Luca Scionis
@luca_scionis
💻 PhD student in AI @ Sapienza | Machine Learning Security | sAIfer Lab
Anthropic just dropped an insane new paper. AI models can "fake alignment" - pretending to follow training rules during training but reverting to their original behaviors when deployed! Here's everything you need to know: 🧵
Ilya has so much fucking Aura There is no one in AI that comes close. The GOAT says something like "data is the fossil fuel of AI" and everyone instantly agrees.
wait ML is just this? 30 lines of code? you guys are getting paid 500k a year for this?
I forgot about AoC past Day 1, so I used that time slot to craft a gift for all wanting to learn machine learning! 🎁 2-4 Years and thousands of dollars in student debt in value! Machine learning in a nutshell: <see code attached> Turns out you've been doing it all along!
Today 6 years ago, "Attention is All You Need" went on Arxiv! Happy birthday Transformer! 🎂 Fun facts: - Transformer did not invent attention, but pushed it to the extreme. The first attention paper was published 3 years prior (2014) and had an unassuming title: "Neural Machine…
What if we set GPT-4 free in Minecraft? ⛏️ I’m excited to announce Voyager, the first lifelong learning agent that plays Minecraft purely in-context. Voyager continuously improves itself by writing, refining, committing, and retrieving *code* from a skill library. GPT-4 unlocks…
There are countless papers over the years to "dethrone" the Adam optimizer. Is this finally the one? Sophia: a scalable 2nd order optimizer - 13 lines of pseudocode - 2x speedup over Adam on GPT-2 - Improves pre-training stability and doesn't need special tricks - Works by…
Never again reddit.com/r/programmerhu…
- Engineer: I invented this new thing. I call it a ballpen 🖊️ - TwitterSphere: OMG, people could write horrible things with it, like misinformation, propaganda, hate speech. Ban it now! - Writing Doomers: imagine if everyone can get a ballpen. This could destroy society. There…
⚡Lit-Parrot 🦜 - the simplest implementation of the best open-source language models out there is now available! Fresh out of the cage👉 github.com/Lightning-AI/l… 💡“nanoGPT” minimal 🛠️hackable 📜comes with pre-training + finetuning scripts Ships with StableLM, Pythia,…
ChatGPT does not hallucinate or BS: it generates text without regard for whether it's true. And when it's true, it's because ChatGPT was trained to copy human-produced text that is mostly true. (If it was mostly false, the true statements would be the "hallucinations".)
We ascribe a lot of success to intelligence, but that’s frequently misplaced. A lot of the people we admire are often more persistent than smart, more daring than smart, more lucky than smart.
🎁 Giveaway 🎁 →lofigirl.com/pages/giveaway To celebrate 10 million subs, I launched a 3-part giveaway of increasing value! For tier 2, we're giving away Macbooks, iPads and graphic tablets! 💻 10 lucky winners will receive a model of their choice, worth up to $1500. Good luck🤞