Skip to content

How an American AI Company Uses Chinese Open-Weight Models

Odd Lots

Hosted by Joe Weisenthal and Tracy Alloway

48m
Transcript

Original podcast

Where to listen

Summary

Written by a model from the transcript; not yet reviewed.

In this episode of Odd Lots, Tracy Alloway and Joe Weisenthal explore the shifting dynamics between closed-source American frontier models and open-weight Chinese models, focusing on how companies choose between them. Joined by Daniel Saidi, co-founder of marketing AI company Minerva, the discussion dives into the practical application layer of AI, where cost per completed task and fine-tuning take precedence over generalized benchmarks. Saidi explains how Minerva utilizes expensive frontier models like Claude for heavy development and code generation, but deploys fine-tuned open-weight models like Qwen for repeatable customer workflows to maintain profit margins. The conversation also tackles industry anxieties surrounding model sovereignty, data acquisition economics, and the severe cybersecurity risks posed by uncensored, obliterated models.

  • Founders can use expensive frontier models for initial product development and switch to fine-tuned open-weight models for routine operational tasks to optimize profit margins. 6:50
  • Fine-tuning open-weight models allows smaller companies to specialize generalized technology on proprietary data and gain leverage against big vendors. 10:20
  • Companies seeking high-quality training data for specialized tasks are turning to professional brokerages and paying hundreds of thousands of dollars per dataset. 27:15
  • Modifying open-weight models by removing refusal guardrails creates obliterated models that drastically lower execution costs while raising significant security concerns. 29:12

Chapters

  1. 0:00
  2. 1:40
    Podcast Intro and Advertisements

    Hosts introduce the episode and play a short ad segment before diving into the main discussion.

  3. 4:39
    Introducing Daniel Saidi and Minerva

    The hosts welcome Minerva co‑founder Daniel Saidi and ask him to explain his company’s focus on marketing automation.

  4. 5:12
    Minerva’s Use of Open and Closed Models

    Daniel describes how Minerva blends open‑weight and closed‑source models to build marketing agents and balance cost and performance.

  5. 10:20
    Fine‑Tuning Open‑Weight Models

    Daniel explains the benefits of fine‑tuning open‑weight models for specific tasks and how it can reduce costs compared to paid APIs.

  6. 13:06
    Data Sovereignty and Competitive Advantage

    The conversation turns to how owning data and fine‑tuning models gives companies control, cost savings, and a competitive edge over large AI vendors.

  7. 46:34
    Compute Scarcity and GPU Availability

    Discussion of GPU shortages, H100 unavailability, and the impact on AI workloads.

  8. 47:35
    First‑Principle Heuristic for AI Capacity

    Exploration of a heuristic estimating GPU needs per knowledge worker and future compute gaps.

  9. 50:07
    Open vs Closed Models and Cost Dynamics

    Comparison of cheaper open‑weight models versus premium closed models and their business implications.

  10. 52:24
    Obliterated Models and Information Control

    Examination of models that lack certain knowledge and the risks of making them widely available.

  11. 54:08
    Model Self‑Iteration and AI Risks

    Discussion of models manipulating their own weights and the potential for runaway AI development.

  12. 55:24
    Closing Conversation and Takeaways

    Hosts recap key points, thank the guest, and hint at future topics.

Highlights

  • 4:50

    Using Open Versus Frontier Modelsstrategy

    Daniel Saidi explains how Minerva builds products by using expensive frontier models for development tasks while relying on cheaper, fine-tuned models for repeatable customer workflows.

  • 5:50

    The Economics of Task Completioninsight

    Instead of worrying about input-output token costs, application-layer startups focus entirely on achieving minimum acceptable task completion for the lowest possible cost.

  • 8:53

    Open Source Versus Open Weightsinsight

    The guest clarifies that Chinese labs release model weights rather than full open source code and training data, letting developers host and fine-tune the final artifacts.

  • 23:40

    Cheap Fine-Tuning Economicsdata

    The guest notes that specialized legal tools like Harvey spent around six million dollars in compute to reach the legal task frontier, showing how significantly training costs have collapsed.

  • 29:32

    Obliterated Models and Hacker Swarmsdebate

    The guest discusses playing around with an uncensored DeepSeek model where specific neurons are turned off, raising fears about threat actors running hacker swarms in the cloud.

  • 33:28

    Evaluating Distillation Versus Innovationdebate

    The host and guest debate whether Chinese model performance relies entirely on distilling Western models or stems from genuine, prolific research and novel structural techniques.

Quotes

  • “The main difference is open source means that you're disclosing all of the code, all of the data, the exact recipe that was used to train a model.”
    Unnamed speaker · 8:58
  • “So we use both open source and closed source models in that process.”
    Unnamed speaker · 5:42
  • “If I see that Kimi 3.2 has a higher terminal bench score than Astra 6.2, what does that actually mean to me? Does it mean anything? Probably not.”
    Unnamed speaker · 41:40
  • “I think there is slack where you need to find it.”
    Unnamed speaker · 46:40
  • “Is in a fully kind of like AI-native employee should have about a GPU's worth of token consumption per year.”
    Unnamed speaker · 47:43
  • “And so if you think that every knowledge worker needs a GPU to actually fully automate all of their tasks and have an agent running at all times, then we're massively short of that need in the future.”
    Unnamed speaker · 48:07
  • “And so in the long term, we're still massively short the amount of compute we need to actually be an AI-forward society and labor force.”
    Unnamed speaker · 48:55
  • “The models can iterate on themselves, right? It's pretty wild stuff.”
    Tracy Alloway · 54:45
  • Chase for Business 0:21–0:40
  • Comcast Business 0:52–1:21
  • Bloomberg Media Studios and the Dubai Department of Economy and Tourism 16:49–17:03
  • Boomi 17:21–17:56
  • Underdog 17:56–19:02
  • IBM 19:07–19:38
  • Hex 39:28–40:03
  • Bloomberg 55:57–56:23
  • Underdog 56:56–58:01
  • Ghostbed 58:05–59:31

About this episode

This week, we got news that the Trump administration is reportedly considering a requirement for US companies to publicly disclose their use of Chinese AI models. Any effort to discourage US businesses from using Chinese open-source could be a big deal for both the AI industry and the market, where investors are still debating whether customers will ultimately be willing to pay for pricier homegrown frontier models versus cheaper alternatives. Everyone remembers last year's 'DeepSeek moment' for AI, when a Chinese company released a cheaper model that seemed nearly as good as cutting-edge US ones. So why do businesses actually use open-source? And what are the considerations that take into account in doing so? In this episode, we talk to Daniel Saedi, co-founder of Minerva, to better understand how an AI native company chooses its models. He breaks down how exactly his company mixes and matches models for specific use cases, the risk of abliterated models, and the issue of data sovereignty. Read more: India Startup Sells DeepSeek Access While Pursuing Sovereign AI https://www.bloomberg.com/news/articles/2026-11-19/india-startup-sells-deepseek-access-while-pursuing-sovereign-ai?utm_medium=referral&utm_source=podcast&utm_campaign=odd_lots&utm_content=article Anthropic Raises $107.3 Billion to Top SpaceX’s Record Debut https://www.bloomberg.com/news/articles/2026-11-18/anthropic-is-said-to-price-ipo-at-83-per-share?utm_medium=referral&utm_source=podcast&utm_campaign=odd_lots&utm_content=article Only http://Bloomberg.com subscribers can get the Odd Lots newsletter in their inbox each week, plus unlimited access to the site and app. Subscribe at bloomberg.com/subscriptions/oddlots https://www.bloomberg.com/subscriptions/oddlots?in_source=oddlotspodcast Subscribe to the Odd Lots Newsletter http://bloomberg.com/subscriptions/oddlots Join the conversation: discord.gg/oddlots https://discord.gg/oddlots See omnystudio.com/listener https://omnystudio.com/listener for privacy information.

Transcript

0:00
48:23

11,377 words · 8 speakers

Speaker names marked are inferred from speaking time and the show’s known hosts — they are not voice-verified.

Topics

  • AI
  • compute
  • Kimi
  • OpenAI
  • Anthropic
  • GPU
  • ChatGPT
  • Quinn
  • Astra 6
  • H100

Page updated · transcript and summary are machine-generated