I’m also sort of curious as to how much of a market research they’ve done if they’re trying to compete with Azure and AWS.
Even before the recent LLM rush took off, AI was a thing. In the city of Copenhagen there was a project to digitalise a few million case files (which is 10-100 documents per case file), and how it was done was basically with an intermediary company who knew the training and a cooperation with Microsoft. Yes, I’m dumping down the complexity of it all, but once the training period of half a year was over, Azure made a lot (and I mean a lot) of infrastructure available for not a lot of money and the process completed in a week or so. Since it had to happen and because it was a PoC the same project was also done by real humans. This was the “actual” project and every time deadline and whatnot the AI project had came from how long it would take X humans to do it. I can’t recall how many X was, but it was enough to meet the legal deadline for when these case files had to be digitised and sorted correctly.
The human project was the result, and then the AI PoC was later used as a lesson on whether it could be done this way or not. It can, it was more accurate and not more expensive.
Anyway… I’m not sure who would’ve been capable of competing with Azure. (Outside the usual suspects). Maybe a company of Hetzner could? But you would need someone who can offer you a massive amount of computing on demand, and the only companies which are going to have that are big vendors.
Maybe it’s different with LLMs because the requirement is a continuous thing rather than something you need for a short period of time?
Assuming: 50 million files, assuming 4x concurrency per CPU, taking 1 second each, would take approx 150 CPU days. Using just 10 machines it could be done in 15 days. This does not fundamentally seem like a massive project in terms of compute? If time per instance would go up factor 10x, if one would allow 30days execution, could be done by 50 machines. I think that most compute providers can do that (given some months notice)?
Even before the recent LLM rush took off, AI was a thing. In the city of Copenhagen there was a project to digitalise a few million case files (which is 10-100 documents per case file), and how it was done was basically with an intermediary company who knew the training and a cooperation with Microsoft. Yes, I’m dumping down the complexity of it all, but once the training period of half a year was over, Azure made a lot (and I mean a lot) of infrastructure available for not a lot of money and the process completed in a week or so. Since it had to happen and because it was a PoC the same project was also done by real humans. This was the “actual” project and every time deadline and whatnot the AI project had came from how long it would take X humans to do it. I can’t recall how many X was, but it was enough to meet the legal deadline for when these case files had to be digitised and sorted correctly.
The human project was the result, and then the AI PoC was later used as a lesson on whether it could be done this way or not. It can, it was more accurate and not more expensive.
Anyway… I’m not sure who would’ve been capable of competing with Azure. (Outside the usual suspects). Maybe a company of Hetzner could? But you would need someone who can offer you a massive amount of computing on demand, and the only companies which are going to have that are big vendors.
Maybe it’s different with LLMs because the requirement is a continuous thing rather than something you need for a short period of time?