Liquid AI's Ramin Hasani on liquid neural networks, AI advancement, the race to AGI & more! | E1928

[@zacharydaniels3186]

I have been waiting for an update on this company for about a year. Thanks for having them on!

[@ashh3051]

I think Jason got a bit confused at 13:20 and the interviewer didn’t do a good job of catching this. I’m guessing that Jason was thinking that current neural networks are like “human brains” and the liquid neural network was based on a worm brain. This is a mistake as current neural networks are not like human brains. I think the liquid guy would claim that his architecture is closer to the human brain than traditional NNs as his design is more based on biology.

[@SteveMayzak]

Hilarious that our current neural networks are named as such. I guess that name was take so Liquid it is! Seriously cool

[@timeTegus]

@@SteveMayzak most neural networks are more like human brains. but trasformers are a bit of a exeption. they seem more distant. but they also seem to scale very good.

[@SB324]

Catch that Joscha Bach name drop at the end! If Joscha is on the team, they’re on an entirely different level

[@circlelabs]

Wow, insane tech

[@DonaldHarvey-yi2tg]

0:22

[@MichaelCoker-st3cd]

Why didn't the liquid AI driving test also contain the same amount of noise as the standard test? It seems slightly unfair and the liquid AI attention map seemed to just have attention *everywhere*. I'd love to see a test where they throw a ball out in front of the vehicle as well. I'm still super interested in this tech to see how it works and the potential for lower power/processing needs.

[@KCM25NJL]

If you watch the actual staged demonstration video, they do go on to add noise to the LNN demo and it does prove to be largely unaffected.

[@JetLagRecords]

This Week in Startups, Your videos always make me happy, so I subscribed!

[@gridvid]

Awesome. Liquid Networks are awesome. Keep it up

[@Renaudpc]

I love this approach. Why try to create the human brain when you can create an earliest version of it, learn the ins and out and evolve.

[@Renaudpc]

I love this.

[@SB324]

Glad you had Ramin on!

[@Mheme-f7n]

Ramin Hasani!!!!!!!!!!!!

[@droh_it]

Technical question for the founder. Can you do multivariate time series anomaly detection with liquid neural networks?

[@roycedsouza2963]

Why don’t you just use tabular analysis?

[@hamandchees3]

Wish jcal asked more about how liquid neural networks work in practice-- how are they pretrained, how sample efficient are they, do they have published scaling laws, etc.-- rather than diverge off into hot topics.

[@petratilling2521]

They should go for the high-end personal market as well. I want a whole-house self contained model that comes with the ability to plug in verticals and learns as we go. I really hope they blow past ALL current architectures.

[@nicktrish2008]

Please give Elon this guy’s number?

[@danielscott2062]

No. I've seen his code. Spent too much time fixing it... His Liquid CNC paper was interesting but ultimately couldn't be reproduced. The best minds create and publish clean code. Sorry Ramin.

[@ps3301]

Since Elon is a first principle analysis kind of guy, if he is presented the liquid neural network which is much better in self drive and consumes a lot less power, Elon will be adopting it if it consumes 10 times less power. Self drive chip consumes a lot of energy.

[@talentsherpa]

This week in startups us my compass. Thanks

[@游淯聖-j5p]

Can someone explain to me what this "liquid" stuff is? I dont see the special of this, just a varient of ODE and even worst experiment result, in the end is this just another RNN like model?

[@petratilling2521]

This is the way.

[@petratilling2521]

This idea that only “bad” jobs will disappear is silly. The “best” jobs will probably be the first to go.

[@PeterNooteboom]

This almost feel to good to be true. But I suppose time will tell.

[@abdulshabazz8597]

This technology is promising. The reason transformer models have so many parameters is for controlled tunability which can always be controlled due to the limited way hyper parameters are employed in transformer models.. What is not advised are proposed advanced architectures with builtin runaway geometric cross-learning capabilities which this model is not.

[@zacharydaniels3186]

It strikes me that this will be the sort of "brain stem" of AI that interacts with the real world while the "right brain" creativity will be transformers. My question is: where does the logical "left brain" architecture lie? Mamba?

[@innomind]

Is he Persian?

[@chrisredlich9086]

Even if AI comes up with the most elegant solutions for energy or other issues, it’s impossible to say how that will be used. If it is politically inconvenient, I guess it will be shut down under the table.

[@hobocraft0]

Finally, some good old fashioned bio mimicry. Who is the reference? God.

[@zacharydaniels3186]

Excellent point.

[@juliusapriadi]

...and as usual, no word about AI Safety. Let's continue the race to AGI with better and better architecture, and then act surprised if we accidentally get killed, since no one bothered waiting for alignment to be solved.

[@rw9207]

It was almost funny watching Jason misunderstand and fail to grasp, almost everything Ramin said. His thinking is so rigid, it was like watching a Grandpa trying to play portal 2.

[@ashh3051]

This guy seems like a quack to me :(

[@SB324]

That’s a good sign, we need a new paradigm

[@kuo-yingwang2273]

Lots and lots of embedded add in this channel.

[@freerockneverdrop1236]

hoax