we should extort free compute with an ai benchmark

[strawberry9]

On Wednesday, 21 October 2026 at 00:48:39 UTC, monkyyy wrote:
> They are dumping millions of dollars on ai slop that matches a 
> benchmark; including 10 scientists making their own problems 
> and being judges on their own pet problems.
> To bait the cultists they added in some cryptography about how 
> they encypted the problem statement before publishing and blah 
> blah blah
>
> https://www.youtube.com/watch?v=2fEb-SJU3TY
>
> we could likely find 15 or so problems, gather up as many phds 
> that will agree to grade a benchmark. and then maybe we get 
> some libs out of it or the ability to declare ai sucks spend 
> more money on dlang.

The problem with AI, is that it will flawlessly execute a bad 
plan for you.

I've used AI for the last month or so, to create my new favourite 
langauge, CobaltC: https://github.com/strawberry9/CobaltC

At first I thought it had done a really great job. Then it became 
apparent (5 weeks later) that the runtime it implemented is so 
bogged down by implicit, invisible bookkeeping, that the language 
cannot be taken seriously (yet).

I've instructed it move to a design that favours guarantees over 
optimisation cleverness. So we'll see what happens there...

The entire project has only cost be a little over $200 so far, so 
I'm not complaining too much about cost. This includes a 
full-featured standard library.

We will still need 'real' mathemeticians and programmers for a 
little while more though.

I'm looking forward to future models that won't just flawlessly 
execute a bad plan for you, but will instead tell you straight 
up, this is a bad plan and here's a better one that will save you 
money in the long run.