Newsletter Archive

Anthropic's Claude Mythos model dramatically outperforms Opus across benchmarks (77.8% on SWE-bench Pro vs 53.4%) but won't be publicly released—instead it's finding decades-old software vulnerabilities so skillfully (181 working Firefox exploits vs Opus's 2) that Anthropic is restricting access to 12 companies under Project Glasswing, backing the security effort with $100M in credits and $4M in donations. Meanwhile, Meta released Muse Spark, a model positioned between Sonnet and Opus, after a year of billion-dollar spending.

Ben's Bites · 5 min read Industry

MIT researchers developed CompreSSM, a control-theory-based method that prunes unnecessary components from state-space models during training rather than after, cutting computational overhead without performance loss. The technique identifies and removes dead-weight model components early, enabling smaller, faster models to emerge naturally through the learning process.

MIT AI News · 8 min read Research

Sentence Transformers v5.4 now supports encoding and comparing text, images, audio, and video in a single API, enabling visual document retrieval, cross-modal search, and multimodal RAG pipelines through unified embedding and reranking models.

Hugging Face · 12 min read Tools

CoreWeave landed a $21 billion contract to supply AI compute to Meta, signaling massive infrastructure bets by Big Tech. Meta simultaneously released a new model competing more aggressively with rivals, while Anthropic's secondary share sale drew so much investor demand that some participants got cut off from their desired stakes.

Bloomberg Tech · 1 min read Industry

Anthropic's Claude Mythos announcement sparked renewed fears about open-weight AI enabling cybersecurity attacks—a cycle that repeats with every frontier model release. The author argues the debate conflates separate issues and that the 6-18 month lag between closed and open capabilities actually benefits security readiness.

Interconnects · 7 min read Community

xAI filed suit to block Colorado's new law mandating safeguards against discrimination in AI-driven employment decisions and other autonomous systems—testing whether states can regulate AI bias at the application level.

Bloomberg Tech · 1 min read Policy

Alibaba leads a $290M funding round for Shengshu, a startup building general world models to move beyond LLM limitations—the capital push signals industry recognition that embodied AI and robotics require fundamentally different architectures than language models.

CNBC Tech · 1 min read Market

Meta committed an additional $21B to CoreWeave for AI compute infrastructure, signaling a dramatic scaling-up of training capacity for frontier models and competition with other hyperscalers for GPU access.

SiliconANGLE · 5 min read Industry