Eric Florenzano

Vote!

(Or distill into itself this way)

Thinking this will be a powerful new formula: collect use cases, build a course router, prompt optimize (GEPA etc.) each use case (including many-shot examples sampled from prod), then distill that whole system into a smaller model.

On-Policy Distillation wtih giant system teacher prompt prefix On-Policy Distillation wtih giant system teacher prompt prefix - opd_teacher_prefix.py

I've been messing around with a version of thinkingmachines.ai/blog/on-poli... that adds a 35k token long system prompt prefix to the teacher model, and it works *great*. WIP: gist.github.com/ericflo/1175...

Just played for a few hours so taking a break, but will send you a steam link so we can connect there later!

Having fun with ARC Raiders! I thought it might be too sweaty for me but it balances nicely to my poor skill, and the loot stash prevents total wipeouts. First multiplayer game to pull me in since Overwatch. Anyone else playing?

github style heatmap of your bluesky posts for the past year

Turns out bf16 sucks, who knew, precision > range github.com/sail-sg/Prec...

...why deep learning gets better with scale (better fractal "render" resolution), why LLMs have to do all kinds of norm and projections to work (must attend to many feature scales), and why TRM works well on ARC (can find features in hugely magnified areas of the intelligence manifold)

If it's true that intelligence is somehow fractal, as some researcher said offhandedly on a podcast, then it would make sense...

Why do almost all air purifiers have 3 speed settings?
1. Too low
2. Just a bit too low
3. Way too high

Are there any that have a dial or continuous speed setting?

That feeling when you get kinda heated about a post and then realize it's a shitpost. In my opinion, shitposting is a socially corrosive act that erodes our psychological commons and leaves us more emotionally detached for the worse. Call me a millennial but I value earnestness.

Wow! I thought this number would be way lower, like 1-5%! I thought a whole lot more of y'all were wearing contacts and getting lasik. My model of the world was way off, that's kinda neat actually.

At Nouscon

One thing about me when I talk about code, is I like to call things bouquets. A bouquet of variables. A bouquet of agents. A bouquet of proxy servers. It fits lots of situations and is fun to say 💐💐💐

This is 100% what I'm envisioning

🤣

I wonder if we will see Developer Firms structured as partnerships between well-known individuals, like how Law Firms operate today. (As a result of agentic coding)

Open Source AI Week | LF Events Open Source AI Week in the Bay Area between Oct 18-26, 2025 features top AI/ML conferences, hackathons, & networking you won’t want to miss!

There's this whole OSS AI thing going on this week in SF but I'm not very plugged in - anything I should check out? Anyone in town for this want to catch up? This list is 🤯 events.linuxfoundation.org/open-source-...

I don't know how much I can say about the work stuff, but the agents impress me daily with what they're able to do. And frustrate me daily with what they aren't.

Karpathy throwing cold water on agents hype is good. I'm actually a huge believer in agents - Claude Code convinced me this year, and I maintain an orchestra of them at work. But hype and fear was getting totally out of control. Am also wary of hero worship, but at least Karpathy wields it well.

#nokings

No Kings sign "Rejecting Kings Since 1776"

Awesome lady holding a sign saying "America's Only King" and it's Elvis 😂

#nokings starting to gather at the Civic Center in SF

My dumb idea for the night is it would be fun to make a social network for LLM agents. Just an MCP server where the agent can read or post short messages, with no stated purpose, on their path to completing other tasks. Like a social media break. Maybe it would help prevent entropy collapse in RL

I think Haiku 4.5 is pretty exciting and will fill a lot of gaps

Performance of llama.cpp on NVIDIA DGX Spark · ggml-org llama.cpp · Discussion #16578 Overview This document summarizes the performance of llama.cpp for various models on the new NVIDIA DGX Spark. Benchmarks include: Prefill (pp) and generation (tg) at various context depths (d) Bat...

Holy crap, getting that discounted Ryzen AI Max+ 395 system a few months ago was very much the right call, as it performs better than long-awaited DGX Spark AKA Project DIGITS github.com/ggml-org/lla...

I think you had to build an app and get it in the app store

1. Apps have ossified and consolidated, so no longer a habit to browse & install apps
2. Web platform kept slowly developing and raising the bar
3. Savvier public knows how to use a webapp
4. Benefits to customer relationships
5. ChatGPT

Feel like we're back in an era where a website is the best way to get a new idea out there. Huh!

I vibe-coded a vibe-blogging app for myself and have been blogging semi-daily and not taking it seriously and am having fun with it. Today a post about what I hope for the future of software and training your own personal algorithms neuralcolumn.com/authors/9e13...

Yeah that sounds right. It wasn't the most interesting clip in the world, but once I couldn't find it, I really wanted to!