Eric Florenzano

I dunno if it's just me or what, but I never enjoy being frustrated by media. Scare me. Make me laugh. Intrigue me. Infuriate me even - I love to hate a good bad guy! But if I'm *frustrated* by stupid actions of characters, that's not fun. This is a Hijack S2 post.

But Anthropic is the most authoritarian and tight-gripped about the use of its models, and it charges the most for them - which seems in direct tension with those two concerns, and with each other.

Watched both Dario Amodei interviews on a flight back to SF tonight, and I have to say I'm really surprised that he seems so concerned about authoritarian use of AI, and its equitable diffusion. Because I agree about these being big risks!

Shower thought: the old concept of a "webmaster" - a single directly-responsible individual who owns a company's web presence. Seems both historical and futuristic at this point.

I bought the domain infinitecontext.net right after the first RLM paper dropped, but eventually came to the above conclusion

Recursive Language Models in ADK Implementing and extending the most exciting Agentic Paradigm of 2026 Recursive Language Models (RLM) is a promising architecture proposed by Alex Zhang and colleagues at MIT. In their paper, publishe...

My RLM hot take: RLMs are a breakthrough pattern that all agent harnesses will adopt and also bake into their RL. But IMO practitioners needn't care or worry or adjust, because the agent harnesses will all make this change transparently, and have started already. discuss.google.dev/t/recursive-...

Have to admit I'm kinda bummed we don't have a local A3B to match Sonnet 3.7 from last year. That model was the bar, in my opinion, at which the coding agent harness could Do The Work. It took more skill back then, however. Nothing in the A3B model size comes close still.

It's only useful because our training data has so much humanity in it. I see it like when you accidentally flip two signs in your math lib and the results come out OK for the most part because the signs cancel each other out, but not on all operations.

It's just this little prompt, but packs so much impact: github.com/obra/superpo...

If you haven't tried the superpowers:brainstorm skill, try it! It FLIPS THE ROLES of the developer and Claude Code. You tell it the feature you want, and it starts working on a spec, interviewing you, asking questions, etc. Hard to believe a two-paragraph promptlet can improve Claude Code this much.

I think this is a manifestation of it: en.wikipedia.org/wiki/Chatbot... as well as people "falling in love" with their chatbots, and then the moltbook stuff over the past few days: how the reddit training data attractor seems to be convincing people of personhood...and this is early days still.

Anthropomorphism is rapidly rising to the top of my list of AI worries, past concentration of power, and past rapid job change. Because we can't fix anthropomorphism - it's a bug in our hardware.

We need this to end now. We need to end this now.

I'm gonna run out this week 😭

Jan 8 Reports A collection of 3 reports

It's so good os.eflorenzano.com/public/news-...

I like the reports my little vibe-coded news app produces os.eflorenzano.com/public/news-...
os.eflorenzano.com/public/news-...
os.eflorenzano.com/public/news-...

The worst part about the ICE execution in MN is knowing they'll use it to punish more people and expand ICE further, and now they'll probably get their beloved martial law they've been salivating over. America's enemies are loving all of this BTW.

Flipping through the TV, Peacock's top movie is Bugonia. Me: "looks interesting and Emma Stone is in it" ... 2 hours later ... W T F did I just watch? Insane movie, great acting.

War is hell and we have opened the gates wide. Another mark of shame we'll bear, forever.

YouTube video by Zeyuan Allen-Zhu, Sc.D. Physics of Language Models: Part 4.1b, Canon Layers & Architectural Principles

And another one youtu.be/ssEmQeXHthA

ericflo (Eric Florenzano) User profile of Eric Florenzano on Hugging Face

Made 4 new hopefully-orthogonal synth pretraining datasets inspired by Physics of Language Models:

- Dyna: track state via non-commutative ops
- Inducto: infer rules from few-shot examples
- Filtro: follow a graph while filtering noise
- Logic: evaluate XOR circuits

huggingface.co/ericflo/data...

YouTube video by Zeyuan Allen-Zhu, Sc.D. Physics of Language Models: Part 4.1a, How to Build a Versatile Synthetic Pretrain Playground

Ooh, new Physics of Language Models is out! youtu.be/x3G8knjPDbM?...

I completely believe that too - I've been doing a lot of Python and data work, which is super well-represented in its training data

I worried when I posted this back in March that it would seem like hyperbole but it has been pretty much spot on:

🎯

"What if we made a movie where everyone is awful and makes nothing but terrible self-destructive decisions?"

Stranger Things S5: 👍
Marty Supreme: 👎
Heated Rivalry: ❤️

Nice job Codex CLI - hit enter over 4 hours ago and it's still going strong.

60 minutes CECOT segment. Recorded through my browser. HD quality. Not phone. Spread it around. If you're press, it's yours. Go nuts.

It's streaming on a Canadian website, but there is no telling for how long.

www.filemail.com/d/wkcdttnacp...

Following ARC-AGI progress has taught me more about humans than computers, I think. The most surprising thing so far is how poor humans are at reasoning about necessary vs. sufficient conditions. Many of the smartest human minds get really tripped up on this!