[netcan]

Anthropic generally seem more into living within market discipline and market signals of some sort. Products with margins, even if it's sort of irrelevant considering R&D costs and capital inflow.

That said, there's nothing like the real thing.

The risk is something like the railroad bubble and the dotcom. Over-investement, circular revenue and a timeline that doesn't work.

Or, maybe it'll work out.

[c3fxx]

The whole premise is based on the fact that over-investing in GPUs and models are a good thing here as it yields more 'intelligence'.

This as it turned out was not true for rail roads - more and more rail roads isnt a good thing.

The real dilemma facing the model producers is that all this money invested for a general model, targeting general intelligence, is a disaster and essentially the investment into existing assets is a write off. Then on top of that if this is true, youve got data centres full of compute that aren't being used up.

[andai]

The weird position they find themselves in now is that they have to keep making it smarter... but they already made it too smart (Mythos). I'm not sure how that's going to work out exactly.

They find an arbitrary intelligence cutoff point between Opus and Mythos, label it "acceptable risk", and then the labs coordinate to gradually nudge that line forward and hope the internet doesn't break?

[notnullorvoid]

> but they already made it too smart (Mythos).

It's largely a marketing tactic. It will be released, and it won't be long before other models show similar capabilities.

If they wanted they could add guardrails. The scales required to brute force search for vulnerabilities like they did would be very identifiable.

[c3fxx]

Scam Altman already pulled this trick numerous times.

Whats wrong with people? Is it really that hard to see the truth?

[ghywertelling]

I think we will see unbundling of large model into submodels: modular, smaller and efficient, only include what you need eg a CUA model, a reasoning model, a legal model, a writing model, a coding model (this could get subdivided into different languages). That way you only update that submodel which needs retraining.

[throwaway173738]

Maybe they’ll figure out how to make an agent train an agent.

[andai]

The labs started doing that in late 2024, they all published research on it.

Curiously, mid 2025, they all simultaneously implemented increasingly bizarre restrictions on "self replication". I don't think there was anything public but it sure sounds like something spooked them. (Or maybe just taking sensible precautions, given the direction of the whole endeavour.)

At any rate, I recently asked Opus about "Did PKD know about living information systems?" and the safety filter ended the conversation. It started answering me, and then it's response was deleted and a red warning box popped up.

But notably, I was given the option to continue the chat with a dumber model (presumably one less capable of producing whatever it thinks I meant by that phrase).

Also, I told GPT-5 about my self-modifying Python AI programmer, and it became extremely uncomfortable. I told it an older version of itself had designed and built it (GPT-4 in 2023), and it didn't like that at all! So something's definitely changed in the safety training there.

[tempodox]

And if they don’t, it won’t be for lack of trying, I promise you. Just like the circular financing, nothing makes more work for “AI” than “AI”.