for my fellow Italian moots

Last night I got vibecodable idea after watching the 3rd pretentious movie of the week. Paste your letterboxd username, get called out for pretentiousness by claude based on your recently watched films. It takes about 10 seconds and it's unfortunately accurate letterboxd-tribunal.pages.dev

Dedication?

@gwern.bsky.social on why human preferences and not synth data lead to modal collapse

has anybody ever done Claude-test driven development? Not just TDD with Claude writing the tests, I mean literally calling Claude code at test time to semantically validate the outputs (for fuzzy data processing things)

On a new machine

Happy to be on the cool side of the right neighborhood (the t̶p̶o̶t̶ tfea one)

Thread on the newly released X recommendation algorithm. I work on LLM-based social simulations so this is directly relevant for me and I spent some time digging through the code. github.com/xai-org/x-algorithm

for some reason by accident i am reading capitalist realism and the thirst for annihilation at the same time

huggingface preaching the gospel on what should actually matter in llm evaluation (especially point 3) huggingface.co/spaces/OpenE...

Thread on the newly released X recommendation algorithm. I work on LLM-based social simulations so this is directly relevant for me and I spent some time digging through the code. github.com/xai-org/x-algorithm

My old wifi contract expired, and the new one struggles to be born. Now is the time of monsters

Funny hot take theory: the first LLMs were bad at arithmetic because Turing wrote they should be, in the Turing Test paper, to appear human

Funny hot take theory: the first LLMs were bad at arithmetic because Turing wrote they should be, in the Turing Test paper, to appear human

Very bsky problem

The more I spend time in Austria, the more this feels true

We are hiring a Senior Scientist! If questions like "how much of Common Crawl can we actually store and process on our infrastructure?" get you excited, you would prob love this job. Come do research at fun scales with us

A thing I've been working on for the past year: an LLM benchmark on the dreaded Italian medicine faculty entry exam. I will present it in two weeks at @ailc-nlp.bsky.social Clic-it in Cagliari!

from this awesome blogpost: snats.xyz/pages/articl...

Life is going from one cool spot to train AI models to another In this fancy Berlin library you can listen to vinyls while you do it!

Me looking very dumb while pointing at things at two academic events: 1. Pointing at a logits inside @repligate.bsky.social's loom for the "Braive New World" conference at @uni-graz.at 2. Pointing at my poster at @ic2s2.bsky.social last week on improving LLM agentic natural conversation synthesis

Only prisoners have time to read, and if you want to engage in a twenty-year long research project funded by the state, you will have to kill someone. Sorry for the Fisher posting but it's so good

I look really bad/funny in this picture but I am glad I was given the opportunity to talk about LLM research to an audience of cool researchers! Fun fact: I also held a hands-on session that involved playing around with Bluesky data!

Karpathy on training Neural Networks: you should go slowly and be paranoid me, while vibecoding torch code: what if i just increase the paranoid part?

how not to name your ai agents

genz semantic embeddings engineering from my collegue

backpropagation was inspired by freud

“How about we pull over for a bit and get some rest?" - GPT4, when it's their turn to drive

I wish all textbooks were written like this one from @jurafsky.bsky.social‬

was true in 2012 and it is true now

Stumbled upon new ammunition for my personal struggle against RHLF: the tendency of models trained this way to give correct sounding answers make us believe in wrong things more arxiv.org/abs/2409.12822

Looks important for my rec algorithm people: hard confirmation that Google uses essentially hybrid rankings for search: page rank, bert embeddings and a secret ingredient (user data)

All of this mentions of Zero make me think about Slothrop

Reposting from the other site to spread it here: apparently, thinking that RLHF irons out creativity in LLMs is now corroborated by this paper arxiv.org/pdf/2505.00047

Protestant-Catholic German beefs are my favorite kind of beefs

At Weizenbaum today, talking about slightly more complicated chatbots!

A friend and colleague at my lab (the CS² under @janalasser.bsky.social ) is organizing a super nice school on democracy and social media here in Graz! If you're interested, check it out and share! Link: sicss.io/2025/graz/

Claude, my friend, are you ok?

Has anybody ever tried GRPO training with annealing wrt the generations temperature? I feel like that could be a way to explore more of the search space at the beginning of the training, to then be less "creative" once the model is hopefully already on the right track

And so it begins

The current state of LLM training getting me to read papers on how to recognize Dante's Inferno poetic metre was not something I would have guessed last year.

I will be hosting a livestream on 10th of this month (in Italian tho) to talk about my LLM structured polling data extraction project!

Happy to be here! I mostly post about AI/NLP/Computational Social Science research/personal experiments but also about my journalism (mostly either data-based or about US politics)