dimmuborgir

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.

From the paper:

"A single run of the DGM on SWE-bench...takes about 2 weeks and incurs significant API costs." ($22,000)

Yup, they are called GCCs (Global Capability Centers).

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.

Dark mode finally!!! This was the only bummer for me all these years. Not anymore!

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.

I think her target audience is extremely broad. From teenagers to middle-aged people. Plus her latest album was marketed aggressively.

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.

AI is bad at music also. Even the state of the art transformer models can't produce more than a few seconds of coherent melodic phrases.

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.