Search for a command to run...
Articles tagged with #llm
How Shopify, Meta, Intercom and Ramp replaced frontier APIs with open models post-trained via SFT and RL, cutting serving costs by up to 96% on domain tasks.
Why multi-turn agents fail past step 20, why most RL rollouts give zero gradient, and the tricks that fix it: CISPO, difficulty filters, gated rewards.
A small GRPO experiment: ISO-AdamW scored 75.8% versus AdamW's 75.4% on GSM8K, with 47% more GPU memory. The idea, the 44x kernel detour, and what four answers mean.
Qwen 3.8 27B throughput from concurrency 1 to 64: Together, Fireworks FP8, Doubleword, vanilla vLLM, and MTP4 speculative decoding, with latency and repeat counts.
Why LLMs master induction and deduction but struggle with abduction, inventing new hypotheses. Insights from DeepMind's Tom Zahavy on AI reasoning limits.
Why physical AI is the next step for reasoning models: long-horizon tool loops, hardware as stateful APIs, humanoid robots, and Isaac Lab gear assembly.
