Agreed. Level 5 autonomy is the only way to go. Will need expensive LiDAR etc for now. Can’t expect humans to suddenly intervene if the level 3/4 system messes up. This cavalier approach from Tesla may kill many innocent bystanders, and may trigger legislation which could kill…

“Facebook’s online help page says that Meta AI will join a group conversation … if someone “asks a question in a post and no one responds within an hour.” So Llama is just a way to make fake activity on FB groups so people check back more often to avoid AI generated FOMO? Will…

arxiv.org/abs/2404.01261
I currently don’t trust LLMs enough to find them useful for my work. Science needs to be accurate. OTOH we need reliable QA and summarization tools to help with the deluge of papers, so the quest goes on…

Good article from WAPo Gerrit De Vynck 🦭 . Serving costs are coming down, but training costs are going up, and the reliability is still low. Image, music and story generation is fun, but has copyright issues. Plus Why automate the fun stuff?? Automate tedious tasks reliably if you want…

Internal Gemini did quite well on this, but still made errors, and is not yet public. So I did it by hand using this useful library
github.com/rrtucci/texnn/…

This is a great podcast! Two really smart and articulate people having a super interesting conversation on AI. I have some issues with what Dario said, but it’s worth listening to. (Plus other shows from ⁦Ezra Klein⁩ :) nytimes.com/2024/04/12/opi…

From the excellent new computer vision textbook by Isola, Torralba, Freeman.

This looks cool - a Jax-inspired PyTorch library for Bayesian deep Learning. It even has EKF (although only diagonal, not our low rank trick …)

don’t you think it’s wildly irresponsible to use the general public - not just Tesla owners who choose this feature - as part of a data flywheel? I love my Tesla S, but FSD needs to be much more robust before releasing in the wild. Humans can’t intervene quickly enough in case…

I assume the purpose of the focus on 'ML for social impact' is to get URMs interested in the field (it would be good to hear from the NeurIPS people who created this...), but I think in reality this track will be used by already high-achieving students (not URMs) to further pad…

This is my semi-periodic request to VLM gurus: when will I be able to upload an image such as the one below and have {ChatGPT, Gemini, Whatever} give me back tikz code that reconstructs the figure? I would pay >$20/month for that! (David Dohan knows :)

Totally agree. Unfortunately for me, the golden hours, where Deepmind work hours in California overlap with work hours in Europe, are during my morning, so my day usually starts with meetings and messages :( Context switching makes it hard to do anything creative or technical.

'NIPS' lives on! Spotted on Google search (and bing.com, for that matter).

I heard an interview with Haidt (on hard fork) and he is very persuasive in his case that social media (but not phones in general, and not the internet) are bad for kids. And he does cite RCTs and interventional experiments in his book, contrary to what this review says, But…

Stanley's probability textbook (probability4datascience.com) is fantastic, so I imagine this tutorial on diffusion is good as well. However, I am no longer of a fan of presenting the discrete time version - the flow matching interpretation seems more elegant.

I'm delighted to present that our AutoBNN library is out. We show how to define a meaningful prior over functions, based on sums and products of base GP kernels, and convert it to a mathematically equivalent (in the limit) neural net architecture. This gives strong OOD…

Indeed! A much needed book.

It’s shocking to me that some people want to attribute feelings and consciousness to AGI but not to our closest animal relatives.