upvote

points

by jeremyjh1 days ago |
comments
upvoteby user439281 days ago|
next
[-]
From the model card:

  Model                AECI   95% CI (global)
  Claude Opus 5.5    174.78   170.33-180.29
  Claude Mythos 5.2  177.12   172.42-182.94
  Internal model     178.78   173.75-185.18
Considering the IPO is planned in a week, I doubt they are holding much back.
reply
upvoteby jeremyjh1 days ago|
parent|
[-]
According to their S-1:

In H1 2026, cost of revenue is $2.3B. That is primarily inference compute, but also includes cost of training models (but not research).

In the same period, R&D was $9.4B. That is primarily spent on compute. That is what they are "holding back" already, in full public view.

This is not a company that is optimizing for OpEx or end user revenue. The only thing relevant to their plans is achieving AGI/ASI first, and controlling enough compute to leverage it to earn the majority of the current economy.

reply
upvoteby sfink1 days ago|
parent|
[-]
"Achieving AGI" strikes me as a mostly meaningless phrase. I'm skeptical that it is a boolean event. They're already smarter in some ways, their impact is large and gradually rising, the set of specific tasks that they aren't capable of is gradually shrinking. If AGI means anything useful, it seems like we'll only see it in hindsight and have to argue over when "it" happened.

Likewise the singularity. While recursive self-improvement is theoretically trivial (just tweak a random subset of weights like a monkey with a typewriter, eventually you'll get something smarter), the resulting exponential improvement is still very hypothetical and feels unlikely. Unless you're ok with an exponent of 1.000000000000000001.

reply
upvoteby jeremyjh1 days ago|
parent|
next
[-]
These are not my stated ambitions, they are the ambitions that Anthropic and Dario have said they are pursuing. Whether or not they are realistic is not relevant to this thread. The credibility of the plans are relevant to the S-1 and anyone who purchases stock from them, but for the purposes of this thread, it is only important that Anthropic believes it has credible chances of achieving this goal, and in particular that it has a competitive advantage, could do so first, and that being first could have a very large upside.
reply
upvoteby sfink1 days ago|
parent|
[-]
Sorry, I didn't mean to judge your post, I was just going off on a tangent inspired by your post. Mostly unrelated to either the S-1 or Fable 5.2.
reply
upvoteby zeroonetwothree1 days ago|
parent|
prev|
next
[-]
"smarter in some ways" != "general intelligence"

I agree we will only see it in hindsight, but I think that's always the case with any technology. It seems worse until it's suddenly better. I don't think it matters to argue the exact moment though. For example when did cars become clearly better than horses? No one cares about the exact date.

reply
upvoteby charcircuit1 days ago|
parent|
[-]
What about the intelligence is not general? What subject is it missing?
reply
upvoteby Jensson1 days ago|
parent|
next
[-]
The stuff that humans are still smarter at, otherwise it wouldn't be "smarter in some ways".
reply
upvoteby econ1 days ago|
parent|
prev|
[-]
We are slowly approaching a real definition of intelligence by chanting that llms aren't.
reply
upvoteby emp173441 days ago|
parent|
prev|
[-]
All tools are better than humans in specific, narrow ways. That’s how tools work.
reply
upvoteby vlyan1 days ago|
prev|
[-]
because they hate money? google is the only frontier lab that isn't hundreds of billions in the red.
reply
upvoteby jeremyjh1 days ago|
parent|
next
[-]
Because they do months of testing. Also, if they have models that are really good at AI research, I’d not expect full versions to ever be released, precisely because of their love of money.
reply
upvoteby thomasmurph1 days ago|
parent|
prev|
next
[-]
Inference on the research models they have is likely more expensive than the released ones.
reply
upvoteby 93po1 days ago|
parent|
prev|
[-]
My assumption is that they're overly expensive to run at scale and/or don't have the same safeguards that public models have and possibly get better results because of that. If a model is 100x more expensive to run but 4x better results, of course you'd use that internally to make your product even better - it's compounding returns. But people aren't going to pay 100x their current bill for 4x results.
reply