Make LLMs Easy to Train

Training language models still means broken SDKs, flaky GPU instances and wrangling terabytes of data. The ask is products that make model training easy.

Gabriel Birnbaum and his co-founder have trained diffusion and language models at Can of Soup for three years, and he says the tooling has barely improved: broken SDKs, GPU instances that turn out to be faulty only after half an hour of setup, bugs in open-source tools, and the work of managing terabytes of data.

Wanted: APIs that abstract training, databases for very large datasets, and development environments built for ML research. As post-training and model specialization matter more, these products could become foundational.

Read the full request on YC →