Data Machina #262

Mistral AI and NVIDIA teamed up to ship NeMo, a 12-billion-parameter language model that punches well above its weight class, at least on paper. It handles multiple languages, stretches to a 128k-token context window, and comes with both base and instruct versions sitting on Hugging Face under the permissive Apache 2.0 license. That's the kind of release that used to be a big deal on its own. Right now, though, some researchers are picking apart Mistral's benchmark comparisons against Llama 3 8B, arguing the numbers look a little too flattering. Open weights are great. Open scrutiny is what keeps everyone honest.

On the research side, two papers stood out this week for solving unglamorous but expensive problems. Stanford's TextGrad borrows the idea of backpropagation from deep learning and applies it to plain text: instead of numerical gradients, an LLM generates written feedback that gets fed back through a compound AI pipeline to improve each component. Early results show real gains in zero-shot performance across a range of tasks, which is notable because most prompt-engineering tricks don't generalize that well.

Tencent, meanwhile, tackled a much blunter problem: token training is brutally expensive once you're processing tens of billions of tokens. Their patch-level training approach compresses several tokens into a single patch and trains the model to predict patches instead of individual tokens. The payoff is roughly half the computational cost, which for anyone running training clusters is not a rounding error.

There's also a free, community-built computer vision course from Hugging Face aimed at bringing newcomers up to speed on the field, and Stanford's STORM system, which tries to automate long-form, Wikipedia-style writing by simulating multiple perspectives interviewing a topic expert before drafting an outline. Different corners of the field, same underlying theme: make models cheaper to train, easier to improve, and more useful for producing things humans actually read.