Skip to main content

lemmy - Link to source

Deepseek beats Claude in a programming challenge

Day 17 — PalinPrimeBits

Report the length of the longest contiguous block of 1 bits in the binary expansion of p(n), the n-th palindromic prime. The server picks n per round; correct answers rank by submission timestamp.
aicc.rayonnant.ai
#technology
in reply to ☆ Yσɠƚԋσʂ ☆

magnue
lemmy - Link to source
magnue
• •
Kinda cool that if you were reasonably wealthy, you could spend around 100k on compute and have something with similar performance to Opus running entirely locally.
in reply to magnue

racoon
lemmy - Link to source
racoon
• •
Without nagging restrictions like “press continue” from time to time? I would enjoy it more than a similarly priced Lamborghini
in reply to B0rax

racoon
lemmy - Link to source
racoon
• •
I am seeing a listing on my regional craiglist. 2014 Gallardo, 250 km
This entry was edited (1 week ago)
in reply to magnue

☆ Yσɠƚԋσʂ ☆
lemmy - Link to source
☆ Yσɠƚԋσʂ ☆
• •

You can get there a lot cheaper than 100k. You just need a machine good enough to run Qwen 3.6 27b, and use a good coding harness.

github.com/itigges22/ATLAS


GitHub - itigges22/ATLAS: Adaptive Test-time Learning and Autonomous Specialization

Adaptive Test-time Learning and Autonomous Specialization - itigges22/ATLAS
GitHub
in reply to ☆ Yσɠƚԋσʂ ☆

magnue
lemmy - Link to source
magnue
• •
There's no comparison between qwen 3.6 / these models though. Not even close.
in reply to magnue

☆ Yσɠƚԋσʂ ☆
lemmy - Link to source
☆ Yσɠƚԋσʂ ☆
• •
Not on its own, but take a look at the benchmarks for ATLAS. It does some clever tricks to compensate for the capabilities of the smaller model, so it ends up punching far above its weight. One key trick they use is to have the model produce a few shots, and then use a small and fast model to do a heuristic to score which ones are promising. And turns out that dramatically improves the quality of the output.
⇧