Nano Banana for me. After the initial wow phase it's meh now. Randomly refuses to adhere to the prompt. Randomly makes unexpected changes. Randomly triggers censorship filter. Randomly returns the image as is without making any changes.
The result announcement blog post sounded too hypey and gave the impression that he changed his tone but for the last one year he has been consistently saying that LLM+"discreet program search" is the way to go. All the top scoring submissions before o3 had followed the same bruteforce strategy. Even o3 is more or less doing the same bruteforce under the hood, maybe not discreet.
We can extend this phenomenon to all of knowledge. You ask ChatGPT something and it gives you some very generic 10-point blogspam-esque answer. You immediately think ChatGPT is intelligent and that it has understood your particular question and that it has given you a tailor-made answer.
His book "The Algebraic Mind" goes into great detail about connectionism (neural networks), symbolic systems, limits of connectionism and proposals to integrate neural networks with symbolic systems (hybrid systems).
The "deep learning alone is enough" camp (especially LeCun) has abused him for years but now slowly coming to the realization that we need to feed neural networks with explicit inductive biases to attain AGI which is exactly what Marcus has been saying since the 90s. LeCun, for some reason, refuses to call these explicit biases as symbols and that's the only disagreement between Marcus and LeCun these days.
Those models are not trained on short loops. They are trained on whole songs just like image generation models are trained on whole images. And yet they struggle to repeat sections, modulate to a different key, create bridges, intros and outros. After a few seconds of hallucinating a melodic line they simply abandon the idea and migrate to another one. There is no global structure whatsoever.
This one suffers from the same problem that previous audio generation methods had. It correctly mimics the piano timbre but there is no global structure in the generated melodic lines. There is some style imitation but no melodic coherence.
Lots of weird artifacts which are very hard to fix.
The argument that users can now generate professional grade art by bypassing artists entirely feels so strange. I have access to Dall-E. To generate images without artifacts, you have to do one of these: a) Do a lot of cherry-picking which can be expensive. b) Prompt should be about an abstract concept which can "tolerate" any number of artifacts. c) Prompt should be about a common/generic concept that you have already seen a lot of times on the internet.
I think the biggest use case of Dall-E will be in removing creative block for artists.
Who gets to decide what is the "point" of music? Music is a twenty billion dollar industry. An AI system that can spit out highly "realistic" and "pleasing" music can change the music industry as we know it.
Such coherent and pleasing melodic phrases in the style of Avril Lavigne. I thought it could be copying wholesale from a song unknown to me. Nope. Shazam doesn't get it.
This can revolutionize song writing/composition/production and soon music listening/consumption.