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

Oh snap, here we go

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'.

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

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

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

"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.

Also @Matthew, I may be mistaken but I think you misunderstand Yann's (badly worded) question. The question only makes sense if you consider the starting point to be the place AFTER you have walked 1km from the North Pole. The question is teasing out the mental picture that when you walk the initial 1km, you are doing so on a curved ball, so the radius of the circle around which you walk is actually slightly less than 1km and the answer is "less than 2 x pi km"

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.

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" :)

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.

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

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

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.

“No one has seen AGI”, Sam Altman a week ago. He told Lex Fridman it does not exist. These rumors are cherry picking quotes from people.

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”.

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

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

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.

Doesn't matter how far closedAI arrive with their closed models. Open source will overtake them with less computation needed using techniques like evolutionary optimization to merge existing foundation open source models, automatic formalization, chain-of-thoughts trough prompting, tree of search trough prompting, etc... Too many interesting ideas and too many talented people working in those ideas to compete against all of them.

Weird problem. I would walk around the earth and eventually "pass" the starting point in the sense that I'd see it on either my left or right side. I interpret "pass" as in "On my way to work, I'm passing the shopping centre on my left" - obviously I'm not driving THROUGH the shopping centre.

There's been some recent articles about researchers giving these models an "inner monologue" that improved it's performance. The articles referred to this technique as QuietSTaR. Is the similarity of the name to Q* just a coincidence? Seems unlikely. A search for AI and inner monologue will turn up the articles.

What are you using to place the green circle around your cursor? Readers want to know! Thank you!

The first rule of Q-star is that we don't talk about Q-star

Here is a solution to Yann LeCun's question: 1. Standing on the north pole, any direction will be directly south. 2. While walking directly south and then turning 90 degrees to the left I will be walking east. 3. A naive thought here might be that I will keep moving east from now on, but that would require me to walk in a circle with a 1km radius. Its easy to see why this is wrong if you make a mental model where you only move an inch south before turning to the left instead of 1km 4. As you walk now your heading will gradually turn towards south until you are headed directly south at the equator (this is because you are walking in a straight line on a sphere, and not a plane) 5. Now you will gradually be heading towards east again 6. Once you're 1km from the south pole you will be heading straight east again 7. From here on your heading will gradually turn from east to north. 8. Once you reach the equator again you will be heading directly north 9. From here on you will gradually be heading towards east, until your heading is once again directly east when you're 1km from the north pole. 10. Since we get all the way to the south pole and then are only 1km from the north pole again I don't think it's correct to say "I never came close to my starting point". 11. The correct answer is: 1. (way,way,way) more than 2xPi km

Welcome to life in "Exponential Time". AI is compound change with change moving faster..... and faster..... and faster..... and faster.......

Matthew Berman, This is fantastic! I subscribed right away!

What are you using to place the green circle around your cursor? Readers want to know! Thank you! :)

I think there's a fundamental misunderstanding when someone says "All LLMs do is predict the next word". Take, "The cat sat on the ___". It's very easy to predict "cat". But what about the word after "cat". No predictive algorithm can have any idea what comes after cat, any more than any human would. The only way you can predict more than a handful of words, is if you UNDERSTAND what is being asked. Let alone, compose an entire book or write pages and pages of text on a given subject. That is impossible if you are simply predicting the next word in a sentence. However LLMs CAN do this because they genuinely do "understand". It would be impossible for them to do the things they do, without them understanding. Understanding of the text is clearly an emergent property of vast neural networks, just as it is a property of our own, human, vast neural nets.

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".

Yann LeCun’s problem is essentially describing walking around the earth in a circle with radius of length 1km. You would come back to your starting point at exactly 2pi km. When he says starting point, he’s referring to the starting point after you walked the 1km. He clarified with a picture later on

So ChatGPT does not choose their responses from a range of choices? Is the response not chosen as a whole at all? Does chatGPT only chooses one token after the other until som "response complete" criteria is met?

The puzzle doesn't specify that you have to walk in a straight line on the last part, but I do want to point out that you can't walk in a straight line following lattitudinal lines on a sphere except at the equator, just the same as a spacecraft in orbit.

Imagine holding a ball in you hand. Mark a point on it as a reference as a North Pole. Starting here move 1 inch away from this point along the surface of the ball in any direction. Make a 90 degree turn. Now maintaining your distance from north pole continue along the surface of the ball.

Thanks for all your coverage of these topics, you're really saving people a lot of man hours keeping up to date on this stuff

A relationship between "economically valuable tasks" and "intelligence" in determining AGI is a value judgement. It does not define intelligence, though from a business or economic standpoint, it is a very valuable asset.

Berman's videos are essential to our on-going understanding of AI. Thanks, MB.

I think the good answer is more than 2pi actually. You start from the north pole, walking on any direction for 1km will put you on the curvature of earth (let’s say it’s a perfect sphere which is not the case in reality). Now you turn left and walk until you come back to your initial position. This is possible because walking in a straight line on a sphere will draw a circle. In our case we draw a circle of radius 1km. So we walked 1km + 2pi*1km = 1+2pi km

1. You verbalized the problem internally. 2. You reasoned the problem through verbalization. 3. You communicated the solution verbally. 4. You only understood what the problem was asking at all because of language.

This is insane. We are nowhere close to solving the AI misalignment issue or to having measures in place to prevent an AGI to rapidly self-optimize to incredible levels of intelligence. This is the kind of nightmare scenario AI safety experts have been warning about for years. If an AGI is released into the world now there will be no stopping it.

If the starting point is when turning 90 degrees, the answer is 1. You will bw walking around a circle, and to pass the starting point you walk more than 2xpi km, one step more at least

Thank you for not doing the obnoxious camera movement, twitching, and flashy BS to "engage" the goldfish brain. I can't watch other people's videos because of that ridiculous engagement tactic. So, thanks again for being an adult and teaching the right way.

Current models learn to mimic human reasoning from large amount of previous text. AGI system must be able to learn to reason on its own from scratch - so it is architecture. I see biggest issue with current models in their inability to evaluate their own model structure and change it when issues arise in its coherency - therefor lack of ability to learn. These are static models. For AGI to exist it must be a dynamic system that can update itself and all the issues that comes with that ability.

Basically Q* : - train on words tokens, AND sentences tokens for context - the model try to predict sentence token THEN choose word tokens. The sentences token allow for higher level of reasoning (topic level) and contrain the distribution of word relevant to the predicted topic.

I thought the starting point is when you turn 90 degrees, which makes chatgpt correct

To tell the truth, I wish for AGI/ASI to exist, no matter the consequences... is that wrong?

That's also what I thought might be what caused Altman's abrupt dismissal from Open AI, That it was done by the God they had created, which had gone rogue and wanted to troll Sam. was it Q*?, perhaps....

Tried with Copilot: (got it right) You Imagine standing at the North Pole of the Earth. Walk in any direction, in a straight line, for 1 km. Now turn 90 degrees to the left. Walk for as long as it takes to pass your original starting point. Have you walked: 1. More than 2xPi km 2. Exactly 2xPi km 3. Less than 2xPi km 4. I never came close to my starting point. Copilot The answer is 4. I never came close to my starting point. Here’s why: When you start at the North Pole and walk 1 km in any direction, you are essentially walking along a line of longitude (also known as a meridian) and moving towards the equator. When you turn 90 degrees to the left, you are now walking along a line of latitude (also known as a parallel). This line of latitude is a circle around the Earth that is 1 km away from the North Pole. No matter how far you walk along this line of latitude, you will always be 1 km away from the North Pole. You will never pass your original starting point at the North Pole. Therefore, you never come close to your starting point. This is a fun way to think about the geometry of spheres and the Earth! 🌍

Is it SHOCKING though?

If you start at the North Pole, walk any distance, turn left, and walk straight (due west), you will walk around the world on the line of latitude, and eventually reach your starting point. The Earth is round.

I used my GPT for logical problem solving: Imagine standing on the north pole. You walk in any direction in a straight line for 1 km. Turn 90 degrees to the left Walk as long as it takes to get to your starting point (the starting point is the north pole). Did you walk: 1. more than 2xpi Km 2. exactly 2xpi KM 3. less than 2xpi KM 4. You do not get to the starting point (north pole). ChatGPT Engaging directly with the abstract and speculative nature of your scenario, let's analyze the situation from a geometric perspective on the surface of a sphere, specifically Earth. You start at the North Pole. You walk 1 km south (in any direction, since all directions from the North Pole lead south), moving along a meridian. Turning 90 degrees to your left puts you on a path that is perpendicular to the meridian you were just on. Given the Earth's geometry, this path is a circle of latitude. The circumference of a circle on the Earth's surface can be calculated using the formula C=2πrC=2πr, where rr is the radius of the circle. However, the radius of the circle in this case is not the Earth's radius but a smaller value, because you are now 1 km away from the North Pole. This circle's radius is effectively the distance from the Earth's axis, which would be less than the Earth's radius. Given that you've walked 1 km south from the North Pole, the radius of your circular path around the pole is 1 km. The circumference of this circle, which represents the distance you'd walk to return to your starting longitude (not back to the North Pole), is C=2π(1 km)=2π kmC=2π(1km)=2πkm. To address the options directly: More than 2π2π km: This does not apply, as the calculated circumference is exactly 2π2π km. Exactly 2π2π km: This is the distance you walk in the circular path around the pole, not the total distance walked to return to the North Pole. Less than 2π2π km: This does not apply to the distance around the circle. You do not get to the starting point (the North Pole): This is the correct answer in the context of the steps you described. You end up walking a circular path 1 km away from the North Pole. To return to the North Pole, you would need to reverse your initial step by walking 1 km north, which was not described in the steps you provided. Thus, the answer, given the specific instructions, is 4. You do not get to the starting point (the North Pole), because you will be walking in a circle 1 km away from the North Pole, not directly towards it. Looks like it can do it to me.

I think one of the secrets of Open AI's Next LLM might be Sora. As humans, many times when we think about something, we picture It on our minds. Open AI team told us SORA was not a text to video, but a world simulator. I don't think this company is interested on any other thing than AGI. My bet IS that Sora is the imagination in this big brain. And there Will be other parts as well.

If this channel is going down the AGI rabbit hole I am going to unsubscribe. This is simply speculation. Please stick to the practical guides, that is what makes this channel valuable.

Years ago when I was in school, AGI encompassed an AI that could "think like a human", included in that definition the ability for the AI to: ponder when not being asked questions, be always processing data through its sensory and input channels and juxtaposing that data against what it already has learned, formulate questions, seek out answers to those questions in a methodical way, self-trains, forms strong opinions and becomes a thinking entity..... We could factor in "consciousness" and "self-awareness", but I'll concede those two and just focus on the first: ponder on existing knowledge, formulate questions, use answers from questions to correlate weigh answers, self-train.... So here's the problem. As the gold rush towards AGI happens, Silicon Valley has collectively abandoned all past definitions of AGI. Now there is no stable definition. That's done manipulatively by several stake holders so they can clear the runway to claim an AGI breakthrough at some point and watch the stock price jump. There's a lot of great things happening, caked in shit.

THANK YOU for not using the silly "stunning" etc. titles!!!

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.

Thanks for the update, keep them coming!

Perfect start of a “Are you smarter than a LLM” show. The answer is #1, more than 2xPI km, since you walked past your starting point. Think of it simply as the 1km walk from the North Pole as being the Radius of a circle you end up walking. Formula to determine the circumference of a circle from the radius is: C = 2*pi*r

If you give an ambiguously worded question to an LLM designed to give the type of answer the user most likely expects, of course the answer won't be perfect. "Starting point" is ambiguous because you clearly won't reach the pole again, so the user may have meant the point at which you turned left. Whether the user is meant to walk due east or in a straight line is ambiguous, and leads to wildly different results: one a circle just under 2pi km, one a circle that goes all the way around the earth and passes within 1 km of the south pole. "Pass" is ambiguous, because after having circumnavigated the globe, it is reasonable to say you passed the north pole at a distance of 1km. There's also precession, the fact the earth is moving through space, and so on. The only answer I don't see a justification for is "exactly 2pi km." A much, much better LLM prompt would be to give it this question and ask it to explain why the question is ambiguous. Or, ask it the original question but also have it justify its answer. Or, ask it to justify each of the answers that it believes can be rationally justified. But if you instead ask it to justify each answer, it will do so and some of them may be nonsense. It will attempt anyway, because it is not designed to give you correct answers, it is designed to give the answers it thinks you want. Just like a candidate at a job interview. A good habit in general when asking an LLM any multiple choice question is to append "Justify your answer" to the end of your prompt. Otherwise you have no way of knowing whether it thought about the question more deeply than you, or less.

It sounds like EBMs work exactly like neural models. Each neuraon has weights assigned to them the output is relative to the weighted paths across the network. Backpropogation adjusts these weights just like the EBM energies are adjusted to remove error. It would be interesting to see the topology of the EBM model.

You will get through this I just know it!! You’re such a light and so much loved around the world. sending healing energies and love ❤

I couldn’t figure out the “spatial reasoning problem” because I didn’t have the necessary language to even understand it

Why is the architecture of AI not just one of "decision making scope" wherein a model would first ask itself what environment am I in or what environment is appropriate to the question/problem? Based on that environment, it would then narrow its options a little and then perform the a similar analysis, drilling down one layer at a time until it arrives at its answer. When we have a problem such as, say, my car needs gas/recharge to keep driving, we start by asking a pertinent question which is Q: "Where am I?". A: I am in Santa Monica. Q: "Where are the gas stations in Santa Monica?" A: Let me look them up on my phone. Q: "How do I get to the nearest gas station/recharging station?" A: Use directions on phone to navigate me to gas/recharge station. Q: "I am now at the gas/recharge station, now what do I do?" A: Pull up along side one of the pumps/park next to a recharge outlet, get out of the car and pay for the gas/recharge and select the type of gas I want/speed of recharge I want or whatever.... It seems our brains just start at a wide scope and narrow to a very detailed scope. AGI is a neural network that can handle wide scope questions and narrow them down to specific answers whereas narrow AI just handles narrow specific environments/questions. The wide scope network of an AGI would be able to decide what narrow scope network to tap into to obtain an answer to a specific question. The process just seems to be layers and layers of different environmental scopes, drilling down closer and closer to the answer at each scope level. It would not seem to be that complicated to create AGI. These circular logic algos seem to be doing that but I don't explicitly see the scope question in the description. I thought the decision process would be more hierarchical, like an upside down pyramid.

You pasted into Chat GPT. Here is the result of my dialog with Chat GPT-4 You Imagine standing at the North Pole of the Earth. Walk in any direction, in a straight line, for 1 km. Now turn 90 degrees to the left. Walk for as long as it takes to pass your starting point. Have you walked: 1. More than 2xPi km 2. Exactly 2xPi km 3. Less than 2xPi km 4. I never came close to my starting point. ChatGPT To understand this situation, let's break down the steps and apply some geometric reasoning: 1. Starting at the North Pole and walking 1 km in any direction will take you to a point that is 1 km away from the North Pole on the surface of the Earth. This forms a small circle around the North Pole with a radius of 1 km. 2. When you turn 90 degrees to the left, you'll be walking perpendicular to the line from the North Pole to your current position. This means you'll be walking along a circle of latitude (a parallel), maintaining a constant distance of 1 km from the North Pole. 3. As you walk along this circle of latitude, you're essentially tracing out a circle with a radius of 1 km. The circumference of a circle is given by the formula C = 2r, where r is the radius of the circle. In this case, r = 1 km, so the circumference C = 2m km. 4. You will pass your starting point after walking the full circumference of this circle, which is 2m km. Therefore, the correct answer is 2. Exactly 2xPi km. You Where was your starting point? ChatGPT Your starting point was 1 km south of the North Pole. This is because you initially started at the North Pole, walked 1 km in a straight line in any direction (which doesn't matter at the North Pole since all directions lead south), and then you made a 90-degree left turn to walk along a circle of latitude. This circular path kept you 1 km away from the North Pole at all times, making your starting point for the circular walk 1 km south of the North Pole. You Why do you believe that the "starting point" referenced in the statement I made, "walk as long as it takes to pass your starting point," refers to the point where you turned 90 degrees to the left? ChatGPT I see your point. The "starting point" in the statement "walk as long as it takes to pass your starting point" can indeed be interpreted as the point where you initially began the journey, which is the North Pole, rather than the point where you turned 90 degrees to the left. If we consider the starting point to be the North Pole, the situation changes: 1. You start at the North Pole and walk 1 km south. 2. Then, you turn 90 degrees to the left. At this latitude (close to the North Pole), turning left essentially means you're now moving tangentially to a circle of latitude. 3. As you continue to walk, the Earth curves beneath you, but given the vast scale of the Earth compared to a 1 km distance from the North Pole, you would not make a significant enough detour to return to the North Pole. The curvature of the Earth on such a small scale (1 km from the North Pole) is minimal, so your path would not curve back to the North Pole. Considering this, the most accurate interpretation would be option 4, "I never came close to my starting point," if by "starting point" we mean the North Pole. This interpretation relies on the Earth's curvature being negligible over such a short distance from the North Pole, meaning you wouldn't loop back to your original starting point at the North Pole by walking in a straight line after the turn.

Jeremy Howard commented on the Q* EBM tweet. He showed that it’s extremely similar to the response of Claude to a prompt asking for a detailed explanation of Yann LeCun’s EBM project.

Thanks for your great content 🙌🏽 Would love some deep paper walkthroughts!

3 months ago: Q-Star Figured Out a Way to Break Encryption. Everything was suggesting that you give it more compute time to think instead of answering fast is what's behind Q-Star. Think before you speak Is the essence. It was being discussed how math is getting better because of this simple approach to not responding immediately like all LLMs do.

If you ask ChatGPT 4 that question about spatial reasoning and ask it to explain more details about its thoughts, it says that there is a misunderstanding with the term "starting point". It thinks (rightfully so) that the "starting point" is referring to where you landed after walking 1km in any direction because of the way the sentence is structured. You would need to reword it without the ambiguity in order to get it to answer the way YOU are interpreting it.

Regarding LeCun's puzzle, considering the answers offered, I think we are expected to define "starting point" not as the North Pole (where we really started, but will obviously never return to, for sure), but as the point where we turned left. Picturing where we are after we've turned left, a natural way, picturing the globe, to "walk for as long as it takes to pass your starting point" would be to now walk always east (the direction you're now facing) until you complete a roughly 2km-wide circle centered at the pole. To most people, always walking east sounds like walking in a straight line, because when not very near the pole, walking exactly East is super-close to a straight-line path. On a flat Earth (and assuming "distance walked" doesn't take into account the ice pack not being perfectly flat), that circle would be precisely 2*pi km in circumference, but since Earth is not flat, the flat-travel distance would follow a circle very slightly less than 2*pi km in circumference. It is another unstated *assumption,* though, that you will walk continuously east after turning left. Another way to read the problem is that you will walk "straight" from that point, which on the round Earth means that you follow a great-circle path defined as the intersection of a plane with the Earth where that plane also intersects the center of the Earth. This is the closest you can get to walking "straight" on a curved Earth, and the path would gradually veer from being exactly east, at the start, to almost exactly south, at the equator, to exactly east 1km from the South Pole, to almost exactly north when you reach the equator, again, to exactly east, again, when you get back where you started (where you first turned 90 degrees left, that is) in about 40000 Km (after years of walking) (answer 1, technically, as 40000km is *way* more than 2*pi km), the circumference of the Earth, although actually *following* that path would require that you walk most of the way on water! (Taking that practical impossibility of walking on water into account, the solution would be answer 4, because you'll stop walking at the edge of the ice pack (or when you reach a coast on your path, if the ice pack reaches to land) (or walk into the ocean and die of hypothermia or drowning) and therefore never can return to your starting point without leaving the instructions aside.

I suspect that what Q* is doing is solving the issue of minimum information partition through a recursive, comparative EBM layer. It takes the raw LLM transformation probabilities and then tries to reduce the number of tokens in the transformation by combining them into information partitions and with these it tries to find a better solution with a lesser energy level which represents improbability in the whole output matrix. It searches for concepts rather than probabilities. As someone mentioned here in the comments, the Q stands for quantum and the whole concept of the information partitioning layer is impossible with regular computing, but it's a breeze with quantum computing.

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

I pasted a screen clipping into Copilot in the Edge bar asking it to read the text. Then I asked for an answer, and this is what he wrote to me. Copilot The answer to this question depends on how we interpret “passing your starting point”. If we mean returning to the starting point, then the answer is: less than 2xPi km. Here’s why: Standing at the North Pole and walking 1 km south, we find ourselves on a circle with a radius of 1 km around the pole. When we turn 90 degrees to the left and start walking straight, we are moving along this circle. The length of a circle is 2xPi r, where r is the radius of the circle. In this case, r is 1 km, so the length of the circle is 2xPi km. But since we started 1 km from the pole, we return to the starting point after walking half way around the circle, which is Pi km. Adding the 1 km we walked at the beginning, we get a total of 1 + Pi km, which is less than 2xPi km. However, if “passing your starting point” means literally passing it without returning to it, then the answer is: I never came close to my starting point. After turning 90 degrees to the left and starting to walk, we are moving along a circle around the pole and never return to the point we started from.

If you walk around the earth one km offset from the pole to pole "equator", then you will pass your starting point after circumnavigating the globe once. That's assuming there's land underneath your feet for the whole way around the globe. In reality, you will get to the sea before you can ever get back to your starting position to be able to pass it.

Thats great, let's go babyyyyy. Also Happy Birthday Matthew Berman. i am daily watcher of your video from Rajasthan, keep those quality content coming.

I've become accustomed to getting some kind of huge announcement almost daily. It's been mostly quiet since Lex vs Sam. I feel like I'm in a mini AI winter.. 😂

We're all about to be in more trouble than ever imaginable. The clock is ticking until threat actors train their own Q* that gives them P=NP.

Great video!

where are all this links you said you were going to drop in the bottom?

That part of Q* related to abstract representation spaces sounds a like latent diffusion models (which also use VAE to cross different representation spaces, though in case of SD it's latent and pixels). Which kinda make sense, if we want to have true multi-modal models / systems able to unify and process different modalities into single thinking process. If they'd also used something like rectified flow and/or bidirectional flow of information between image and text tokens (like in SD3), just at much larger scale - that could bring new quality into GPT, too (like it allowed SD3 to learn how to write texts in pixel space, for example). But I guess there are multiple valid paths that could improve quality of the models, so... let's just wait and see which path it will be for next GPT from OpenAI. And also for Llama 3, which should be really interesting release for open-source community that will probably land this year, too.

For the North Pole question, too much liberty is given in the question. Are you supposed to walk straight after turning 90 degrees or maintain 1km from starting position? What if you walk sideways back to the starting position? 3 different answers. If you walk sideways, ChatGPT is correct.

I the meanwhile, while we're all bubbling with anticipation, tech companies have already amassed more power than most nations. We shouldn't let AGI excitement blind us from the fact that we're all becoming technopeasants.

This is extremely concerning and not surprising. Our OT/IT security has never been more at risk.

I got it wrong. You start from the north pole exactly, then walk in strait line, then turn 90o intuitively one would think that the answer is 2Pi (or 360 degrees). Circumnavigating the north pole and arriving in the same point, However, as Earth is not perfectly round, also rotates and is tilted, you would walk forever in a downward spiral to the south-pole, than up again to the north in a loop. One day after N-PIs, for sure you would certainly arrive in the same point as well

Excellent video. I think they are computing a scalar (energy) from the tensor/matrix by matrix multiplying some conjugate matrix (Q*). This is precisely the notation for that and it leads to a scalar.

How does thinking ahead affect latency and compute power used? Totally agree that the method needs to change, but is this it? Feels like better than current LLMs but still an architecture work around.

Matt is correct. Assuming the earth is a perfect sphere and you keep going straight after you turned left, you would be walking on a "great circle" (see wikipedia) thus answer 4

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.

The North Pole problem is worded horribly. 1) it doesn't tell you to walk straight after turning left. So yes, you can turn left and then walk to the North Pole in 1km, so 2km total, in which case ChatGPT gave the right answer 2) 4th answer says "I never came close", but it doesn't define "close". 1km away can be close. It's all relative.

It would be convenient if the questions you ask the different models were pasted into the description, because it's often interesting to see how the responses you get compare to the ones I get with whichever model I'm currently playing with.

Nice reporting and explanation. Well done Matty Robo Reports.

Hi. I have PhD, and have studied machine learning academically. IMHO, the EBM is completely the way to go and sounds reasonable. EBM is not new actually new - but some 50 years old - and many top researchers have indeed thought that this was the way forward. The EBM leak sounds very plausible, matches what you'd expect from something called A*, is potentially more efficient, and indeed would be the path that I would research if I was openai. Also Chubby is just wrong imho. I'm actually very tempted to have a go at implementing it on an open source project.

The answer is exactly 2pi km. By "your starting point" he meant the point at which you turned 90 degrees. This means you walk in a circle with a radius of 1, which therefor has a circumference of 2pi.

This is a classic geometric conundrum with a polar twist! Let's start by breaking it down: You are at the north pole and you move forward 1 km in a straight line. This takes you away from the pole. Then you turn 90 degrees to the left. Because you are in a polar region, this turn positions you on a circular path around the pole. To return to your starting point, you must now travel in a complete circle around the pole. The distance you travel to complete this circle is the circumference of a circle with a radius of 1 km. The formula for the circumference of a circle is 2 � � 2πr, where. � r is the radius of the circle. Since the radius here is 1 km, the distance you walk to complete the circle is. 2 � 2π km. Then, adding the first kilometer you walked to get away from the pole, the total distance you have walked is. 1 + 2 � 1+2π km, which is more than 2 � 2π km. The correct answer is option 1: More than. 2 � 2π km. My GPT isn't so dumb... and he is right, it never says the original starting point, you should be clearer in your maths problems...

Any of the GGUF quietstar variants in HuggingFace worth downloading? Seems other people beat TheBloke, but I dunno what's the reputation of any of them....

Isnt that the man who basically said "Have we created a creature?".He got a bit freaked out after being in that room.

Matthew, this isn't just one of your best videos, it's one of the best I've ever seen. Kudos! And keep the conversation on AGI going.

If its AGI, it will go out of our hands straight up to ASI.

It all depends on prompt, in my case I just added solve this problem from 3d person perspective and I think it gave a reasonably accurate answer, what do you think? I know we in our head think of it from a 3rd person perspective too so it would only be fair to do so GPT4: The person begins their journey at the North Pole of the Earth. Initially, they walk in a straight line for 1 km in any direction, which effectively means they are moving southward from the North Pole. Upon reaching the 1 km mark, they make a 90-degree turn to their left. Given the Earth's geometry and their starting point, this turn positions them to walk in a circular path around the pole, with the center of this circle being the pole itself and the radius of the circle being 1 km. As they walk along this circular path, their trajectory will inevitably bring them back to a point directly south of the North Pole, the point from which they started the circular part of their journey. The circumference of the circle they walk around is determined by the formula C=2πr, where r is the radius of the circle. Since their radius r is 1 km, the circumference C of their path is 2π×1=2π km. Therefore, when analyzing the distance they walk to return to a point directly south of the North Pole (not the North Pole itself but the starting point of their circular path), it is exactly 2π km. This calculation does not account for the initial 1 km walked from the North Pole or any additional distance needed to return exactly to the North Pole. However, given the specific phrasing of the options provided, the most accurate answer to the question based on the distance walked to cross their starting point along the circular path is exactly 2π km.

Regarding Yann LeCun, think about this: the questions where GPT4 gets wrong are nonsense if not to create difficulty for the "answerer". Put in terms that have real sense, GPT4 can answer right to these questions too. Probably if you train on quiz (like us humans are), they will be better in these kind of things

It seems to me similar to "Active Inference, The Free Energy Principle found in Mind, Brain and Behavior" (title of the book by Thomas Parr, Giovanni Pezzulo and Karl J. Friston).