Author’s Note: I’ll be traveling the next two weeks, probably dropping to 1x / week while I’m out!
The Facts
Here are a few loosely related pieces of news from the last few weeks:
Earlier this month, Mistral AI announced their €105M initial funding round, and their strategy memo (an interesting read if you have 5 minutes) details their plan to build models that are competitive with GPT-4. The team behind Mistral (according to a few sources) is an incredible group of engineers and researchers.
Yesterday, Inflection AI released results claiming that their Inflection-1 model (which will soon be available as a conversational API) outperforms GPT-3.5 on a number of benchmarks. As we’ve detailed here a few times, there are good reasons to be suspicious of benchmark results, but Inflection certainly has the chops to train a competitive model.
Why it matters
Despite all of the recent model releases (Anthropic 100k, the dozens of recent open-source model releases), one fact remains true: nearly every LLM developer is still relying on the OpenAI endpoints to power their applications. Despite all of the investment and effort, other models have not meaningfully cracked into real workloads.
It’s worth checking in — how do we feel about the chances for other providers to challenge OpenAI? A few thoughts:
Cost
Training GPT-4 quality models remains incredibly expensive. Mistral estimated needing an additional $200M to train a competitive model in 2024 — I think that is optimistic and factors in increased GPU availability as H100s come online.
Anthropic and Cohere announced $450M and $270M fundraises this year — probably enough to feasibly train a large model.
Data
Quality data is one of the biggest bottlenecks to training GPT-4 quality models, and OpenAI has already acquired that data. They’re also expanding that data advantage daily via ChatGPT.
Distribution
OpenAI certainly has an early distribution advantage, but I don’t think they have a great go-to-market motion yet. They’re juggling research, consumer products, and developer products, and I think there is room for competitors to win by focusing on one of these categories.
Model Size
All of this talk is about GPT-4 quality models, but fine-tuned smaller models can outperform larger models on specific tasks. This isn’t yet a large market, but there is room to win here with specialized data and by understanding the needs of a niche market
My thoughts
I think there are ~4 companies adequately funded to train GPT-4 quality models right now, and it seems likely that at least one of them will succeed in the next 12-18 months. OpenAI does not seem to be building GPT-5 yet, so there should be room for a competitor to emerge. Considering how much value OpenAI has created with GPT-4, seems like a worthwhile investment!