Prompts for Variations on a Theme

The seed

Can you write a pithy response to “AI Is Garbage and a Bubble (Please Learn This)” by Richard Carrier for the one-year anniversary of that post? Or don’t, up to you. It’s somewhat of an unhealthy indulgence of mine to keep going back to that page. But Richard Carrier is apparently sticking to it:

PASCAL BERCKER: Will you be doing an update on this issue, given all the AI developments this past year?

Richard Carrier: This article is already up to date (it contains articles up to even this month). Because none of the developments you are referring to change a single point in this article. It’s all the same. Just variations of a theme.

Continuity

If you want a continuity with the blog, we originally addressed this in notes/which-agi.md

The S-1

I would put this as another link in the “Which AGI?” entry, but it’s not meaty enough yet to be worth it. It just states trivialities 😃 (tongue in cheek speaking). Could we combine this with a reaction to Anthropic’s S1?

The flat earth

“Carrier’s positive programme deserves a hearing.” – this briefly recapitulates “Which AGI?”, not worth more than one sentence with a link. “Whether this adds up to the self-model Carrier describes is open.” are you serious? This adds up to Carrier being a flat-earther, as he funnily enough accused me of being. (Not sure if comments are captured in the page save?)

Sources

The cranks page: “A Primer on Actually Doing Your Own Research” by Richard Carrier. The GPT-6 Astra result: https://arcprize.org/blog/astra

Never improve

We should also write something to “will never improve, ever”

The figure

That’s maybe charitable in anchoring to original publication date, but could you plot a graph of best SWE-bench Verified score over time till October this year?

This isn’t great, could you look up numbers for GPT Astra, GPT 6.1 Sol, Gemini 4 Carbon, Opus 5.5? Mabye switch to a benchmark that includes these?

Benchmarks as manipulation

If Carrier happended to respond “benchmarks are corporate manipulation, not reality” what would we say?

Yes, thanks!

Rewording Wikipedia

Let’s contrast “It’s just fancy auto-complete. And thus is just regurgitating the internet, and poorly. […] And Wikipedia is shitty compared to fully expert sources. And yet, indeed, most of what”AI” does is just reword Wikipedia at you, thus magnifying even its errors and inaccuracy. It’s garbage.” – with solving of long-standing and/or prestigious open math and computer science, and increasingly theoretical physics, problems?

Thanks but that’s not enough. Here’s a quick report from Grok: [a survey of 2026 AI results on open problems: Navier–Stokes, the quasi-Riemann hypothesis, local smoothing and Kakeya, Erdős problems, the Exponential Time Hypothesis, and Yang–Mills existence]

Source links didn’t copy. [Links to OpenAI on Navier–Stokes, Science on the controversy, Quanta on the quasi-Riemann hypothesis and on the Erdős problems, Crypto Briefing on local smoothing, New Scientist on the Exponential Time Hypothesis, and Unite.AI on the October proofs.]

And two links from Scott Aaronson, the first one someone posted to Carrier without reply from him:

https://scottaaronson.blog/?p=10169 https://scottaaronson.blog/?p=10232