Post #2937828
2026-05-14 05:16 UTC
@issackelly@mastodon.social Interesting. It's been a minute, but I recall someone getting pretty decent results off a 4-up 4080 rig and... maybe one of the deepseek models plus some custom wrapper agent code.
Replies (2)
-
@issackelly@mastodon.social 2026-05-14 05:20
@cda@macaw.social yeah, this is Qwen + pi on per hour rented hardware and I have seen similar benchmarks on the nvidia spark box. Unfortunately I don’t know if they’re subsidizing that, but I do know it’s power draw.
-
@issackelly@mastodon.social 2026-05-14 05:24
@cda@macaw.social also reading through the pi source code has been interesting. Turns out an agent is like…Llm inference in a for loop. Spending a couple days learning all the glossary words and basic architectures for the current boom has been pretty enlightening for where the complexity is.