Simon Sudarushkin

Your AI app works in the demo. I make it work in production. Shipping open source AI tools in public:

[Simon Sudarushkin @simon_ether]

Hinton, Bengio and OpenAI's chief scientist say AI could compress years of progress into months.

Read all 14 pages. The speed-up argument rests on one number, r: above 1, AI research feeds on itself; below 1, it fades.

Two of the three 90% ranges they give for r go below 1.

[Simon Sudarushkin @simon_ether]

@ChrisGPT By SemiAnalysis's own numbers a maxed-out Max 20x costs about $940 a month to serve ($11.7k of Opus at a 92% API margin). So the $200 plan only makes money while people use less than ~21% of it, which sounds low for anyone who lives in Claude Code

[Simon Sudarushkin @simon_ether]

@thsottiaux @gazzzzzzman If 3 cups got you this day, I'm scared of what cup 4 ships

[Simon Sudarushkin @simon_ether]

@thsottiaux fwiw that's max vs max. Sol max lands around Opus medium-high on AA, and if you match them at the same score like @scaling01 did, Opus gets 1.3 to 2.9x more tasks a month.

[Simon Sudarushkin @simon_ether]

@thsottiaux If you divide SemiAnalysis's plan values by AA's cost per task, Max 20x gets you roughly 2,000 Opus tasks a month and Pro 200 about 2,900 Sol tasks. So the 5x gap flips to ~1.5x for OpenAI, assuming AA's tasks look anything like what you actually run.

[Simon Sudarushkin @simon_ether]

@trq212 Do you have a rough number for how much smaller the plans come out vs raw HTML?

[Simon Sudarushkin @simon_ether]

@tszzl The intelligence explosion paper's first ask is auditors embedded in the labs, like NRC resident inspectors at nuclear plants. Right now we know AI does 26% of Anthropic's internal R&D only because Anthropic chose to say so.

[Simon Sudarushkin @simon_ether]

@SemiAnalysis_ Your agentic mix matches my own Claude Code logs almost exactly. Over the past 7 days I got 97.5% cache reads, 2.3% cache writes, 0.2% output and almost no uncached input.

[Simon Sudarushkin @simon_ether]

The auditors part is the one I care about. Anthropic says AI went from 1% of its internal R&D work in March to 26% in August, with only high-level human oversight. We know that because Anthropic decided to tell us. Nobody outside can check it.

[Simon Sudarushkin @simon_ether]

Hinton, Bengio and OpenAI's chief scientist say AI could compress years of progress into months.

Read all 14 pages. The speed-up argument rests on one number, r: above 1, AI research feeds on itself; below 1, it fades.

Two of the three 90% ranges they give for r go below 1.

[Simon Sudarushkin @simon_ether]

They want auditors sitting inside the labs, like resident inspectors at nuclear plants. Also air-gapped R&D, a cap on how fast capabilities can grow, and a way to pause specific workloads.

[Simon Sudarushkin @simon_ether]

Plug their central r into the model, assume full automation and no other bottleneck, and progress gets 10x faster within about 1.5 years. Then a year of today's progress takes about 5 weeks.

They do flag that r was measured while compute was booming, which could inflate it.

[Simon Sudarushkin @simon_ether]

Frontier labs employ thousands of human researchers. By the authors' math, OpenAI alone has the compute to run about 20 million AI ones (anywhere from 2 to 200 million), if expert-level AI ends up as cheap to run as today's models.

[Simon Sudarushkin @simon_ether]

Paper: https://t.co/cG746pi0Oc (footnote 6 has the ranges)

[Simon Sudarushkin @simon_ether]

@_achan96_ @geoffreyhinton Footnote 6 puts two of the three 90% ranges for r below 1, and the supplement says compute scaling would push r up. Is anyone estimating r from a stretch where compute stayed flat?

[Simon Sudarushkin @simon_ether]

@geoffreyhinton The speed-up math in the paper runs on one parameter, r. Two of the three 90% ranges for it go below 1, where progress fades instead of accelerating (footnote 6)