Kevin Patrick Murphy
@sirbayes
Research Scientist at Google DeepMind. Interested in Bayesian Machine Learning.
Great article.
Steven Pinker injects some facts and much-needed sanity in the debates around Harvard and American academia. nytimes.com/2025/05/23/opi…
The rise of tool-using LLMs is good news if you like to develop fast and reliable algorithms: the LLM converts raw text x into structured form x', calls your method f to get y'=f(x'), then converts it back to text y. We need more tools f!
Don’t just predict the mean of your clean data given your noisy data, predict the full distribution.
Distributional diffusion models with scoring rules at #icml25 Fewer, larger denoising steps using distributional losses! Wednesday 11am poster E-1910 arxiv.org/pdf/2502.02483 @agalashov @ValentinDeBort1 Guntupalli @zhouguangyao @sirbayes @ArnaudDoucet1
I don't know why AI doomers keep going on about "bio risk" (which is rate-limited by many physical constraints). AI agent powered cyber attacks to power grids, satellites, etc. seem more likely, since these are digital systems, but directly affect the physical world.
It’s funny how creating “radical abundance” has replaced “democratizing AI” as the new euphemism for getting rich in the tech sphere.
My colleague and former intern @liusiqi42 reminded me that we did RLFT for LMs almost 10 years ago - back then it was for an img2text model based on CNNs and RNNs. But same basic recipe - pre train with MLE then fine tune with PG. arxiv.org/abs/1612.00370
This is a very thought provoking interview with my former student. I do think AI personas (esp multimodal and real time) may be addictive and seem better than humans - but so is heroin (albeit heroin has less useful applications than AI).
AI will probably be better spouses and parents than us someday, by our own standards. Me and @danfaggella talked about what this might look like, how it'll likely feel at the time, and how hard it is to think clearly about these painful ideas. youtube.com/watch?v=XPpg89…
Well, it seems that the Elon / Trump bromance is finally over, as I predicted… 🍿
It’s delicious to see the in-fighting on the right has started before Trump has even taken office. How long will Ramaswamy and Elon last before falling out of favor with DJT? mediaite.com/politics/vivek…
I think it's quite misleading for the big labs to be promoting how well their VLMs work on pokemon, given how much (game-specific) manual annotation is required behind the scenes. Solving general tasks from pixel input is much harder than coding ("Moravec's revenge").
Yep that's exactly what i expected. They test VLMs on gameplay without scaffold: image in, action out. And they can't play. One small feedback to the authors: include a random agent (maybe best of 100 or so) as a baseline.
100%.
The full video of my Upper Bound 2025 talk about our research directions should be available at some point, but here are my slides: docs.google.com/presentation/d… And here are the notes I made while preparing, which are more extensive than what I had time to say:…
Does anyone know if ChatGPT keeps some kind of context or user profile across sessions? If i ask it to derive mathy things related to online Bayes, it often asks me if I want to see a low-rank version of it, or a Thompson sampling version. How does it know I care? Spooky.
I am pleased to announce a new version of my RL tutorial. Major update to the LLM chapter (eg DPO, GRPO, thinking), minor updates to the MARL and MBRL chapters and various sections (eg offline RL, DPG, etc). Enjoy! arxiv.org/abs/2412.05265
This was a great talk (*) on using (proper multi-turn) RL for training LLM agents to reason and use tools. Very bullish on this "Generative Agents" direction! (* Audio was very bad; fortunately brains are good at source separation :)
nice gemini summary/transcription of the talk from Friday :)
I dont know why singapore air is rated number 1 in world. their business class beds are much less comfortable than united/ polaris, because they are narrow and not straight. Food is good but not amazing. IMHO Emirates is best, then KLM & United (but grateful not economy class :)
I had a great time diving at @Wakatobi in Indonesia (although unfortunately I got an ear infection and had to skip the last couple of days). Tomorrow off to Singapore for #ICLR2025 (DM me if you want to meet).
I am pleased to announce that I have updated the online versions of my 2 textbooks (see probml.github.io/pml-book/): I fixed all issues listed on github, added some new references (esp on LLMs), and made a few other small tweaks.