Today's News
Liberty's Highlights
Anthropic published research showing Claude Mythos Preview autonomously discovered improved attacks against HAWK, a post-quantum digital signature scheme under NIST review, and found a new attack on round-reduced AES. The HAWK attack reduces effective key strength by half—requiring key sizes to double to maintain security—despite the scheme surviving two years of expert human review. The AES attack improves previous best-known attack speed by 200–800×, though neither finding affects production systems (HAWK is not yet deployed; the AES attack targets a weakened variant). Each result cost approximately $100,000 in API compute to develop. Anthropic followed responsible disclosure procedures, sharing findings with NIST, government partners, and industry before publication. The research demonstrates frontier AI models' potential to stress-test cryptographic algorithms before and after deployment, and Anthropic partnered with academics at ETH Zurich, Tel Aviv University, and TU Berlin to build CryptanalysisBench, a benchmark for evaluating LLM cryptanalytic capabilities.
Nathan Lambert: RL data quality bottleneck and Jevons paradox reshape AI stack competition
Interconnects by Nathan Lambert
Nathan Lambert argues two structural shifts are reshaping competitive dynamics across the AI stack. First, despite leading labs seeing clear ROI from reinforcement learning data purchases, average output quality from RL data vendors remains "remarkably low"—representing significant low-hanging fruit for improvement as the field scales agentic systems. Second, as cost-per-inference declines dramatically, competitive advantage will shift from raw model performance toward understanding how to build and orient agents for specific use cases (citing Meta's Muse as an early example), rather than pushing frontier model capabilities. This signals a meaningful reorientation of where value creation happens across the stack, from model labs toward application and agent infrastructure layers.
AINews
TypeSafe's Jev decision model API reached >$100M ARR and $7.5B valuation within 3 weeks of launch, with the decision model category rapidly consolidating as the reference point. Competitors including OpenAI (Decisions API on GPT-6 Luna), Microsoft, Perplexity, Cloudflare, and others shipped similar products simultaneously, signaling explosive demand for typed outputs over free-text generation and potential winner-take-most dynamics in a newly crystallized product category. Concrete infrastructure gains are already measurable: LangChain reports that routing coding tasks to cheaper decision models cut median cost per task by 64% compared to always using frontier models, with no measurable quality regression (29.2% merged PR rate routed vs. 27.3% control). Additional signals include vLLM's locality-aware MoE achieving 1.2× faster decode on GB200 and SGLang achieving 5.9% end-to-end gains through MoE tail fusion—demonstrating that the decision model pattern is reshaping both application economics and inference system design.