Most of these signatories have a distorted view of what is coming next with AI. The distortion is due to their inexperience, naïveté on how difficult the next steps in AI will be, wild overestimates of their employer's lead and their ability to make fast progress, and financial incentives to hinder open source AI platforms (since almost all of them work for providers of proprietary AI systems). As you and I know, generations after generations of AI researchers have made the same mistake: thinking that human-level AI is "just around the corner" because of some new paradigm they are working on. They have consistently underestimated how difficult it is because they have an incentive to dismiss the limitations of their favorite paradigm as mere "engineering problems," and they can't foresee the obstacles they'll bump into before bumping onto them. We are making progress towards human-level AI, but we're still far from it. Some of us may have vague roadmaps for how to get there. But no one has a credible blueprint, let alone a demonstrable prototype. Before such a prototype exists (perhaps with the learning abilities of a house cat), regulating AI R&D because of a fear of existential risk is highly premature. There is nothing wrong with sensible legislation to regulate the *deployment* of AI application. But legislation that hinders open research and open source AI platforms, and that makes open source AI developers liable for what people do with their code is extremely regressive.

RT @TheBrianMcManus: I try my best to not talk about Irish politics, but getting sent things like this by my Garda friends just angers me t…

Emergent in-context learning with Transformers is exciting! But what is necessary to make neural nets implement general-purpose in-context learning? 2^14 tasks, a large model memory, and initial memorization to aid generalization. Full paper arxiv.org/abs/2212.04458 🧵👇(1/9)

Just watched a great interview with Monica Daley @birdBiomechy , super interesting. youtube.com/watch?v=xnjeVdLH…

Next week I'm going to be joining @3M as a Senior Research Scientist for AI in dental applications. Can't wait, sad to be leaving @VisualAIPeople after 4 great years.

One difference between experienced ML researchers with those new to the field is the amount of visualization that one does including visualizing training data, augmentations, parameters, gradients, predictions, etc. A codebase with only end metric plots almost always means bugs.

Looks to me like stage 2) argument: x.com/woj_zaremba/status/151…

This is unbelievably based

Has anyone got any advice/links on optimizing v,v large (2gb ) TF vision models? - for inference/training

What's the differences among ... Latent space, feature space, embedding space, representation space, latent feature, feature embedding, latent representation, embedding representation, latent embedding, and feature representation? 🤔

AI isn’t “hijacking” art history, it is providing the tools it needs to scale! The first automobile was not faster than a horse and carriage, but it had the attributes that allow it to scale.

#COP26 has just begun while I started working in the renewable industry today. As a welcome gift, I'd really appreciate if all the world leaders agreed to dump 30 years of work on my desk.

Big fan of @ramin_m_h talk on the paper 'Liquid Neural Netwoks' about a 'new class of time-continuous recurrent neural network models'. Really good visualisation. youtube.com/watch?v=IlliqYiR…

Does my GAN's loss go down every day? No. But does it try its hardest to make its image quality go up over time? Also no

The ability to wrap arbitrary python/NumPy operations within a Tensorflow graph - and save - then serve it, is a massively underrated benefit of Tensorflow. Main changes include changing from python lists to TensorArrays.

So many exciting new frontiers in ML, it's hard to give a short list, particularly in new application areas (e.g. in the physical and biological sciences). But the Big Question is: "How could machines learn as efficiently as humans and animals?" This requires new paradigms.

We do not have an answer to that question, and the gap to bridge is enormous (how can people learn to drive a car in 20h of practice?) Decisive advances towards an answer will mark a new era in AI. That's why I work on self-supervised learning. It's our best shot at the moment.

those who attack space maybe don’t realize that space represents hope for so many people

The conversations around Falcon and the Winter Soldier being “too political” when it’s just detailing Black people’s existence in the US shows us that the “no politics at work” policies will reprimand Black people for just sharing our experiences & existence

New preprint: "Impact Invariant Control with Applications to Bipedal Locomotion," by William Yang. During impacts, robots undergo large and rapid changes in velocity. State estimation, and thus control, in these periods are incredibly difficult. (1/2) arxiv.org/abs/2103.06907