Harvey spent about $6 million in compute to bring a fine-tuned Chinese open-weight model to the frontier of legal work, Daniel Saedi, co-founder of Minerva, told Odd Lots. He said fine-tuning has become far cheaper than building a model from scratch.
Most of the debate over cheap Chinese AI models asks whether they will undercut Anthropic and OpenAI. Daniel's startup pays for both kinds. It builds its product with frontier closed models and runs its high-volume work on cheap Chinese models it has fine-tuned on its own data.
"I actually think that this is the primary benefit to the application layer right now — the ability to fine tune on top of those open weights." — Daniel Saedi
Daniel spent four and a half years at Bridgewater Associates before co-founding Minerva, which builds AI agents that run marketing campaigns from planning through execution. Because it mixes open and closed models in production, he sees the price of each one on its own bills.
The full interview is covered here so you can skip it. 48 minutes of audio, 10 minutes of reading.
Here are the 10 insights that matter.
Odd Lots · Nov 20, 2026 · 48m · 38m saved
With Daniel Saedi, Co-Founder of Minerva
Key Takeaways
Minerva uses frontier models to build its product and cheap open-weight models to run it
Chinese labs release weights, not the code, data or recipe, so "open weight" is not open source
Fine-tuning has become cheap, with some teams training usable models for as little as $90
Daniel ranks the benefits of fine-tuning as cost first, then control of a company's own data
The biggest security risk of open weights is a model with its refusals switched off, not data flowing to China
Distillation alone does not explain how far Chinese open models have come
Daniel thinks the US is still short of the GPUs that full AI adoption would need
1. The DeepSeek Question
Joe Weisenthal and Tracy Alloway, who host Odd Lots for Bloomberg, opened on the worry the DeepSeek release created: that near-frontier Chinese models, given away as open weights, could take pricing power from the closed US labs.
Joe framed cheap Chinese models as good news for chip sellers and a problem for the closed US labs
"This was the core anxiety of the deep seek moment... if you're selling chips, then that's great... but for the big American closed source companies, specifically Anthropic and OpenAI, there are some big questions right there." — Joe Weisenthal
Joe and Tracy listed the open questions. Are Chinese labs mostly distilling Western models, possibly against those models' terms of service, or doing original work? And do benchmark scores overstate how the models perform on real jobs? They brought on Daniel for the view of a company that uses both kinds of model rather than one that builds them.



