Interesting that this got posted now: the project is receiving increasingly more skepticism lately in the Dutch tech scene [0], and I think that’s fully justified.
What is the exact skepticism? The only thing I could get from that was from some "tech entrepreneur":
> GPT-NL was never built to compete with Claude or ChatGPT. It was trained exclusively on licensed data, and is intended more for governments and companies where privacy and compliance matter more than raw performance.”
That's it? That it didn't aim to compete with SOTA models? Maybe this is something you have to start with something, then ramp up, rather do what only a select few labs been able to do, start with really big models. Especially if you're resource constrained, which since this is a government project, I really hope for the sake of the tax payers it was.
TNO is something like semi-DARPA. It gets a lot of stuff tax free and a lot of gov funding, but a lot of their budget is from getting businesses to hire their R&D teams.
They do really good R&D on a lot of stuff. This is just their attempt at public credibility/internal skill building to enter the LLM business.
Doubt its going to be successful, but they "waste" a lot more money on other things that you never heard of. Its not fraud, its just R&D dressed up a little too much too early.
Common approach I've seen is having workflows with paid/larger/hosted models for some workflow where you don't quite know exactly how it'll be when you first put it together, then with time you've locked down how things more or less work yet you still need free-form text parsing of some kind, so you end up replacing the bigger models with carefully post-trained small models.
Besides that, there is a ton of use cases for smaller models for a bunch of different things. We'll be unlikely to be able to run LLMs (actually Large) on smartphones for a while, while the smaller LLMs seem to run already on-device in experiments.
> GPT-NL was never built to compete with Claude or ChatGPT. It was trained exclusively on licensed data, and is intended more for governments and companies where privacy and compliance matter more than raw performance.”
That's it? That it didn't aim to compete with SOTA models? Maybe this is something you have to start with something, then ramp up, rather do what only a select few labs been able to do, start with really big models. Especially if you're resource constrained, which since this is a government project, I really hope for the sake of the tax payers it was.