Post #2601073

2026-03-20 09:18 UTC

@anttipeltola@mastodon.world according to the linked article they only ran a benchmark *after* the LLM rewrote SQLite. Yeeeaaah, that’s not gonna work. Here’s what happens when you tell the LLM to run a benchmark continuously after each change and discard any patch that makes things slower: https://awesomeagents.ai/news/shopify-ceo-ai-agent-liquid-engine-53-faster/ Feedback loops matter. They’ve always mattered for hand written software but they matter more for AI because LLMs are good at reacting to corrective instructions.

Replies (4)

@trademark@fosstodon.org 2026-03-20 11:44

@kumarvibe@mastodon.social @anttipeltola@mastodon.world The guy who did the benchmarks tried to see if an LLM could fix the problem itself https://substack.com/home/post/p-191241059 he tried three times with differently skilled hypothetical personas.

Open ##2601074
@MikeStok@mstdn.ca 2026-03-20 17:01

@kumarvibe@mastodon.social @anttipeltola@mastodon.world is that the same Lütke who’s happy to host Nazi stores on his platform?

Open ##2601075
@emma@orbital.horse 2026-03-21 01:11

@kumarvibe@mastodon.social @anttipeltola@mastodon.world okay, so now you're no longer programming, you're trying hill climbing and depending on what your stochastic text prediction loop does, you can end up at a local maximum because at the end of the day, it's a stochastic text prediction engine financed by nazis.

Open ##2601076
@anttipeltola@mastodon.world 2026-03-21 10:29

@kumarvibe@mastodon.social Shopify worker glazing over Shopify CEO in my post. I'm thankful I retired from tech in 2019. Things are really going to shit.

Open ##2601081