34 chunks · 24 episodes
The model providers are still training the models to know how to do arithmetic. Only if you're Seeing Like A Frontier Lab would you do it that way. That's insanely wast...
The incentives of the model providers are not fully aligned with users. Over time, those incentives will become increasingly misaligned. Ideally you want your harness to be made by someo...
Seeing like a model provider: look at every problem like an eval problem. The answer to every problem becomes "simply get a better model," and the tactic to do that is "simply de...
Model providers keep on deprecating old models. Even if the old model was good enough for your use case, you might not have access to it. They have a limited amount...
...alternatives are largely good enough for most workloads. This is one reason why model providers are desperately trying to move to harnesses that store state.
...s, which requires leaving money on the table for the fat tails. Two options for model providers: 1) Move to a sales-gated API for all uses. This would allow detecting the fat-tail use and bucketing them into the higher margin buckets based on u...
Whenever model providers complain about others distilling their models, I think about the Project Panama images. A warehouse full of books deliberately destroyed by the inge...
... minutes to closer to 5 minutes to get more efficiency. If you want to cost the model provider a ton, send a single question about 5 minutes after the last one finished, to stay permanently in the cache.
...elected for. If vertical integration is selected for, it's conceivable that the model providers stop giving access to the models for the frontier models. Mythos appears to be going that way. Conveniently, they can tell a story of national secur...
A situation good for society is "all of the model providers have to compete and none of them win" for LLMs. Great for everyone but the LLM model providers, who are in a never-ending red ocean battle. But the ...
The model providers seem to be in a meta-stable equilibrium. None of them have any differential pricing power, since the models are practically commodity. But they do a...
The LLM model providers are like electricity providers back when electricity was new. Competing to get better quality for cheaper. Innovating on new techniques to do so. Bu...
...'t. They are bound by their fiduciary duty to continue pushing, since the other model providers are too. One apparently told him that they were hoping for a Chernobyl style disaster that would get governments to step in and stop the competition...
I continue to think the best business parallel for LLM model providers is cell phone networks. Extremely capital intensive to build out, but then much lower marginal cost to operate. Though inference has much more margi...
Tying models to UX from that model provider is dangerous. If the models are tied to the UX from the vertical integration, users get stuck to a single model. That requires that one model to be "...
...y way users interact with models is not via a UX owned by one of the foundation model providers.
...dels of generally competitive quality and lightly differentiated abilities. The model provider should not also own the context layer. That's like the cell carrier trying to dictate you can only use their super-app on your phone.
...e race between a non-trivial number of powerful LLMs, which prevents any of the model providers from trying to flex their muscle too much.
...t consumers have a subscription to that then allows them to use any of the main model providers with their context and not get stuck with any of them. But it's unclear which of those offerings will be the schelling point that starts getting com...
...heir API before ChatGPT got big. Because they set that precedent, the other top model providers also added a public API. Now, if any one of the providers got rid of their API, their competitors would push forward and scoop up the market share. ...