Reactions from ranked influencers
6 posts
too much wordcelism around Dean's Decelgate I want estimates of 1) how much profits on inference do "The Labs" (don't fucking call them labs please, these are corporations with political ambition that makes East India Company blush) need to maintain progress within 90% of this pace 2) the maximum margin and volume compression incurred by existence of Chynese open weights within 6, 4, 2, 1 months of the AI corps' current public frontier 3) interactions between those
Creating better AI models costs money. If you give the models away for free, you do not get money to train new ones. It’s insane that this is even a conversation, but people want to hurt OAI/Ant for political reasons, others want to pump their open source / inference bags, and some just want to be popular — and giving things away for free is always popular.
@AndrewCurran_ The key thing to understand is that tokens are priced against hours of white collar labor automated, and the premium on better, lower-error, higher value labor can scale infinitely The frontier can be marginally ahead and drastically more profitable it's all bs
@teortaxesTex In the alternative path who is going to build Stargate? Who is going to build Starmind? We absolutely can't slow down or stop here. Maybe I just don't understand the third path being proposed?
Who is going to build Starmind? Let's see. Sol on DeepSeek costs to serve with $2/hr GPUs: V4‑Flash: $0.035 per 1M output tokens V4‑Pro: $0.10 per 1M output tokens 70–85% margin. OK, maybe 60-70 if we choose higher interactivity (70 tps). These are models of the ≤1e25 scale, maybe <4e25 for the entire project. Anthropic/OAI can overtrain smaller models on more tokens, use faster hardware, and frankly design more efficient architectures, generate stronger data with their triple Ph.Ds and AGI in the cuck shed… not to mention superior mech interp augmented training projects. They can, if they so choose, SERVE EVEN CHEAPER. FIVE CENTS FOR A MILLION TOKENS OF FABLE-TIER MODEL IS POSSIBLE ALREADY OR IN THE NEAR FUTURE. They don't need to subsidize shit even on maximally utilized subscriptions. They are printing cash. Elite data annotation also rounds down to dust, compute for next training runs and gobbling up the economy are their main expenditures. They're so insufferable on the TL because they're power tripping. Inference is virtually free money firehose for their companies, tens of billions of dollars annualized and going up. Margins are above 80% No Matter What They Do, and this is bought with the simple and brutal merit of standing on the frontier of capability, speed and reliability, the fact that *I paid for Sol* to make this illustration. Mere 1 or 2 months of lead is enough to earn the commanding share of the expanding market with rapidly improving hardware and exponentially improving algorithms. This is what Dario meant by top companies accelerating away from the competition, he just didn't price in that the competition would also follow on an exponent. You should PISS on frontier employees who whine and cry and poast Arrow about muh chynese open source denying them capex for reinvestment into glorious progress and bringing forth "communism". Distillation doesn't even dent their bottom line, it's a pathetic fast following hack, and fast followers don't win in a superstar-dominated market. Frontier companies are *economically invincible*. The frustration of these dorks exposes merely their grandiose dreams and neurotic hopes to become Cosmic Overlords living on equity priced in parsecs, owning large swathes of the lightcone where nobody else has any agency or perhaps even exists. That is, by the way, why everyone above 3SD in SF is trying to get into a "lab" (reminder, it's not a fucking lab). The only reason to tolerate their noise is that they'll likely fail. That, however, depends on hard and currently poorly modeled physical constraints. We shall soon learn which prevails between atoms and bits.
@teortaxesTex In the alternative path who is going to build Stargate? Who is going to build Starmind? We absolutely can't slow down or stop here. Maybe I just don't understand the third path being proposed?
@AndrewCurran_ I tried to estimate this, and it seems that models within 2 months of the frontier can't impose >12% of cut on frontier CAGR and capability progress. I think this is all bullshit neuroticism of wannabe feudal lords. They need to show the math or get bullied off the website.