Aldo Cortesi

I make software, break software and make software that breaks software.

[Aldo Cortesi @cortesi]

Announcing spacecurve, a space-filling curve library with a web + native interactive playground.

[Aldo Cortesi @cortesi]

Disappointing. Timid, backward-looking recommendations completely inadequate to the moment, with a side-helping of peevish, finger-wagging self-interest. If this is how the mathematical community chooses to respond, it will be completely overtaken by what's coming.

https://t.co/hBJl8M3Dl6

[Aldo Cortesi @cortesi]

Grok CLI is an excellent harness, and Composer 2.5 is a good, fast workhorse for quick refactoring (though too weak for deep work). I'm reaching for it more and more, which is surprising me.

[Aldo Cortesi @cortesi]

Second, Codex used to be the model I could set to work without worrying about running out of a $200 subscription. Now, my data shows Claude is the model that I can set to work at a rate of about 10% of weekly quota per day, while for the same work Codex will eat 30% and run out.

[Aldo Cortesi @cortesi]

Two flippenings, in two directions. First, Claude used to be more personally pleasant than Codex. Now it's a nitpicky, unenthusiastic, perpetually caveating scold. Codex is still a low-affect worker bee but I can work with that.

[Aldo Cortesi @cortesi]

https://t.co/31uNrnKq9Z

[Aldo Cortesi @cortesi]

And here are the same graphs in term of wall-clock time. Interpret with caution because a) GPT got to reap the big wins early on, b) I stopped Claude 4.8 often in its early run for subjective code evals. I'd say this nets out to roughly the same progress slope.

[Aldo Cortesi @cortesi]

One fascinating finding. Both agents are on their respective $200 tiers. Opus is using about 1/10th of the weekly quota per day on max thinking, while GPT 5.5 is using about 1/4 weekly quota per day on xhigh. Unexpected.

[Aldo Cortesi @cortesi]

The longer term of my Opus 4.8 comparison actually looks a bit more flattering. Clearing issues roughly inline with GPT 5.5, in a domain where all the big wins have been reaped.

[Aldo Cortesi @cortesi]

https://t.co/JnT6C24SZf

[Aldo Cortesi @cortesi]

More data in my ongoing Opus 4.8 vs GPT 5.5 task clearing runoff. Agents are doing a very large C++ to Rust port. The tasks are extracted unit tests from upstream that need to give the same result in our Rust type checker. Deep in diminishing returns now.

[Aldo Cortesi @cortesi]

This can be completely explained by how the interaction is framed in terms of the training corpus and doesn't require any reasoning about model agency, consciousness or personality.

[Aldo Cortesi @cortesi]

It's my firm belief that many people get sub-optimal results because they're rude or abusive to the models. In the age of AI, nicer people also produce better code.

[Aldo Cortesi @cortesi]

@snesworld90 I mean, this is surely because of some nonsense you have in your system prompt or memory, right? Did this 10x, and all responses were reasonable.

[Aldo Cortesi @cortesi]

And here are the same graphs in term of wall-clock time. Interpret with caution because a) GPT got to reap the big wins early on, b) I stopped Claude 4.8 often in its early run for subjective code evals. I'd say this nets out to roughly the same progress slope.

[Aldo Cortesi @cortesi]

More data in my ongoing Opus 4.8 vs GPT 5.5 task clearing runoff. Agents are doing a very large C++ to Rust port. The tasks are extracted unit tests from upstream that need to give the same result in our Rust type checker. Deep in diminishing returns now.

[Aldo Cortesi @cortesi]

@hen0s1s

[Aldo Cortesi @cortesi]

@hen0s1s No fast mode.

[Aldo Cortesi @cortesi]

@Xxi5olc For this particular project, doing this particular piece of work, running only a single agent... yes, that appears to be the case.

[Aldo Cortesi @cortesi]

@antor D'oh. Of course. Let's just say I had some other things on my mind! :)

[Aldo Cortesi @cortesi]

Early data on Opus 4.8. I switched a task queue for a complex project over from GPT 5.6. Case resolution progress slowed down... BUT the patches read very well and show taste - often including strong consolidation and code quality improvements.

[Aldo Cortesi @cortesi]

The longer term of my Opus 4.8 comparison actually looks a bit more flattering. Clearing issues roughly inline with GPT 5.5, in a domain where all the big wins have been reaped.

[Aldo Cortesi @cortesi]

This is Claude trying to warn you that you're about to blow through your whole token budget in 30 minutes.