This prompted Gary Marcus to post "'Sales is All You Need' is Dead" garymarcus.substack.com/p/breaking-n...

This prompted Gary Marcus to post "'Sales is All You Need' is Dead" garymarcus.substack.com/p/breaking-n...

#aihype #ai

Most notably Richard Sutton (father of #reinforcementlearning) in his keynote said researchers should focus on inventing new AI that can learn from new information after it starts handling real-world tasks

that’s cool that you included Sejnowski! With all due respect, Hinton gets all the press these days because he’s more associated with the practical AI successes of deep learning. But Sejnowski was there back at the beginning!

So what do we see? In the last year performance went from abysmal to poor. The best agents can now complete just under 50% of the tasks. If the trend continues in the next year, we will see agents able to complete most of the tasks. My bet is that performance plateaus somewhere in the 70% range.

In a simulated small company, agents try to complete tasks like emailing, web searches, arranging meeting rooms, and analyzing spreadsheets. Its the only OBJECTIVE measure of how far agents are progressing.

the-agent-company.com#/leaderboard

I’m doing turkey meatballs lol

Extreme learning is using a neural network with a completely random hidden layer and only training the output layer, so its much faster - and this works! See Extreme learning machine: theory and applications GB Huang, QY Zhu, CK Siew  for details

GenAI needs to be complemented by other, arguably deper, AI forms like Causal Machine Learning, Neurosymbolic AI and World Models. What AI needs to really reach the potential is deeper types of AI. Interesting areas for future research include Causal Machine Learning and Neurosymbolic AI.

My POV is that GenAI is not going to change the world as much as the proponents think. Particularly, language models are going to continue to be limited by hallucinations because at their core they only predict next tokens based on probability. Language models will not, on their own, lead to AGI

It's reasonable to expect that both AI will change the world, and most of the current AI companies will fail and much of the current spending on AI will be a gigantic waste.

The opposing POV's in the article make sense by comparison to the dot-com bubble: The internet eventually changed the world, but not before the dot-com crash took down most of the first generation of internet companies.

I think these all miss the real problem: Investors are now throwing money primarily at generative AI and not investing enough in complementary AI forms like Causal Machine Learning, Neurosymbolic AI and World Models. When the crash comes, funding will be withdrawn from everything.

IMHO its clear LeCun made a significant contribution by actually getting convolutional NN to work (as even Marcus acknowldeges). At the same time, LeCun does seem to have a long track record of not crediting those whose shoulders he stood upon, and flip flopping on his positions about GenAI.

I did my PhD on extracellular fields with Christof Koch and we analyzed ephaptic effects - fields seem too weak to be functional. So on the substance of the debate I think I’m with KK. That’s the biophysics I found when I analyzed it back in ‘07. Has our understanding of the biophysics changed?

Hang in there! It’s amazing that the surge in AI interest has not led to more funding for research on real brains.

IMHO there is a key difference: banks need healthy balance sheets to keep lending, without which the economy grinds to a halt. Google, Microsoft and Nvidia can all loose $500B each and we'll still be able to use the internet and office tools. Am I right?