Top story
OpenAI previews GPT-5.6 Sol amid US AI regulatory drama. OpenAI unveils its next-generation model on a Cerebras deployment promising up to 750 tokens/sec, while the US government effectively gates individual access and OpenAI publicly warns restrictions shouldn't become the norm. OpenAI Blog · The Verge · TechCrunch · Zvi Mowshowitz
Research
Combining Language Models Has a Co-Failure Ceiling. Empirical analysis across 67 frontier models shows ensembling, routing, and mixture-of-agents hit hard limits from correlated failures. HuggingFace Papers
Information-Aware KV Cache Compression for Long Reasoning. A compression technique tailored to long-reasoning workloads that preserves critical context. HuggingFace Papers
Neglected Free Lunch from Post-training for LLM Agents. Argues post-training offers an underexploited performance boost for agentic LLMs. HuggingFace Papers
CoffeeBench: Long-Horizon Agents in Multi-Agent Economies. New benchmark for evaluating LLM agents across extended, heterogeneous multi-agent settings. HuggingFace Papers
Tools
DeepSeek-V4-Pro-DSpark Released. DeepSeek ships V4-Pro-DSpark alongside the DeepSpec paper on Hugging Face. Reddit/LocalLLaMA
audio.cpp: 12 TTS Models in One C++/ggml Runtime. Unifies Qwen3-TTS, PocketTTS, VeVo2 and others, claiming up to 5x speedup over Python on CUDA. Reddit/LocalLLaMA
llama.cpp Vulkan Tensor Parallelism. New PR makes tensor parallelism viable over Vulkan backends. Reddit/LocalLLaMA
Reward Hacking Debugger for RL Training. Tool that flags reward hacking behavior in RL reward functions during training. Reddit/MachineLearning
Industry
Trump Admin Releases Anthropic Mythos to 100+ US Companies/Agencies. Government clears Mythos 5 for trusted organizations including Cisco and JPMorgan, weeks after export block. TechCrunch · The Verge
OpenAI's Jalapeño Chip Marks Spiciest Move Away from Nvidia. OpenAI's custom silicon signals intensifying pressure on Nvidia's AI hardware dominance. TechCrunch
Mistral Previews New Open-Weight "Fat But Sparse" Model. Mistral announces open-weight family coming this summer with partner early access in July. X/Grok
Community
Google Losing Top AI Researchers Reveals the Real Moat. Departures of Shazeer, Jumper, Adler, and Pritzler argue the moat was human capability, not weights. Reddit/artificial
The Underrated Value of Open-Weight Models. Argues the real win is being allowed to post-train, distill, and own derivative artifacts, not local inference. Reddit/artificial
100% of CIOs Budgeting for AI Despite Public Skepticism. RBC survey of 100+ CIOs shows enterprise adoption is diverging sharply from public AI discourse. Reddit/artificial
AI Deciphers Burnt Herculaneum Papyrus Scroll. Machine learning helps read a Vesuvius-era scroll carbonized nearly 2,000 years ago. CNN via Reddit