Right, the Costco problem. A small boutique eg wine store might be able to do better for picking a very specific wine for a specific occasion, but Costco is just so much bigger that they can make it up in Volume and buy cases and cases of everything with a lower markup, so it ends up being cheaper to shop at Costco, no matter how much you want to support the local wine boutique.

In Norway there is a state-owned monopoly on selling wine and liquor (anything above 4.75% ABV). They have 350+ physical shops, a large online shop and around $2bn annual revenue. This makes them one of the largest purchasers of wine and spirits in Europe, and they can get some very good deals.

So even though you have high taxes and a restrictive alcohol policy, the end result is shops that have high customer satisfaction because they have very competent staff, excellent selection and a surprisingly good price for quality products.

The downsides are the limited opening hours and the absence of cheap low-quality wine - the tax disproportionally impacts the low quality stuff, almost nobody will buy shitty wine at $7 per bottle when the decent stuff costs $10, so the shitty wine just doesn't get imported. But for most of the population these are minor drawbacks.

The advantage of small purpose-specific models is that they might be much more robust i.e., unlikely to generate wrong sequences for your particular domain. That is at least my experience working on this topic during 2025. And, obviously, smaller models mean you may deploy them on cheaper hardware, latency is reduced, energy consumption is lower, etc. In some domains like robotics, these two advantages might be very compelling, but it's obviously early to draw any long-term conclusions.

I second this. Smaller models indeed may be much better positioned for fine-tuning for the very reason you point out - less noise to begin with.

> If/when frontier model development speed slows down

You do not believe that this has already started? It seems to me that we’re well into a massive slowdown

Not the OP but I use AI all day every day and have noticed substantial improvements in the models over the past ~6 months. GPT-5 was a huge leap (contrary to reporting) and so was Sonnet 4.5.

GPT5 was by no means a huge leap. I’d be willing to believe that you prefer it, or that you found it an improvement, despite both of those being wildly contrary to my experience (and most of the rhetoric online). But objectively speaking it was a small improvement, even going by OpenAI’s marketing claims.

In practice, I upgraded everything to GPT-5 and the performance was so terrible I had to rollback the update.

> GPT-5 was a huge leap (contrary to reporting) and

Depends on what you compare it to. For us who were using o3/o1 Pro Mode before GPT-5, the new model isn't that huge of a leap, compared to whatever was before Pro Mode existed.

It's hard for me to say. I don't think you know you're on the S-curve until after the fact.

On the one hand, most models are "good enough" for chatgpt-like usage, and there it's hard to see/feel generation-to-generation improvements. On the other hand, if you look at instruction following, dealing with long context windows, >200 tool call interactions while staying on track, there's still plenty of improvements to be had. So, hard to say where we are.