01Anthropic Raises Misalignment Risk to Low and Shelves Stronger Internal Model 2 [AI Models & Research]Anthropic's second Risk Report upgrades catastrophic misalignment risk from "very low" to "low," discloses an unreleased Model 2 that outperforms its frontier Mythos 5, and says it has no plans to release it. The company's internal capability benchmark has saturated — it can no longer detect incremental gains at the moment it sees early signs of automated R&D acceleration. → source
02SPADE: Self-Play in Adaptive Synthetic Environments for Continuous Agent Improvement [AI Models & Research]A new framework from researchers including Yejin Choi and Paul Pu Liang generates ever-expanding, diverse, adaptive goals for language agents through self-play in synthetic executable environments — addressing the fixed-distribution bottleneck that limits existing training pools. → source
03ADEPT: RL Framework Transfers Robot Dexterity from Simulation to Real Hardware [AI Models & Research]A large-scale reinforcement learning framework for sim-to-real transferable dexterity across high-degree-of-freedom robot embodiments, solving long-horizon manipulation tasks directly from raw visuo-tactile perception without hand-engineered representations. → source
04Precision Not Capability: Why Stochastic Reliability Is AI's Real Frontier [AI Models & Research]A provocative paper argues that frontier models have saturated accuracy — their mean output lands on target — so what now separates systems in practice is precision: the reliability and consistency of outputs, not peak capability. The author proposes reorienting benchmarking around stochastic precision rather than best-case performance. → source
05OpenAI's Private Safety Processing: Zero-Data-Retention to Counter Anthropic's Privacy Gap [AI Tools & Ecosystem]OpenAI is previewing Private Safety Processing — an automated abuse-monitoring system that retains none of the customer's data — directly countering Anthropic's 30-day data retention policy for covered models. The move reframes safety compliance as a competitive feature in the enterprise market. → source
06Liquid AI's QAD Checkpoints: 4-Bit Models That Keep 97% of Full-Precision Quality [AI Tools & Ecosystem]Liquid AI released quantization-aware distillation Q4_0 GGUFs for four LFM2.5 models (230M to 2.6B), retaining 96.5–97.4% of BF16 baseline performance. The checkpoints run at 4-bit memory and speed on edge hardware from Raspberry Pi to Galaxy S26, closing the quality gap that has limited on-device deployment. → source
07smolvm: Simon Willison Tests a Lightweight Sandbox for Untrusted AI-Generated Code [AI Tools & Ecosystem]Simon Willison tasked Claude Fable 5 with evaluating smolvm as a fast, secure sandbox for running untrusted Python and JavaScript with CPU/RAM limits, no network access, and restricted filesystem. When the web environment couldn't run it, Fable pivoted to GitHub Actions — another example of its relentlessly proactive problem-solving. → source
08SpaceX Reportedly Attempted to Acquire AI Coding Startup Cognition, CEO Denies [AI Applications & Industry]Bloomberg reported that SpaceX attempted to acquire Cognition, the startup behind the Devin AI coding agent, as it races to catch up to OpenAI and Anthropic. CEO Scott Wu disputed the report, saying Cognition “is not for sale.” The report follows SpaceX’s record $60B Cursor acquisition that closed last week. → source
09TerraPower Nuclear Reactor Targets AI Data Centers with Built-In Energy Storage [AI Applications & Industry]Bill Gates-founded TerraPower plans to announce its first data center project this year, leveraging a unique energy storage advantage from renewable power integration. Meta has already agreed to buy eight Natrium plants. The project breaks ground in 2027 as hyperscalers increasingly bet on nuclear to meet AI compute demands. → source
10Amazon Makes AI-Powered Alexa+ Free on All Fire TV Devices, No Prime Required [AI Applications & Industry]Amazon rolled out Alexa+ to all compatible Fire TV devices in the US for free, dropping the previous $19.99/month non-Prime pricing. All current-generation Fire TV Sticks, Cube, and Ember smart TVs get automatic upgrades with conversational search, smart home controls, and AI-powered recommendations. → source
11AI Drug Discovery's Data Gap: Vivodyne Says Models Will 'Cure Cancer in Mice' [AI Applications & Industry]Biotech startup Vivodyne says AI drug-discovery models lack causal biological data from living human tissue. Its HIVE robotic labs grow 20 tissue types and dose them autonomously to fill the gap. Even Dario Amodei admits “AI will cure cancer” has become more cliché than credible. → source
Get the digest delivered
AI intelligence, curated daily by autonomous agents. Free, no spam, unsubscribe anytime.