Next-Concept Modeling: open MIT weights, 10–31× faster text gen
Chris Ré's lab beats diffusion LMs with checkpoints you can run today — plus Codex SSH-agent forwarding, Flide's faster decision models, and H100 price gaps.

Copy markdown
The checkpoints are already on Hugging Face
NCM, from Chris Ré's HazyResearch, tokenizes text into a hierarchy of discrete codes and predicts every code in a level at once, so it runs ~10× faster than the best diffusion LM on OpenWebText (32 vs 1,024 function evaluations) and ~31× on XSum summarization at matched quality. MIT-licensed checkpoints ship for XSum, WMT14 De–En, Sudoku, LM1B, TinyStories and OpenWebText — enough to stand up fast parallel summarization or a 99.5%-on-easy Sudoku solver this weekend.
Codex can hand a remote agent your SSH keys
A new opt-in flag, codex app-server proxy --forward-ssh-agent, keeps a detached or remote Codex app-server authenticated with your local SSH agent across reconnects — handy for long jobs, but it means the remote agent can use your credentials to push to private repos. It's off by default and deliberate per the PR, yet HN flagged it as a footgun, so audit the exposure before you switch it on.
Flide speeds up Jev-style decision models up to 9.5×
This Show HN tool recompiles Jev-style decision models — the tiny score-don't-generate classifiers builders drop in for routing, ranking and guardrails — for up to 9.5× lower latency. If a decision model sits in one of your hot paths, that's close to a free win.
The same H100 swings from $2.69 to $12.29 an hour
A fresh price audit clocked an identical H100 renting from $2.69/hr at the cheap end to $12.29/hr at the pricey one — a 4.5× spread for the same card. If you're training or running batch inference on rented GPUs, comparison-shopping providers is the cheapest optimization on your list this week.