This started as a couple of comments under George Batchelor's musings on AI on Linkedin, but I thought it was worth setting out in a standalone post. So here are my current tuppence on the value(s) of AI models:
(The future is inherently uncertain. All views my own. Etc.)
- There is a real chance that many investments in AI models will turn out to be NPV-negative for investors.
- This is not necessarily a bad thing socially. If none of the frontier labs achieve the dominant market position that, say, Google achieved in search, the result could be cheap, widely available AI capabilities as profits are competed away and consumer surplus is created.
- Many leading Chinese models, as well as Meta's Llama and Mistral's models, are open-weights. Everyone with sufficient hardware can run them locally. Everyone can, in principle, tailor them according to their needs. For example, last year, Perplexity released R1-1776, an adapted version of DeepSeek-R1 that answers more freely on topics the original model treats more cautiously, illustrating how open-weight models can be modified for different markets and norms.
- So, in theory, the amount that frontier labs such as Anthropic can charge is constrained by (i) how much better they are than the open-weight models; (ii) the costs of operating them locally; and/or (iii) how much the open-weight labs charge for their interfaces.
- Open-weight models are often estimated to be 4-8 months behind the frontier closed-weight models. How much are people willing to pay for a model that's just a few months ahead?
- This is not an either/or choice either. A couple of months ago, a developer told the FT that he spends $50 a day using Moonshot’s Kimi model for roughly 80% of his work, reserving Anthropic’s Claude for more complex tasks because using Claude alone would cost him $900 a day.
- And the open-weight competition is real globally. It was reported that US technology companies have been adopting Chinese models due to cost concerns. I can see why: if these models can do most tasks sufficiently, if not equally, well, but for less than 10% of Claude's token price, they are viable alternatives in many cases.
- If many developers end up using cheap open-weight models for most tasks and reserving Claude / GPT only for the hardest ones, that will have an impact on OpenAI and Anthropic's financial positions and valuations.
- There have been a lot of discussions about AI safety recently. Any analysis that does not address China and/or open-weight models that can be run locally is at best incomplete.
- The following statements are not mutually exclusive: (i) competition is good because it increases total welfare; and (ii) much of that welfare may show up as consumer surplus instead of profits or anything measured in GDP.
(proofread by GPT-5.5. All errors are human and mine.)
 | |
| Generated by ChatGPT-5.5, with the lyrics of Eason Chan's 人工智能 ("Artificial Intelligence", 2001) as the prompt. |
