², a recursive agent harnessing framework that improves the prompts, tools, memory, skills and orchestration around a frozen model, with no ground-truth verifiers needed. A contrastive Propose-Probe-Compose step refines the harness per task, while a second recursion level distills reusable lessons for future tasks. One step lifted GLM-5 on WorkBuddy-Bench from 64.2 to 75.7, with gains on all four model-harness setups across three professional benchmarks.
Most agent gains are expected to come from bigger models. Harness² bets the cheaper lever is the scaffolding around the model, and ships it as Apache-licensed code anyone can run. If the bet is right, the next round of agent progress looks like better harnessing, not just bigger weights.
Tencent Hunyuan released Hy-Memory-30B-A3B, a 30B-parameter MoE model with about 3B active parameters trained specifically to write long-term memory: deciding what to retain, how to represent it, and how new evidence updates stored state. Trained with attributed reinforcement learning across the Mem0, MemOS and Hy-Memory frameworks, it ranks first on all four evaluated memory harnesses with a 55.49% macro average, ahead of Claude Opus 5 at 54.62%. Weights are open on Hugging Face under Apache 2.0.
Every agent framework today bolts a general-purpose model onto the memory write path and hopes prompting is enough. A small, open, specialized writer that outperforms frontier models at this one job could become standard plumbing for agents that need to remember across sessions.
NVIDIA and IIT Madras's AI4Bharat lab introduced Nemotron-pretraining-Indic-v1, a synthetic pretraining corpus of about 2.8 trillion tokens across the 22 constitutionally recognized languages of India: 2,601 billion native-script tokens plus 224.6 billion romanized tokens. Documents were translated with Gemma-4-31B-it and Bodhan AI's Indic Translate, quality-scored against MQM-inspired rubrics, and repaired in a feedback loop that recovered 438.4 billion tokens. The GPU-accelerated generation pipeline, built with NeMo Data Designer, is open for others to reuse.
Frontier AI still speaks a handful of languages fluently because pretraining data does. An open, reusable recipe for high-quality synthetic corpora is how the other 7,000 languages get a fair shot at the AI era, and this is the largest such recipe ever published.
Editor's note: no new computational pathology papers in this morning's sweep. The journal scan returned only off-topic keyword matches, arXiv's Monday announcement landed at press time with nothing in window, and no lab announced new pathology research. For recent pathology AI, see No. 019: Beyond the Dominant Label, Evidence Field, and the chest X-ray foundation model benchmark.
Anthropic's Frontier Red Team reports that an internal Claude research model traced the origins of the NIST ECDSA curve seeds, a provenance question open since the NSA supplied the values without explanation. The model hypothesized a 1990s-vintage pseudorandom generator, scanned the seeds for its signature, and identified the C library's lrand48 seeded with Unix timestamps: the P-256 seed matched candidate 1,842 from a run starting 1997-08-16. All in under half an hour. Anthropic stresses this is not a cryptographic break and no action is required.
A model doing original historical-cryptographic research in 30 minutes, with reproducible code, is a working demo of AI as a research instrument. It also quietly retires a decades-old backdoor anxiety: the seeds look like timestamps, not hidden messages.
Using the Epic Cosmos database of over 300 million patient records, the UTHealth Houston team tracked cervical cancer screening among 9 to 16 million patients per year from 2019 to 2024. Screening rose across every group, with Hispanic patients gaining most (9.85% to 15.51%), but American Indian, Alaska Native, Native Hawaiian and Pacific Islander patients had the lowest rates throughout, and Western and Southern regions lagged.
Screening is one of the most effective cancer interventions ever devised, and this is the largest longitudinal look at who actually gets it. The pattern, broad gains with stubborn gaps, is the baseline against which every future equity intervention, AI-assisted or not, will be measured.
Editor's note: no new generative-model papers in this morning's sweep; arXiv's Monday announcement landed at press time with nothing in window, and the journal scan's diffusion hits were off-topic. For recent diffusion work, see No. 019: RFdiffusion3's enzyme designs, CellDiffusion, and TIPS for diffusion language models.