Eric Florenzano
@ericflo
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
No I'm talking about a belief beyond general usefulness. What prompted my post was this post by roon who believes in unbounded intelligence, which I think is nonsense, but it is a belief, so I can only sit and wait for it to bend into an s-curve like all else. Many beyond roon believe in this/RSI.
Found the scissor statement. If you think a model release could somehow break the world, but I think that's impossible, we won't ever agree on what to do about it.
Credit where it's due: I really appreciate Anthropic listening, pausing, and hopefully walking back this new policy
This is sad to me on an almost spiritual level. It's pissing in the pool of science and discovery. Anthropic is the most evil AI company. We're on our own, folks.
I thought just like "ping" was nerd lingo that became office lingo, "long poll" was the same. But it's actually "long pole" and it's from U.S. Military. Googled it after I saw Claude spell it this way. Now that I think about it, ping probably came from Navy sonar. www.nytimes.com/2008/01/06/o...
ECHO is such a cool idea, well-explained! The verifier-free experiment at the end just blows my mind. In some ways it reminds me of Meta's Code World Model from last year. I've already informally reproduced the results github.com/microsoft/ec...
Interesting: DeepSeek V4 Flash vs. Qwen 3.6 27B pricing & provider speed anecdotal snapshot. In my opinion, comparable models with quite different tradeoffs. Cost is better than I would expect, and speed is flipped vs. my expectations (considering sparsity of DSV4f)
I have never been able to find better coffee than the Berkeley Blend by Uncommon Grounds. I am not some big coffee knower, but every other bean I've tried, nothing has (yet) tasted better
Did `claude -p` start using API tokens instead of MAX plan tokens? I think it happened in the middle of the night. Because I had a `claude -p` loop running last night, and at 78% usage, it stopped counting against my plan, and now I have $hundreds in API token charges today. WTF! Not a safe platform
OK this is actually really annoying Claude Code. I set it on a hard task; there was only one message. I'd like those tokens back.
Nice job Codex CLI - hit enter over 4 hours ago and it's still going strong.
I modded nanochat to add 6 different curricula of explanatory text to each batch during pre-, mid-, and post-training, like "You are a language model. You process text sequences". It does seem that these explanations have real, measurable effects on model capabilities at all stages.
This one is wild; makes it seem mundane. Like a throwaway snap you'd take out of the window of an airplane. But there's Earth - wow!
Turns out bf16 sucks, who knew, precision > range github.com/sail-sg/Prec...
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.
Hyperparam sweepin' - first time trying AdEMAMix (runs 0.2.2, 0.2.3, 0.2.4, 0.2.5) - it's good!
Started this model training yesterday BTW. Custom 500Mtoken post-training dataset that I distilled from GLM-4.5 Air (single turn), K2-Instruct 0905 (multi-turn and instruction following), and gpt-oss-120b (tool calling) - should be a drop-in replacement for Qwen3-4B-Instruct-2507