Hard to see how can Mistral compete with Meta, they have order of magnitude less compute, their models are only slightly better (at least on the benchmarks) with less permissive licenses?
In general I feel like all model providers eventually become infrastructure providers. If the difference between models is very small, it will be about who can serve it reliably, with the most features, with the most security, at the lowest price.
I'm the head of R&D at Rev.ai and this is exactly what we've seen in ASR. We started at $1.20/hr, and our new models are $0.10/hr in < 2 years. We have done human transcription for ~15 years and the revenue from ASR is 3 orders of magnitude less ($90/hr vs $0.10/hr) and it will likely go lower. However, our volumes are many orders of magnitude higher now for serving ASR, so it's about even or growth in most cases still.
I think for Mistral to compete with Meta they need a better API. The on-prem/self-hosted people will always choose the best models for themselves and you won't be able to monetize them in a FOSS world anyways, so you just need the better platform. Right now, Meta isn't providing a top-tier platform, but that may eventually change.
Yeah, the license thing is definitely a problem. It's hard to get excited about an academic research license for a 3B or 8B model when the Llama 3.1 and 3.2 models are SO good, and are licensed for commercial usage.
to be clear- these ministal model are also licensed for commercial use, but not freely licensed for commercial use. and meta also has restrictions on commercial use (have to put “Built with Meta Llama 3” and need to pay meta if you exceed 700 million monthly users)
They can't since Meta can spend billions on models that they give away and never need to get a direct ROI on it. But don't expect Meta's largess to persist much beyond wiping out the competition. Then their models will probably start to look about as open as Android does today. (either through licensing restrictions or more 'advanced' capabilities being paywalled and/or API-only)
> But don't expect Meta's largess to persist much beyond wiping out the competition
I don't quite follow your argument - what exactly is Meta competing for? It doesn't sell access to a hosted models and shows no interest of being involved in the cloud business. My guess is Meta is driven by enabling wider adoption of AI, and their bet is more (AI-generated) content is good for its existing content-hosting-and-ad-selling business, and good for it's aspirational Metaverse business too, should it pan out.
I'm arguing that Meta isn't in this for altruistic reasons. In the short term, they're doing this so Apple/Google can't do to them with AI tech what they've done to them with mobile/browsers. (i.e. Meta doesn't want them owning the stack, and therefore controlling and dictating, who can do what with it) In the longer term: Meta doesn't sell access... yet. Meta shows no interest... yet. You could have said the same thing about Apple and Google 15+ years ago about a great many things. This has all happened before and this will all happen again.
Hard to imagine anyone competing with AWS/GCP/Azure for slices of GPUs/TPU. AFAIK, most major models are available a la carte via API on these providers (with a few exclusives). I can’t imagine how anyone can compete the big clouds on serving an API, and I can’t imagine them staying “non compliant” for long.
Maybe, but when selling a SAAS here, big clients will always ask what cloud provider you use. Using an European one is always a plus, if it isn't simply required.
