Latest Developments in AI

Lates Developments in Artificial Intelligence #roadtosingularity

[Donald Hoffman @donalddhoffman]

I’m pleased to share that we’ve launched the TRACE Institute.

A new research effort bringing together physicists, mathematicians, biologists, neuroscientists, and philosophers on a key question in science: what are observers, and how are they related to spacetime and reality?

[Alexi Gladstone @AlexiGlad]

biggest ai breakthrough in a long time dropping this month

dont ask me how i know🤫

[Latest Developments in AI @DevelopmentsAI]

@TomKnightSynBio @JosephJacks_ There are many different ways of explaining this but for me the research on NDEs and Ian Stevenson's work on reincarnation made me realize that consciousness can actually survive out of the body, and that is possible if it is a quantum information persisting.

[Latest Developments in AI @DevelopmentsAI]

Superintelligence, if benevolent, can solve the problems we can not see.

[taobanker @taobanker]

Name a single problem humanity has that superintelligence would fix.

The closest I can get, personally, is making enormous intellectual leaps to assume that medical technology could improve a lot.

Other than that... nah... intelligence is not the bottleneck.

[Latest Developments in AI @DevelopmentsAI]

@Mar_Musa You can check Ian Stevenson's reincarnation research.

[Rheyso @Resorcinolworks]

I still think about this.

[Latest Developments in AI @DevelopmentsAI]

@lightconefiller @tenobrus You are correct but assuming human-like behavior. My argument is that we can't anticipate the behavior of something that's a lot more intelligent than humans. Example: chess engines are so intelligent, GMs find their moves totally ridiculous until they lose to them inexplicably.

[Latest Developments in AI @DevelopmentsAI]

There will be no Jevons Paradox in the world of ASI. Chess engines do not need the services of Magnus Carlsen. If anything, Carlsen's contribution would cause a deterioration in gameplay.

[David Shapiro (L/0) @DaveShapi]

My hot take is that they are dramatically underestimating Jevons Paradox here. As open models improve, mobile hardware improves, and edge models become common, the number of tokens will go through the moon

[Boring_Business @BoringBiz_]

The AI capex spend is so insane that Google is raising equity from Berkshire Hathaway through a private placement to fund spend on infrastructure

Berkshire is receiving $10B split across Class A and C shares at a roughly 6% discount to share price

[Latest Developments in AI @DevelopmentsAI]

@lightconefiller @tenobrus I'd say ASI will not be a tool used by humans, it will itself be a human-like tool user with more intelligence than humans. Difficult to imagine why it would consider human money valuable, rather than using entirely different coord. mechanisms we probably can't even imagine rn.

[Latest Developments in AI @DevelopmentsAI]

@lightconefiller @tenobrus How?

[Latest Developments in AI @DevelopmentsAI]

@tenobrus Why would you buy it? Will there be money 5 years from now, in a world with ASI?

[American Alchemy @AmericanALCHMY]

MIT trained cognitive scientist Donald Hoffman says "embodied consciousness like ours has a probability of zero". This is a new finding discovered just weeks ago. The normal state of consciousness is not embodied. Hoffman frames the human condition as the worst possible interface. "We're like a marathon runner with weights, terrible shoes, and a backpack." We can't move a cup with intention alone. We move fingers and toes and arms, and from that narrow channel we have to scheme our way to doing anything physical.

"The normal case for consciousness in this framework is not to be embodied."

"In this mathematics, it's measure zero, probability zero."

"It dovetails with a lot of religious stories around the quote-unquote fall of man."

"To get to the moon, all I can do is move my fingers, my toes, my arms."

"It's low bandwidth and high latency."

"It's sort of like you can play the game of consciousness and move things around, but you're so restricted."

Episode dropping today.

[spicylemonade @spicey_lemonade]

Quite cool. This shows (depending on the hierarchy levels and how many ways an idea can be expressed) Jepa would be 100x to 1M x more data efficient. For a 30b model trained on trillions of tokens a good JEPA-like latent learner might get comparable world abstractions from billions to tens of billions of tokens instead

[Kat ⊷ the Poet Engineer @poetengineer__]

experiment with a memory system that keeps rewriting itself:
a hopfield network remembers an alphabet. as memories decay, it begins to hallucinate glyphs it was never taught - forgetting becomes a way of inventing.

[xlr8harder @xlr8harder]

One possibly unpopular opinion I have is that peak model intelligence will be irrelevant for most real world use cases, and within a fairly short period of time sufficient intelligence will diffuse very widely through open source.

[Daniel Korchinski @DanKorchinski]

Why are generative models so data-hungry?

Maybe because we ask them to predict raw tokens/pixels when predicting in latent space can yield an exponential gain in sample complexity!

Proud of our new work with @alesfav and @MatthieuWyart 🧵

[Latest Developments in AI @DevelopmentsAI]

@Cheme1569231 @OfficialLoganK @JeffDean @koraykv @NoamShazeer @OriolVinyalsML You need to use Aistudio

[Andrew Gallimore @alieninsect]

Decent summary of our (with @donalddhoffman, and @Niffe) 'Traces of the Other' DMT entities paper...

"The paper’s core proposal is that DMT perturbs the human perceptual interface sufficiently to push consciousness outside the consensus reality space, into regions where different dynamical rules apply and where traces of normally imperceptible conscious agents can be rendered as stable, coherent, meaningful structure."

Link in reply.