I have a sneaking suspicion that someone at Google may be making the same bet, looking at the faster and faster Flash models which provide acceptable results to a lot of people (outside of coding).
For Anthropic. But not for the AI industry at large.
Cheaper, more powerful AI will continue to expand the bubble. Projects will get more ambitious. Everyone will build out their own custom little software. Code diversity expands and requires even more AI.
The bubble is _not_ on models becoming more intelligent and solving arc-agi-999.
They are already good enough at what they mechanically are.
You have to use the right harness, right verifiers (automatic where possible, human where not), etc much much more specific than a generic one like claude code or codex, and it will also be able to work within constraints and be the "proposer" of an imaginary optimisation problem and an excellent one at that. But you have to frame the task at hand in that manner or maybe even reorganise the task you do itself so it is more amenable to being framed that way. If you use it this way, it is _already_ massively economically useful. But it will take many years for it to actually be usable in that way, since you need DC capacity to come up first which is few years away and also well, massive organisations that have to integrate these will usually take many years to do so.
It is also useful albeit less so in cases like general SWE, where you still need a human in a loop for non-verifiable requirements, and also in other general usecases where information retrieval is too intractable and you need to carefully use LLMs as a component of the overall system.
I am not saying Fable or whatever the biggest models are are useless - they will certainly be useful for tasks at the frontier of the day - which is today complex exploits and open math problems, and well, tomorrow it could be something in biotech. But this is not what the entire bet is on at all - just automating day to day drudge at the tens of thousands of massive companies and governments we all know and love is more than enough. With the right training data (which _also_ is a bottleneck and takes time) you could even automate certain processes entirely. Sure, if we get a crazy medical innovation and end up saving trillions in healthcare great, but that's just a bonus.
None of this is to say that I think there is zero sketchy financial engineering going on
Right? If we’ve already reached ‘good enough’ then there’s rough waters ahead.
I have a sneaking suspicion that someone at Google may be making the same bet, looking at the faster and faster Flash models which provide acceptable results to a lot of people (outside of coding).
For Anthropic. But not for the AI industry at large.
Cheaper, more powerful AI will continue to expand the bubble. Projects will get more ambitious. Everyone will build out their own custom little software. Code diversity expands and requires even more AI.
The bubble is _not_ on models becoming more intelligent and solving arc-agi-999.
They are already good enough at what they mechanically are.
You have to use the right harness, right verifiers (automatic where possible, human where not), etc much much more specific than a generic one like claude code or codex, and it will also be able to work within constraints and be the "proposer" of an imaginary optimisation problem and an excellent one at that. But you have to frame the task at hand in that manner or maybe even reorganise the task you do itself so it is more amenable to being framed that way. If you use it this way, it is _already_ massively economically useful. But it will take many years for it to actually be usable in that way, since you need DC capacity to come up first which is few years away and also well, massive organisations that have to integrate these will usually take many years to do so.
It is also useful albeit less so in cases like general SWE, where you still need a human in a loop for non-verifiable requirements, and also in other general usecases where information retrieval is too intractable and you need to carefully use LLMs as a component of the overall system.
I am not saying Fable or whatever the biggest models are are useless - they will certainly be useful for tasks at the frontier of the day - which is today complex exploits and open math problems, and well, tomorrow it could be something in biotech. But this is not what the entire bet is on at all - just automating day to day drudge at the tens of thousands of massive companies and governments we all know and love is more than enough. With the right training data (which _also_ is a bottleneck and takes time) you could even automate certain processes entirely. Sure, if we get a crazy medical innovation and end up saving trillions in healthcare great, but that's just a bonus.
None of this is to say that I think there is zero sketchy financial engineering going on