Experimenting with a novel decoding method to get LLMs to be more creative (not just fiddling with temperature)! countbayesie.com/blog/2026/7…

Save 45%, only on Sept. 1st! Today, you can save on Local Image Generation by @willkurt and other related titles: hubs.la/Q04vZdgV0 Learn how to generate, refine, and customize images locally without needing a background in data science or machine learning.

I knew this day would come. Looks like my dwindling reddit usage will drop to zero.

Live look at mathematics research.

This remains a magical place

YC founders moving to sf be like “settlers of g’tan”

AI is crazy. Before I was running this with a small model using HF, getting about 7 token/s. I was able to get a full transition to using Llama.cpp using Qwen3.6-35B-A3B-Uncensored-Q4_K_M, and getting 23.7 tok/s (Impressive given the extra computation required for this decoder)

This is the most accurate description of the current state of agentic coding in practice. I remember you used to solve a big, challenging brain problem and then know you had a few days to implement it. This this was boring, but heads down, headphones on work. It was relaxing.

Whipped up a UI for my future-entropy based decoder! By weighting the base probability against future entropy with a single parameter, you can then just apply a sine wave to give more variety to the decoder itself. Really impressed by how configurable this simply design is!

It’s curious that I’ve had a bunch of requests for the source code for my most recent “entropy decoding” experiment. This means there are still many people, interested in AI, whose first response is still not “point an agent at the article and ask it to recreate it”

What a strange time in software. My last round of interviews I used "leetcode style coding challenge" as an autoreject signal because it meant the team was not in touch with the current direction software is moving. It is hard to test signal now, but leetcode is not it.

This is a great exploration Need more of this to free us from slop jail

Interesting to explore the different possible paths different LLMs would take. Branch for the top 10 initial tokens and then doing greedy sampling comparing - Qwen2.5 0.5B - Qwen3.5-9B-Base

Books are so OP

It was one thing when "literally" lost its meaning, but I have an even harder time accepting that "nondeterminism", one of my favorite concepts in computing, has not only lost its meaning but lost it in within the context of computing itself.

The thing about Fable I don’t see anyone discussing: it’s obviously not just “more parameters go brrrr…” which makes me thinking open models might be less far behind than it seems.