Eric Florenzano
banner
ericflo.bsky.social
Eric Florenzano
@ericflo.bsky.social
Building assistance! Used to make AR/VR. Before that made mobile apps. Before that made web things. People first, then products, systems, data. he/him
Vote!
November 4, 2025 at 8:16 PM
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...
On-Policy Distillation wtih giant system teacher prompt prefix
On-Policy Distillation wtih giant system teacher prompt prefix - opd_teacher_prefix.py
gist.github.com
November 4, 2025 at 5:17 AM
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?
November 1, 2025 at 8:38 PM
Reposted by Eric Florenzano
github style heatmap of your bluesky posts for the past year
yep! I've never used this native sharing method, but let's try it:

chatgpt.com/canvas/share...
October 31, 2025 at 8:28 PM
Turns out bf16 sucks, who knew, precision > range github.com/sail-sg/Prec...
October 31, 2025 at 10:14 PM
If it's true that intelligence is somehow fractal, as some researcher said offhandedly on a podcast, then it would make sense...
October 31, 2025 at 1:23 AM
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?
October 25, 2025 at 9:57 PM
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.
October 25, 2025 at 6:29 PM
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.
October 25, 2025 at 4:35 AM
At Nouscon
October 25, 2025 at 2:52 AM
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 💐💐💐
October 23, 2025 at 6:14 AM
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)
October 22, 2025 at 7:05 PM
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-...
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!
events.linuxfoundation.org
October 19, 2025 at 11:29 PM
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.
October 19, 2025 at 7:22 PM
#nokings starting to gather at the Civic Center in SF
October 18, 2025 at 9:47 PM
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
October 18, 2025 at 4:54 AM
I think Haiku 4.5 is pretty exciting and will fill a lot of gaps
October 15, 2025 at 6:26 PM
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...
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...
github.com
October 14, 2025 at 5:10 PM
Feel like we're back in an era where a website is the best way to get a new idea out there. Huh!
October 13, 2025 at 5:53 PM
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...
Your Algorithm, Your Rules: The Dawn of Truly Personal AI - Eric Florenzano
Picture this: You open your social feed and every post feels like it was handpicked by someone who knows you better than you know yourself. Not because some Silicon Valley giant built a better algorit...
neuralcolumn.com
October 12, 2025 at 3:36 AM
That feeling when you can't find something that you saw online recently! I don't even want to admit how long or stubbornly I've been searching today for this cool artistic video I remember seeing.
October 11, 2025 at 7:47 PM
Everyone talking about TRM and it is definitely interesting. Thoughts:
1) disappointing that it doesn't scale up
2) still needs plenty of FLOPs
3) would be cool to project into LLM embed space
4) unclear how it could be generalized
5) feels either foundational or dead-end
arxiv.org/abs/2510.04871
Less is More: Recursive Reasoning with Tiny Networks
Hierarchical Reasoning Model (HRM) is a novel approach using two small neural networks recursing at different frequencies. This biologically inspired method beats Large Language models (LLMs) on hard ...
arxiv.org
October 8, 2025 at 8:57 PM
Drop the "vibe" - just coding. It's cleaner.
October 8, 2025 at 2:47 AM
I've defaulted to Ubuntu/Debian for so long for Linux stuff that I'm out of my wheelhouse on other distros, but I needed a newer kernel and was having trouble with mainline so I switched to Manjaro. Manjaro has been absolutely great so far! Try new things out y'all!
October 4, 2025 at 6:07 AM
Nuremberg-style trials if we make it through this as a country
October 4, 2025 at 6:03 AM