Monicaxie (@immonicax) started following @NeelNanda5 on Oct 16, 2024

Note: Being badly named doesn't mean current "unlearning" work is useless, I just think it should be renamed to eg "knowledge suppression" and people should be realistic about its limitations and how robust it's likely to be

This looks like a lovely paper! I've long thought that the term "unlearning" was BS. IMO unlearning means removing knowledge from the model's weights. But people tend to measure behaviour, which is indistinguishable from suppression in later layers. I'm glad someone showed this!

Highly recommend this resource on evals from @clefourrier !!

RT @hugobowne: 📺 Going live in 15 with @haileysch__ (@AiEleuther) to chat about The Future of Open-Source AI and Research Infrastructure in…

🚀 Exciting internship opportunity! Join the Llama team @AIatMeta and help redefine what's possible with large language models—from pre-training to post-training. Be part of our 2025 research internship and help shape the future of LLMs. Feel free to email or DM me 📩 Learn

🚀 New paper from our Llama team @AIatMeta! We discuss "cross capabilities" and "Law of the Weakest Link" of large language models (LLMs): 🔹 Cross capabilities: the intersection of multiple distinct capabilities across different types of expertise necessary to address complex, https://t.co/YP4m4koxln

RT @VictoriaLinML: 1/n Introducing MoMa 🖼, our new sparse early-fusion architecture for mixed-modal language modeling that significantly bo…

So excited for the open release of Llama 3.1 405B - with MMLU > 87, it's a really strong model and I can't wait to see what you all build with it! https://t.co/9Bg6m3HOFQ Also check out the paper here, with lots of details on how this was made: https://t.co/JXJn3d9p5H

On-device AI Processing, blockchain Integration, and a vision to challenge and decentralize the technological monopolies dominated by large corporations are just a few of the reasons @DispersionVC backs PIN AI. Read more here: https://t.co/Fe7hSQPV07

"Trying to learn how to be as human as you guys are" is the hardest thing about being a robot, says @Tesla's Optimus bot. AI has come a long way.

RT @AmitMandelAI: אם אתם מפתחים בעולם ה AI יש לנו יום מעניין מאוד בשבילכם, ואתם מוזמנים להירשם אנחנו @AI21Labs ביחד עם @nvidia יצרנו יום מ…

Come join us at #SFTechWeek! Sign up now: https://t.co/5BTA3L8hFB

it's actually crazy how much a man's incentives dictate his worldview; therefore, we should all etf maxx and have equity in everything so we can let true rationalism flower

>hmm openai says don't build features on our existing products because we'll obliterate you >they say instead build features for next-gen ai >but wait >next-gen products come >i built features for them >next-gen ai is now existing ai >they're gonna obliterate me >mfw

RT @sainingxie: Representation matters. Representation matters. Representation matters, even for generative models. We might've been tra…

my top complaint about @COLM_conf is it set the bar too high for future AI conferences p.s. if we didn't talk and you're in the bay let's still chat - thanks @adamfungi for 📸! https://t.co/1e4ck4vJta

excited to talk about self-taught optimizers 🛑 @COLM_conf this afternoon! was a really fun collaboration, and especially grateful to @adamfungi & @LesterMackey for being amazing mentors summer 2023 CE at MSR

RT @denny_zhou: Instead of using the terms “system 1/2,” for LLM reasoning, it is more appropriate to view reasoning as a continuous spectr…

RT @WilliamWangNLP: 📢 We are officially retiring the original 2018 WikiHow dataset [Koupaee and Wang, 2018], as requested by the copyright…

RT @shuyanzhxyc: 🦾#NeurIPS2024 Introducing Synatra, a data synthesis approach for web agents that can generate demonstrations for a wide ra…

RT @gneubig: We just released a paper Synatra, a method for generating synthetic data for web agents by: - Generating synthetic actions fr…

@HazanPrinceton (sanity checks on the formula collected in https://t.co/WqlV93ISr5)

@HazanPrinceton For that form, loss after t/2 steps of GD with step size 1 has a nice form in terms of the complementary error function. https://t.co/ClDuRK3im6

RT @rose_e_wang: AI has the potential to transform real-world domains. But can AI actually improve outcomes in live interactions? We condu…

Taking a quick break from COLM to promote this

RT @COLM_conf: Hello world! https://t.co/5AmTYrHEia

Impressive https://t.co/hyoynRZsr0

RT @srush_nlp: Really enjoyed this presentation at COLM. Such a dense set of experiments. https://t.co/kekuS7jecs https://t.co/Lm0YANQJxu