Show and guest
This episode of Bloomberg Podcasts’ Odd Lots is hosted by Joe Weisenthal and Tracy Alloway, with Minerva co-founder Daniel Saedi (a former Bridgewater employee) as the guest. Original English title: How an American AI Company Uses Chinese Open-Weight Models | Odd Lots.
Key takeaways
- Open-weight models let users download clean weights and run them on their own computers or chips. For application-layer companies, the real value lies in fine-tuning: using their own data to fine-tune open weights is currently the main source of returns. 9:08
- Fine-tuning Chinese open weights such as Kimi can produce better results at lower cost, while also capturing an additional layer of margin. Other benefits include data sovereignty, technical flexibility, and avoiding dependence on large suppliers that may raise prices. 12:32
- He personally tested the latest uncensored version of DeepSeek: someone had turned off the neurons in the model that say “no”—what is known as “model erasure.” This is the biggest risk of open frontier models circulating outside their intended channels. The model costs about $0.30 per million tokens, around 10 times less than Sonnet, yet its capabilities are close to those of the mythic version from February. 25:20
- The benefits of distillation can be measured: Reflection AI, one of the top U.S. open-weight model companies, has just released its flagship model, which trails Chinese open-weight models by about 3–6 months. It is backed by Nvidia and was trained using 12 thousand–14 thousand GPUs; Nvidia is reportedly acquiring the company. 29:54
Background and notes
This episode of Odd Lots discusses reports that the Trump administration is considering requiring U.S. companies to publicly disclose their use of Chinese AI models. It also considers whether, after the “DeepSeek moment,” customers are willing to pay a premium for domestic frontier models. Subtitle coverage: the video has 978 timestamped subtitle segments, covering 0:00–48:59 (total runtime 48:59), with full-video coverage. The views and timestamps in this article are based on the original subtitles; each timestamp marks the start of the corresponding spoken segment. Publication time: 2026-11-20 17:48 Hong Kong time, converted from the relative time displayed by YouTube when captured, with hour-level precision.