🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

interesting blog post from Roy Rinberg on considerations for DP fine-tuning of LLMs++

The Conference on Digital Experimentation (CODE@MIT) is happy to provide free, open access to the Fireside Chat, "The Changing #Data Environment" this Saturday, October 19 at 11:45am ET. Tune in for the livestream or watch on demand. Dean Eckles will be joined by expert panelists: ⭐Malika Korganbekova

Please retweet. I need help getting to deaf & insensitive ears. NJDPB illegally terminated my chronically ill wife's insurance on Aug 31st for lack of her SSN on file. Her quality of life relies on a constant medical attention, including her neurologist appointment tomorrow.

Good question — anyone have a default stack?

You trained your model and then realized some of your data was bad/copyrighted - what can you do? Fear not, we have (some) answers! Check out this new work on *Un*learning by me + my collaborators. Sam Park Kristian Georgiev Shivam Garg Andrew Ilyas Aleksander Madry Seth Neel

Really excited about this new work using data attribution to improve unlearning!

We also wrote a more informal blog post about our new results 👇🏼

I've been thinking about Privacy & LLMs work for 2025 - here are 5 research directions and some key papers on privacy/memorization to get started: 🧵

📢📢 Happy to share that our paper on unlearning evaluations has been accepted to ICLR 2025 🇸🇬 📜: arxiv.org/abs/2406.17216 Thanks to my great co-authors Seth Neel Gautam Kamath Jimmy Di Ayush Sekhari @yiweilu

New paper alert!🚨 What do LLM reasoning, diffusions, & jailbreaks have in common? 🤔 All exhibit critical windows📈--a sudden formation of distinctive features during sampling, e.g. correctness or toxicity. We present a unifying theory of critical windows for diffusion & LLMs.

Science is the engine behind American prosperity. Government funded research led to many of the breakthroughs we use today in AI. AI may well supercharge scientific productivity and the way to take advantage of it is to increase investment in scientific research. Unfortunately

A guess: The uncertainty in science funding in the US will -already- have a huge affect on PhD admissions for engineering schools this cycle, even before funding reductions hit (if they do). I expect historically low class sizes in CS PhD programs. Not good for US AI leadership.

🥁Introducing Gemini 2.5, our most intelligent model with impressive capabilities in advanced reasoning and coding. Now integrating thinking capabilities, 2.5 Pro Experimental is our most performant Gemini model yet. It’s #1 on lmarena.ai (formerly lmsys.org) leaderboard. 🥇

I recently joined @Nudge as medical director. We're building a world-changing device, with the goal of precisely and non-invasively modulating brain function to improve the lives of as many people as possible.