AI Daily Digest · 2026-11-18
🔬 New AI Craft
1. Together Compiler puts real forward-pass measurements ahead of kernel tuning
Together traces optimization failures to benchmarks that miss the cache behavior and inputs of a running model. Its Llama 1B experiment on H100 recovers 98% of a hand-tuned megakernel baseline using Inductor kernels, CUDA Graphs and PDL. The practical lesson is to measure the full forward pass before committing to custom kernels or agent-generated replacements; the reported ratio belongs to that configuration.
https://www.together.ai/blog/road-to-together-compiler
🛠️ Tools & Tips
1. TokenEconomizer learns repository-specific ways to spend fewer tokens
The plugin turns past coding sessions into concise repository notes, selective tool loading and prompts that interrupt unproductive loops, while leaving the base agent and model unchanged. On 277 SWE-Bench tasks across six repositories and five agent-model combinations, the authors report 18.4–30.9% lower average costs with broadly stable success rates; initial learning can cost extra.
https://www.microsoft.com/en-us/research/blog/tokeneconomizer-learning-from-experience-to-make-ai-coding-agents-more-efficient
⭐ Open Source Highlights
(no items this day)
📰 Industry News
(no items this day)
🚀 Major Releases
1. OpenAI releases GPT-6.2 Astra and Sol across ChatGPT Work, Codex and the API
2. Microsoft launches MAI Code 2 and Code 2 Flash in Copilot; Flash has limited Foundry preview access

