Article archive
Every article, newest first.
- Haiku 5.5 makes small models more capable—and their handoffs more consequential 2026-10-11
- Mistral Large 4 makes seeing the evidence part of the reasoning problem 2026-10-11
- Reflection’s Beam asks whether expensive training can buy cheaper reasoning 2026-10-11
- EmbeddingGemma 2 makes one model the meeting point for text, sound and images 2026-10-11
- DeepSeek splits the work of reading and writing—and challenges how AI models are priced 2026-10-11
- GPT-6.1 Sol moves the model frontier from maximum intelligence to affordable intelligence 2026-10-11
- Gemini Robotics 2 makes the cost of changing bodies an AI-model problem 2026-10-11
- Physical Intelligence’s π0.7 asks whether better instructions can replace another round of robot training 2026-10-11
- Mecka’s human demonstrations test what robot models can learn before touching a robot 2026-10-11
- TwelveLabs targets the gap between human video and usable training data 2026-10-11
- Snorkel’s training environments turn expert judgment into a model curriculum 2026-10-11
- Datology puts the training budget inside the dataset 2026-10-11
- NVIDIA turns synthetic training data into a specification to inspect 2026-10-11
- Mistral’s training recipe makes the curriculum a moving target 2026-10-11
- NVIDIA’s four-bit training makes precision a budget decision 2026-10-11
- FlashAttention-4: Faster GPUs need smarter software to speed up AI training 2026-10-11
- Thinking Machines’ SQL study puts expertise inside the training signal 2026-10-11
- vLLM’s split serving stack makes predictable inference the real prize 2026-10-11
- SGLang’s smaller cache could make long-context inference cheaper—but accuracy sets the limit 2026-10-11
- DFlash 2 makes faster inference a question of better guesses 2026-10-11
- PrismML’s 5.9 GB model tests whether compression can move inference off the server 2026-10-11
- OpenAI’s Ironclad test reveals why a better AI score is not a finished job 2026-10-11
- Surge’s spreadsheet benchmark makes missing evidence an AI capability problem 2026-10-11
- Mercor’s accountant test changes the question from AI accuracy to the cost of accepted work 2026-10-11
- Harvey’s new AI test shows why completing the checklist is not enough 2026-10-11
- Factory’s legacy-code test reveals the missing ingredient in autonomous coding 2026-10-11