@aliceisplaying so they prioritize Opus 5 or whatever, but feel that they should give access for a week once they can release it again, but then also they don't want to lock down compute for that long.
I treat this like early high end reasoning models where access was painful until it got cheap
@aliceisplaying I mean, that isn't odd to me, if they were internally going "we're going to start training a next tier model on June 22nd when we have more compute available, hype people" as they announced first time, with the obvious training target being Opus 5 to serve cheaper
@distributionat the repetitive tone of (ex) 4.6-4.8 is the opposite extreme and probably works less well since it grinds with repeated focus/phrasing. Fable is more ~manic, but also this is talking to someone discussing math, Fable talks very reasonably on mundane topics with less tropey-voice
@nosilverv #22 is the default in normal spheres, of "find your calling", jumping between majors, etc. Though even there, we fall short of generally useful instrumental goals.
In tech areas #1 is moreso because people know the sheer possibility, their (idealized) limits, and dreams
@nosilverv Hm, the opposite, it influences sure but it makes us less provincial than perhaps we should be, less adapted to the odd times we are in as the rapid growth and wealth grows mundane in the light of 2000 years since AD;
A year since (idk) 1776 would make us marvel more at ourselves
anthropic delayed fable just long enough that my extra usage credits they gave out a while ago finally disappeared
right when I actually needed them most
@NeoliberalAkiho most people just don't have the bandwidth for tracking politics
and this would allow for smarter coalitions and smaller independents to pop up and attract voters as dominant news cycle constrains less when you have a "voting research" class that isn't just... pundits or news
@NeoliberalAkiho tbh we should've had voting advice for a smallish fee for ages. Go pay the neolib-guy who spends all his time investigating candidates, give him your particulars, he sends you a readout of how they fit your values and a suggestion of who to vote for
@seconds_0 If LLMs stopped around here, I think we'd see those applications moreso along with people falling further into reliance. But the people who would probably be most drastically accelerating their own knowledge realize that it doesn't stop here, so boosting oneself isn't as valuable
@BjarturTomas which leads to market failure for individuals getting past their local flaws
and for society to get past coordination problems against companies/goods that they want to discourage but is harder to affect anything unilaterally (vegan food; random sketchy company like Nestle; etc)
@BjarturTomas Generally lack of good commitment mechanisms especially for money. Both for individual ("bank card, please make so I can't buy cigarettes without dumping 200 into my savings") and also wider social (no way to sign a vegan petition/whatever that stops your card from buying meat)
@fleetingbits That is, reputational/ease decide quite a lot for how companies manage themselves even beyond capacity.
Though of course capacity matters a lot here, they want to keep free as much beefy machines as possible for future training runs, or even simply internal Mythos deployments.
@fleetingbits and so they burn their usage limits and complain.
One route is to make a Fable-Only usage limit separate from normal usage, but that will also engender complaints because of the desire to have a universal medium of account for usage
So a bad branch either way for Anthropic
@eternalism_4eva Strongly privileged notion of counterfactual, such that the question of "what if X didn't happen" makes sense in some independent fashion 100% of the time.
Or fundamental randomness as the prime mover is one people go for a lot I guess.
@MagusGenji@goblinodds And I'm saying those effects make your indirect "they develop secretly" substantially less relevant. That indirect effect is also obvious...
@MagusGenji@goblinodds short timelines clearly make indirect effects less relevant, especially if one views that once you reach a threshold that there's no putting the genie back in the bottle with a surprised "oops". If the company has to work in secret that does actually hamper them quite a lot.
@nosilverv (and books give you time to idle over the insights, which lets you internalize things you learned and heuristics you update; whereas shortform throws the next piece of Content in front of you, effectively shaping your instincts without much reflection)
@nosilverv Shortform you can just get jerked around by enemies and allies. It is more adversarial and tuned to grinding against your specific mind without even a remotely coherent backing philosophy like a book might have, books usually have some vaguely underlying logic at the least.