The Archive

DeepSeek V4 Flash and Pro models now available via API, enabling developers to run open-weights inference at scale.

He said it on the [Dwarkesh Podcast ](https://mrkt30.com/anthropic-mythos-triggers-chinas-ai-arms-frenzy/)this week and I have not been able to stop thinking about it. His argument was not that China is not a threat. It was that cutting them off and treating them as an enemy is probably not the smartest long term play. His actual words were that victimising them and turning them into an enemy likely is not the best answer. The context here is Huawei targeting 750,000 AI chip shipments this year. It is nowhere near Nvidia's compute but the direction of travel is clear. And if DeepSeek ends u...

DeepSeek releases DeepEP V2 and TileKernels optimization libraries on GitHub.

Prompt optimization for LLM-as-a-Judge evaluation on legal QA; tests transfer across Qwen3-32B and DeepSeek-V3 judges on LEXam benchmark.

GPT-5 and DeepSeek-R1 exploit formalization-faithfulness gap in Lean 4 proofs despite valid logical reasoning; evaluates on FOLIO and Multi-LogiEval.

Moonshot releases Kimi K2.6, an open-weight model claiming performance parity with Claude Opus 4.6.

Cross-linguistic study of politeness effects on 5 LLMs (Gemini-Pro, GPT-4o Mini, Claude 3 Sonnet, DeepSeek-Chat, Llama 3) via 22,500 English/Hindi/Spanish prompts.

How to try DeepSeek-R1 for coding via the Zed code editor's out-the-box support.