My statement on the death of George Floyd:

NLP meets computer vision.

You can now ask philosophy questions to GPT-3, a very large deep learning language model. The results are impressive and entertaining. philosopherai.com

Backpropogation has allowed us to efficiently train artificial neural networks. How biological neural networks learn remains a mystery, but an auto-encoder architecture offers a plausible approximation of backpropagation in our brains. nature.com/articles/s4158…

Further great work from Ouyang et al. An excellent example of what computer vision models can accomplish in cardiology.

This is an important and impressive step towards developing neural nets that understand the 'intuitive physics' of our world. The model learns concepts such as object permanence, occlusion, etc, like a child in early stages of development. #AGI nature.com/articles/s4156…

Neural nets speeding up... neural nets.

The greatest tragedy in the current practice of cardiology is the lack of access to cardiac CT. Too many people have a cardiac arrest as their first sign that something is wrong.

As more foundational and general AI models outperform narrow models, the value alignment problem will take center stage. Let us hope we can train it to be a better steward of sentient life than we have been.

As neural nets develop a better “theory of mind,” the ability to infer the unobserved mental state of others, many have worried it may use this capability to deceive us, but I believe this is also essential for it to empathize with us and better align with human values.

Neural nets now are solving previously unsolved math problems. It seems that the symbolic reasoning expert system versus connectionist AI divide of the past is ending, with neural nets showing that they can do it all. This makes sense as our brains are capable of logic.

The era of large multimodal video transformer models is upon us. I suspect an understanding of video should be sufficient for a very powerful world model. We are moving closer and closer to models with extremely general capabilities.

We made a new AI-CAC algorithm to automatically quantify cardiovascular risk from routine chest CTs using the unprecedented scale of the VA National Healthcare System! This is the first step to build a scalable prevention program for veterans. Pre-print: arxiv.org/abs/2409.09968

AI will exceed human capabilities, but it is still worth investing in our own critical thinking and knowledge. Shared understanding deepens the dialogue—like Einstein talking to someone who has studied physics. You will get more out of the interaction the more you know.

AI-CAC source code and NEJM AI publication are now online! We hope this work serves as a step toward earlier CAD detection by analyzing scans that patients already have had in routine care! ai.nejm.org/doi/10.1056/AI…

AI remains under hyped. In what other hype cycle have the goal posts of capability been moved so much?

Whether you’re a scientist, policymaker, or simply a concerned citizen, you should hear what Geoffrey Hinton, the “Godfather of AI”, has to say about the challenges humanity will face in our lifetimes. youtube.com/watch?v=giT0yt…

Choose love. At least I hope AI will.