[lumost]

the counter point is that building or selecting the specialized model may cost as much as the lifetime inference costs of the task with the specialized model.

If I need to pay someone 300k to make the model and infrastructure... then I would need to process many documents to recoup my OCR costs compared to asking claude code nicely.

Perhaps the model zoo is becoming good enough that the cost to find a specialized model is not so high?

[berkes]

> If I need to pay someone 300k to make the model and infrastructure

I was arguing for the existing AI-companies that already make and offer niche models. Like Mistral. But AFAICS, all AI companies have and offer such models.

So all you need to do, is use the existing models. And, yes, select it.

Which, ironically, I would highly value as a niche model myself. I spend way too much time following the breakneck race of the various companies just to pick the right models for my tasks at hand. "Just pick the latest" often yields worse results, or is magnitudes more expensive, or significantly slower, or all of it. "Just pick the most popular" can prove expensive, inefficient for some task etc. This investment, also ironically, has proven something of a "moat" for me. I know very well what Mistral and Anthropic offer. So I won't even bother with OpenAI, Google, X, Tencent etc etc etc models. I just don't have the time to keep researching the latest offers for their pros and cons.

A model that acts as "decision maker" and as proxy, as a conductor, that directs and transforms my questions and sends them off to the right model, right tools, right MCP etc, would be very welcome for me. So that I can just pick that one, and have the highly dynamic world of LLMs and other models shift like undercurrent beneath the surface of this One Model To Rule Them all.

[notahacker]

Sure, but things like "OCR tool" or "agentic calendar assistant" or "agentic interface to CAD" don't feel like things I'd need pay 300k to build out custom infrastructure for, they feel like something lots of people/companies would rather pay an appropriate-sized subscription to a model optimized for that use case rather than burn ludicrous amounts on generalist models optimized for burning as many tokens as possible reasoning about how to update a calendar from first principles.

I mean, I also I think the "agentic calendar assistant" is more likely to come out of the Valley than Europe[1] and that the big AI startups will buy some of the niche AI startups gaining traction in strategic markets, but there are lots of niches that notionally benefit from transformer-based models

[1]probably a calendar assistant with a ludicrously inflated valuation based mostly on the assumption that Google/Meta/Anthropic will buy them to shut them down three years later...