23
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
this post was submitted on 16 Aug 2026
23 points (96.0% liked)
TechTakes
2654 readers
94 users here now
Big brain tech dude got yet another clueless take over at HackerNews etc? Here's the place to vent. Orange site, VC foolishness, all welcome.
This is not debate club. Unless it’s amusing debate.
For actually-good tech, you want our NotAwfulTech community
founded 3 years ago
MODERATORS
Bioman has already pointed out the "Economic Value" numbers they are using for these tables are probably bullshit (based on self reported run-rate extrapolations that are deliberate distortions at best, based on VC valuation at worst). To add to this... the compute values are also probably bullshit. They are likely based on data center announcements and not confirmed totally complete data centers (Ed Zitron has ripped into how much bs there is in data center announcements). "Coding Time Horizon" is probably METR, which, while some of the best numbers for estimating actual AI improvement for practical purposes, are still really bad in several key ways. (They don't have enough human task performers for the longer duration tasks even if everything else was right, because they aren't, and there are several ways systematic bias could have leaked in and compelted distorted the constructed measure of task duration.)
"AI Software R&D Uplift" is the single most important category to their scenario of recursive self improvement... and they have it at a small fraction of what they estimated.
According to one of the writers apparently 17% is supposed to be 'progress is 17% slower than expected', so the 34% figure here implies progress is 34% slower than expected, so that's actually worse
That wouldn't make sense, because 1.07 would mean progress is negative. I think they meant it as a multiplier? So 1.07 is almost exactly what the predicted, .17 is only 17% of what they expected.
Anyway, it doesn't really matter, because so much of the input numbers to these calcs are garbage.
Yeah that's what tripped me up, because if that was true then it'd make no sense why the percent of progress went up while the uplfit number went down.
The percentage isn't total progress to some key point of AI 2027, it is progress relative to their timeline. A constant 1.0 would be staying on track with their predictions, numbers less than that would be falling behind. So they are admitting the real numbers are falling behind their predictions more and more (while still not acknowledging their entire timelines was bs in the first place).
I read their spreadsheet and their % progress numbers in terms of AI capabilities is MASSIVELY skewed by Claude Mythos's METR score. The rest of it is below the mark if you exclude that.
They also admitted that the “China wakes up” prediction hasn’t happened yet, and Kokotajlo still claims that the scenario is “roughly on track” even though the US/China race is important for the scenario, and if China hasn’t woken up, that derails things pretty hard imo. So does OpenAI’s slowing down announcement
I took a quick look over the article again, this spreadsheet contains how they measure their metrics. "Compute" seems to be based off number of chips, which is probably in part based off data centres anyways
And apparently they have ditched METR and are now using "coding uplift (i.e., how much of a speedup AIs are providing to software engineers at AGI companies) and revenue."
Ed Zitron has also explained his suspicion that lots of GPUs are sitting around in warehouses waiting to be installed, in some cases sold (to juice NVIDIA's revenue) but not even shipped yet.
And I'm really skeptical speedup from AIs claimed by the LLM companies is in anyway related to reality.