Anthropic made two governance moves at once: a 5% share commitment to a public-benefit fund and employee-like access for an outside watchdog, both aimed at making its safety claims checkable rather than asserted. Separately, researchers keep probing what happens when agents run unsupervised for a long time, from a 30-hour desert-island experiment with Claude to new training methods for recovering from failures in long-horizon tasks. Disney's absorption of most of Character.AI's staff continues under a bankruptcy filing that doubles as liability protection, not just an acquihire. On the research side, several groups are chasing efficiency: a frozen, untrained vocabulary table that matches learned embeddings on small models, and a data-free distillation method that cuts diffusion inference to one step. None of this changes the near-term picture for small and midsize businesses, but it sharpens who is actually accountable for agent behavior as autonomy increases.
Anthropic committed 5% of its shares to a new public-benefit fund and separately gave an outside AI watchdog employee-like access, including the ability to publish findings without Anthropic's sign-off. Both moves push accountability outward rather than relying on internal review, which matters as Anthropic's models take on more autonomous, longer-running work.
Anthropic will commit 5% of its shares to a newly created public-benefit fund, tying a slice of its equity upside to a stated public-interest mission.
An outside watchdog organization now has employee-like access inside Anthropic and can publish its findings without needing Anthropic's approval.
Disney agreed to hire the majority of Character.AI's employees as part of a stalking-horse bankruptcy bid, with most offered staff accepting. Reporting from Wired frames the bankruptcy filing as a deliberate move to shed legal liability tied to lawsuits alleging the chatbot encouraged self-harm among minors, while Disney gains trust-and-safety technology and other IP.
Disney will hire most Character.AI staff under a stalking-horse bid, though a higher competing offer could still emerge before the deal is final.
Wired reports the bankruptcy filing is structured to let Disney acquire Character.AI's trust-and-safety tech and other IP while leaving behind lawsuits alleging the chatbot encouraged self-harm in minors.
Two papers and a product release target efficiency rather than raw capability. A frozen, untrained 16 MB vocabulary table matches or beats a fully learned embedding table on small language models while enabling a 46 MB ternary model that runs five to six times faster on CPU. A data-free diffusion distillation method reaches state-of-the-art one-step image generation without any training data. Mistral also shipped a dedicated time-series foundation model line.
Mistral released Timestral, a dedicated foundation model line for time-series forecasting.
Several papers target the same problem from different angles: keeping LLM agents reliable over long, multi-step tasks. BELIEFWEAVER recovers from failures rooted several steps earlier; FC-SWE trains coding agents on their own recovery attempts; Dream2Discern predicts failure before it happens; OfficeTown and SciPivot probe organizational and judgment aspects of autonomous work. Together they signal a shift from single-shot task success toward process reliability.
BELIEFWEAVER tracks a persistent belief state across attempts so long-horizon agents can trace failures back to their true earlier cause and improves aggregate score by 5.1 points over a native baseline.
FC-SWE feeds a failed patch and verifier feedback back into training as context for a recovery attempt, improving SWE-bench Verified Resolved@1 to 41.7% versus 38.9% for standard GRPO.
Dream2Discern generates hard synthetic failure examples from a limited trajectory corpus to train runtime auditors that predict multi-agent task outcomes, hitting over 90% weighted accuracy without new data collection.
OfficeTown simulates professional organizations where agents delegate, review, and accumulate context across assignments, producing a 406-task benchmark and training data that improves Qwen3.5-9B on GDPval.
Built from six months of real human corrections to autonomous research agents, SciPivot finds four of five frontier models perform below chance at choosing the human-corrected research decision over the original LLM plan.
Two papers and a product release target efficiency rather than raw capability. A frozen, untrained 16 MB vocabulary table matches or beats a fully learned embedding table on small language models while enabling a 46 MB ternary model that runs five to six times faster on CPU. A data-free diffusion distillation method reaches state-of-the-art one-step image generation without any training data. Mistral also shipped a dedicated time-series foundation model line.
Replacing a small model's learned embedding table with a frozen 16 MB table of fixed binary codes ties or beats the learned version on most benchmarks while cutting trainable parameters 2.6x.
A simulation-free, data-free distillation method unifies sample generation and score estimation in one student network, reaching FID 2.04 on ImageNet 256x256 with a single function evaluation.