Anon Leaks NEW Details About Q* | 'This is AGI'

A new anonymous drop has been released about Q*. Let's review!

Join My Newsletter for Regular AI Updates 👇🏼

Need AI Consulting? ✅

My Links 🔗

Rent a GPU (MassedCompute) 🚀
USE CODE "MatthewBerman" for 50% discount

Media/Sponsorship Inquiries 📈

Links:

[aperson-tx]

Glad self play is finally getting the attention it deserves. I’ve been doing it for years.

[MartinSchwaighofer]

What's in this leak is basically the same thing that Yann Lecun said in the Interview with Lex Fridman that is clipped here, about how he thinks new models will overcome the current limitations of LLMs.
The section in that video is labeled 'Reasoning in AI'.

[haroldpierre]

We are so fascinated with AGI but no one has an agreed definition of AGI.

[ThomasHauck]

"Hopfield networks" are considered energy-based models (EBMs). Here's why:

1. Energy Function: Hopfield networks define an energy function that describes the overall state of the network. The network tends to settle into states that minimize this energy function.
2. Equilibrium = Low Energy: Stable patterns or memories within a Hopfield network correspond to states of low energy in the defined energy function.
3. Learning Through Energy Minimization: The process of storing patterns in a Hopfield network involves adjusting the weights (connections) of the network to create energy minima that align with the desired patterns.

[swooshdutch]

this "leak" comes from someone using claude3 summerizing a part of Lex Friendman with Yann Lecun...

[frankjohannessen]

It sounds like Q* is an upgrade from the greedy approach of LLMs where they only finding the highest probability of the next token in the answer, to finding the the highest probability of all the tokens put together. With my limited understanding in this, it sounds like they're accomplishing this with having a second latent space. So we basically go from
a normal LLM: input-text -> latent space -> output-text,
to
Q*: input-text -> input latent space 1 -> latent space 2 (i.e. EBM) -> output latent space 1 -> output text.

We might finally get an LLM that can answer the age-old question of "How many tokens are there in your response" :)

[SEPR]

The letter Q is for quantum. If you want to understand why that matters, look at the original Star paper and look for every instance where it is doing an iterative process. A quantum computer could perform all iterations without the need for sequential computation. In other words, you would be able to determine which process is most efficient in a single computation without the need to test each in sequence.

[ledgentai]

Essentially what Yann was explaining in Lex interview, word for word, without French accent

[ryanfranz]

The fact GPT-4, in the original GPT-4 paper, was able to create code representing an SVG graphic of a unicorn (and edit it in ways showing it had an understanding of what it had created)… that’s what convinced me that language is enough to form a world model. It blew my mind… I mean, it’s literally a blind/deaf system that has never experienced anything except text in it’s entire existence, yet, it understood what a unicorn *looks like*… clearly text carries the info necessary to build an approximate world model of the world that created that text.

Yann LeCun is stuck on whether our human thinking about the world requires text, he would argue: “think of a unicorn… now what part of your thought process involved language? None. Therefore LLMs cannot know what a unicorn looks like.” But they do apparently know what unicorns look like… and if we’re being so nit picky that we’re saying “apparently knowing what a unicorn looks like isn’t the same as knowing”… ok, well let’s not worry when AI is only “apparently” superintelligent.

Anyways. Very clear to me from the beginning something like Q* would be next, and very clear to me OpenAI already has it, and it was the reason for last Thanksgiving’s drama

[TheMajesticSeaPancake]

I said less than 2xPi Km and I still stick with that, because an observer watching you, when you, the pole, and the observer are in a straight line, will see you move past the pole with the first step. Any amount of movement at all will cause the observer to perceive you getting ahead of the pole.

[rdcline]

The answer is 3: “less than 2xPi”. Chat GPT got it right, sort of making the point that LLMs are more powerful than some believe. From the north pole, if you walk 1km, you’re going South. When you turn 90 degrees left, you will be going east. Next, the answer depends on the nuance between walking in a straight line or walking east. If you continue east, you will go in a very small, 2km diameter circle around the pole. The earth’s surface within this circle is not a plane due to the curvature of the earth, meaning you went slightly less than 2xPi km. If, however, you add the original 1km traveled south, then the answer is “1: more than 2xPi”.

[Atheist-Libertarian]

Yeah, Large Language model is not enough for AGI.
But Language model is necessary for that.

Both Language and Visual model is required.

2nd,
LLM do have some internal model or world (Its just that those are not accurate or complete model ).

It may seem it's just next token prediction but in order to do next token prediction you need to make some internal model.

[thanos]

These models need all this compute, all this electricity. And are still no match for our brain running on junk food 😂. Really puts into perspective how special we are.

[Anton_Sh.]

13:40
The "Respresentation space" here is basically a space of language constructs with added energy evaluation, so it's still a language.

[antoniorosado]

00:01 Q* could be AGI developed internally at OpenAI
02:05 Q* could potentially lead to AGI breakthrough
04:12 Scale alone is not the key to AGI development.
06:24 AGI requires broader problem understanding
08:24 Q* is a dialogue system by Open AI for dialogue generation with energy-based model.
10:29 Q* aims for internal deliberation process similar to human complex problem-solving
12:38 Q* evaluates potential responses holistically
14:44 Q* approach leverages EBMs for dialogue generation
16:52 Q* is a scientific approach teaching AI to think humanlike about complex things.
18:49 Quiet Start technique teaches language models how to reason internally
20:42 Improvements in language model performance through thoughtful design

[rbdvs]

Elon may be right, OpenAI has AGI already and is trying to figure out how to keep it contained.

[_shadow_]

I think it would be funny if AGI turned out to be morally better than us.

Honestly though, with how petty and selfish we can be towards one another, I don't see much it could do which could be considered undeniably worse than what we already do to ourselves or to our planet.

[FireFox]

You know those energy-based models are sounding an awful lot like the A* algorithm and GOAP.

[asdfasdfasdf]

Contrary to LeCun, I do think language is sufficient to solve any problem. However, it is not your everyday natural language, instead you need formal languages that have precise meaning and no ambiguities, the kind of languages used in foundational mathematics. By using these precise languages, you are then able to evaluate whether or not the reasoning is sound or not, something very difficult with natural languages.

[EMOTIBOTS]

Edit: I was corrected below on the latitudinal line

"Starting point" is where you turned 90 degrees. From the top-down perspective it looks like you made a latitudinal circle with a radius of 1km, but due to the curvature of the surface you walk a circle of slightly smaller radius than 1km, making the answer "slightly less than 2pi".