A dip in the token price index can imply the scenarios mentioned in this post. OR it can also imply that tokens are being produced more efficiently.
Aschenbrenner's Situational Awareness outlined 3 sources of progress towards AGI/ASI. 1) compute (more GPUs). 2) algorithmic efficiency (more tokens per watt or $). 3) Unhobbling (smarter use of the intelligence to actually solve problems).
From 2019-2024 we witnessed 1-2 orders of magnitude (!) improvement in just algorithmic efficiency. Aschenbrenner forecasted a 0.25-0.75 ANNUAL continued improvement in efficiency between 2023-2027.
Falling token prices are a well understood phenomena, expected to continue, and contrary to Burry's post, a signal of strength for the AI thesis. Falling costs per tokens is ESSENTIAL to achieving AGI, the intelligence explosion, and superintelligence, because the end-goal is VERY long running tasks, weeks, months, years, of exploratory hypothesizing and problem solving. Per token costs should fall from dollars, to dimes, to pennies.
The metric that matters most is the RoI hyperscalers/neoclouds are earning on their equipment. So far, H100s which should be obsolete by now (released in 2022 with an expected 4-year useful life), are seeing rental rates INCREASE as demand for intelligence explodes.
The key thing that Burry ignores is that AI is as dumb today as it will ever be, and it's more capable than most outside of software engineers realize. How much demand will there be for tokens when you can spool up a marketing agent that manages your entire company's $100M marketing budget, decreasing CAC by 50%, and saving the salaries too? It will run autonomously 24/7 exploring opportunities, tracking results, and optimizing ads.
Yea, that's going to require a few more memory chips. buckle up.
@mikepat711 The thing I care most about, which I'm sure we won't learn, is how much revenue robotaxis are earning per month. A very important unknown for modeling.
@Yuchenj_UW From the perspective of Aschenbrenner's Situational Awareness, this is an instant 0.5 OOM efficiency gain. We're on track for AGI in 2027. 🤯
@AlternateJones Since the S1, the two deals $SPCX signed with Anthropic and Google have increased revenue by ~2.5x. So one more double and they're there.
I'm the anti-Farzad. I think we'll see model Ys produced with no steering wheels or pedals, and we'll see conversion kits for existing model Ys.
The economics of turning the entire rolling stock of HW4 vehicles into cash cows is too great to resist, and it increases the market value of new and used teslas.