This episode continues our discussion with AI researcher Aleksa Gordić from DeepMind on understanding today’s most advanced AI systems.00.07 This episode builds on Episode 501.05 We start with GANs – Generative Adversarial Networks01.33 Solving the problem of stability, with higher resolution03.24 GANs are notoriously hard to train. They suffer from mode collapse03.45 Worse, the model might not learn anything, and the result is pure noise03.55 DC GANs introduced convolutional layers to stabilise them and enable higher resolution04.37 The technique of outpainting05.55 Generating text as well as images, and producing stories06.14 AI Dungeon06.28 From GANs to Diffusion models06.48 DDPM (De-noising diffusion probabilistic models) does for diffusion models what DC GANs did for GANs07.20 They are more stable, and don’t suffer from mode collapse07.30 They do have downsides. They are much more computation intensive08.24 What does the word diffusion mean in this context?08.40 It’s adopted from physics. It peels noise away from the image09.17 Isn’t that rewinding entropy?09.45 One application is making a photo taken in 1830 look like one taken yesterday09.58 Semantic Segmentation Masks convert bands of flat colour into realistic images of sky, earth, sea, etc10.35 Bounding boxes generate objects of a specified class from tiny inputs11.00 The images are not taken from previously seen images on the internet, but invented from scratch11.40 The model saw a lot of images during training, but during the creation process it does not refer back to them12.40 Failures are eliminated by amendments, as always with models like this12.55 Scott Alexander blogged about models producing images with wrong relationships, and how this was fixed within 3 months13.30 The failure modes get harder to find as the obvious ones are eliminated13.45 Even with 175 billion parameters, GPT-3 struggled to handle three digits in computation15.18 Are you often surprised by what the models do next?15.50 The research community is like a hive mind, and you never know where the next idea will come from16.40 Often the next thing comes from a couple of students at a university16.58 How Ian Goodfellow created the first GAN17.35 Are the older tribes described by Pedro Domingos (analogisers, evolutionists, Bayesians…) now obsolete?18.15 We should cultivate different approaches because you never know where they might lead19.15 Symbolic AI (aka Good Old Fashioned AI, or GOFAI) is still alive and kicking19.40 AlphaGo combined deep learning and GOFAI21.00 Doug Lennart is still persevering with Cyc, a purely GOFAI approach21.30 GOFAI models had no learning element. They can’t go beyond the humans whose expertise they encapsulate22.25 The now-famous move 37 in AlphaGo’s game two against Lee Sedol in 201623.40 Moravec’s paradox. Easy things are hard, and hard things are easy24.20 The combination of deep learning and symbolic AI has been long urged, and in fact is already happening24.40 Will models always demand more and more compute?25.10 The human brain has far more compute power than even our biggest systems today25.45 Sparse, or MoE (Mixture of Experts) systems are quite efficient26.00 We need more compute, better algorithms, and more efficiency26.55 Dedicated AI chips will help a lot with efficiency26.25 Cerebros claims that GPT-3 could be trained on a single chip27.50 Models can increasingly be trained for general purposes and then tweaked for particular tasks28.30 Some of the big new models are open access29.00 What else should people learn about with regard to advanced AI?29.20 Neural Radiance Fields (NERF) models30.40 Flamingo and Gato31.15 We have mostly discussed research in these episodes, rather than engineering