I heard from Ed Zitron that you can't really turn off "training" on these models even after they have been released because they need continuous re-training to keep up with the changes in the world and the language etc. Is this correct? if it is, can we even say that research can be turned off entirely?

it's completely false

I do at least think its certainly the case they cannot stop R&D: no moat without being ahead of open-source right?

So model drift is not really a thing?

You don't need to retrain model from scratch to add small of additional knowledge like recent events.

Thanks

The moment they‘d stop development would put marketing and sales in a very dire position, regardless of how good gpt 5.5 is.

Anthropic claimed themselves that AI progress will slow down effectively calling fable the pinicle of what we can do today which is true, they've trained the largest possible model with the most advanced system from nvidia. GPT 5.5 would likely last a year or two because I don't see chinese labs investing that many billions in compute since at the end of the day it's a compute x intelligence graph and the only way to catch up is to quite litereally spend more money, that's realistically where the moat is. Of course distilling techniques have been proved to be very powerful, but I am still not convinced they're able to replicate the nuanced behavior these models exhibit. GLM 5.2 will be interesting to see as it is the most promising model that might blow up the entire argument of spending billions in training.

That said this is obviously not a strategy, but rather an observation that this is not a flawed, impossible concept.

Yes, but I don’t see either of those scenarios giving them the growth needed to justify a trillion dollar valuation. Especially with the current competition (a competitor releases a much better model, and they have to respond or see their revenue drop).

So I don’t see a company in immediate danger of collapsing, but I also don’t see a great investment at that valuation.

but why is that a feasible hypothetical?

Just 10x your revenue without increasing R&D cost? where's this 1 magic trick and why can't every company just use it?

R&D cost is static, converting all subscription customers to api customers would yield 10x boost in revenue immediately so the demand is there, of course we will probably see demand expand more than that the question is if it's 10x (break even) or 20x (justifying trillion dollar evaluation.

> R&D cost is static

since when? do you think their R&D in 2026 is the same as in 2022?

22->26 was scaling up, right now they're at their peak R&D spending from today onwards for the next forseeable future until nvidia starts rolling out their new room-sized systems this will be the upper limit of what can be trained without diminishing returns to where you're just burning money for less than single digit gains. R&D at the end of the day is what they can afford to spend. R&D grows because you have more money to spend, not because the R&D costs are increasing, therefore assuming current R&D costs are enough to keep up with competition (which they very likely are) R&D will remain static or (ideally) shrink.

95% or more of current R&D costs are compute, same as cost of revenue.

I doubt it lets see who's right. 2028 r&d is it higher or lower than 2026.

> converting all subscription customers to api customers would yield 10x boost in revenue immediately so the demand is there

If OpenAI started charging all subscription users API pricing they would lose the majority of users.

If they increase the price demand will plummet. They have no moat and there are competitors.

If they stop R&D right now they'd have to fire most of their staff to realize the savings and sell off their datacenter capacity.

They would show a profit for 1 year and then become completely irrelevant.

they aren't suggesting it as an actual strategy, merely an indicator that the fundamentals seem realistic