Open AI's Q* Is BACK! - Was AGI Just Solved?

00:12 New Q* Paper
01:27 Q*
03:09 New Paper
05:17 AlphaGo Explained
08:35 Alphago Search
10:59 Alphacode 2+ Search
14:24 Noam brown On MCTS
17:59 Sam altman Hints at search
19:15 New AGI Approach
20:01 AGI Benchmark
22:20 AGI Benchmark Solved?
24:40 Limits
29:05 Predictions for Future

Links From Todays Video:

Welcome to my channel where i bring you the latest breakthroughs in AI. From deep learning to robotics, i cover it all. My videos offer valuable insights and perspectives that will expand your knowledge and understanding of this rapidly evolving field. Be sure to subscribe and stay updated on my latest videos.

Was there anything i missed?

#LLM #Largelanguagemodel #chatgpt
#AI
#ArtificialIntelligence
#MachineLearning
#DeepLearning
#NeuralNetworks
#Robotics
#DataScience

[spacepikelet]

Hype counter:
37 actually
31 basically
29 very
28 really
28 pretty
15 truly
13 crazy
Todays Total: 181

[klinglt]

That’s basically, genuinely and truly, pretty pretty impressive use of words (mostly).

[ReflectionOcean]

By YouSum Live

00:00:00 Q* resurgence in AI community
00:00:22 Large language models excel in math
00:01:24 GPT 0 project for reasoning tasks
00:02:41 Importance of search in AI systems
00:06:13 AlphaGo's self-improvement strategy
00:07:00 Challenges in advancing language models
00:09:43 Combining language models with search
00:11:20 AlphaCode 2's success with search
00:13:00 Potential of search in AI advancements
00:15:12 Generalizing search for broader applications
00:17:39 Sam Elman hints at improving AI reliability
00:18:02 GPT-4 questioned 10, 000 times for best response
00:20:00 AR AGI Benchmark sets AGI standard
00:21:00 Core knowledge crucial for AGI tasks
00:23:00 Leveraging LLMS for program search in AGI
00:24:00 Vision limitations hinder GPT-4's performance
00:26:00 GPT-4's weaknesses in coding and context understanding
00:29:00 Future models may surpass AGI benchmarks

By YouSum Live

[Killua]

I'm literally truly truly impressed

[JamesJon]

Prediction: AIGrid will "genuinely and truly believe" around 10 things in this video

[igorkudryk]

Hey man, if it helps, here is why we love your videos:

- They report news
- They discuss possible outcomes

Here is NOT the reason why we watch them:

- Hype
- AGI solved
- Impressive
- WOW

So the most annoying part is when you try to make it more sensational then it is

[pmiddlet]

To be fair, the tool integrated with the LLM is more than just 'search'. Monte Carlo simulation in the way they're using it is 2 fold: 1) generating spaces of possible outcomes within s steps of the game given the current state space of the game. 2) generating code samples with some level of noise (quasi-randomness) to present (likely) a (non-exhaustive) number of possible states of the game from which to allow the LLM to take a sequence of positions (I thought I heard alphago had been looking at 50 moves at a time) against a final binary outcome (win/loss), optimize the loss function (so it learns), and then have it examine de novo board configurations, so that it can learn from those as well. This fits well into Bayesian'esque learning (I'm not even sure the 'esque' is appropriate here, as this is almost textbook use of Bayesian methods and simulation to make reasonable decisions to optimize, in this case, win/loss ratio. The sequences of positions are also somewhat hierarchical (strong vs weak plays, vs weak plays as setups to strong plays).

I do wonder however, given the amount of time they had to train this mode, l if the training burden might have been reduced by allowing the model to work from the discriminative on say 'medium' levels of players who have a lower win/loss rate, to the generative, where the sample of losses were a better catalyst for improvement. The synthetic data it would have produced would have had arguably a far more balanced training set to improve itself upon versus the more 'unbalanced' data sets that would be exhibited by the highest level players (i.e. many wins, few losses to learn from). This is assuming that the MCT Self-refine method doesn't look at both sides of the board for each round of player moves.

[mickelodiansurname]

For the last time Q* is a function in the Bellman Equation, how everyone is missing this is beyond me since its been a thing since the 1950's. The Bellman equation helps solve decision-making problems by breaking them down into smaller, manageable parts. It calculates the best action to take by considering the immediate reward and the future rewards of subsequent actions, thereby optimizing long-term outcomes.

[Mimi_Sim]

Ai Industry is the cooler sibling of celebrity gossip…. and I am here for it 🤫

[viddarkking]

So this plus Mixture of Agents could get us quite close to a Pocket AGI that you can run in a 2K to 4K ish PC.
Nice.

[Aggielife]

They often say that Hollywood has inside information on certain tech and what might be possible in the future. I think this is very true. In 1984, the movie terminator came out. I was a little kid 7 years old. Even if you told me back then that in 40 years, we would be close to this, I would have looked at you crazy! “What you talking about Willis?”😂😂😂

[gunnerandersen]

Am I the only one that is gonna point out that you make assumptions as states over and over, like what Ilya worked at and on and on and on? XD

[johntdavies]

Great video, I work in AI, mostly in FinTech and open source models (for privacy). You are a regular part of my AI news and update.

[Gee]

Q* is for complex math like breaking AES 192 encryption. GSM8k (grade school math) is very basic questions like "Claire makes a 3 egg omelet every morning for breakfast. How many dozens of eggs will she eat in 4 weeks?"

[ChristianIce]

AGI is basically Text Prediction + Magic.
We didn't solve the "magic" factor yet :D

[Nova-Rift]

Where did you find that andrej video? I can't find it. please link to it in description.

[brucehorton]

I hope AI companies never trains on AIGrid's transcripts. If they don't blacklist they'll actually, truly, basically start getting really really crazy responses to very very very simple prompts.

[austinrusso]

Thank you for posting the video links.

[devlogicg]

The Robozen Android Game is actually pretty pretty hard if you need something to do whilst listening to AI videos!

[elsavelaz]

Wait …. So the guys wanting more compute to get a computer closer to agi by giving it more compute to get it there … are claiming they didn’t get enough compute to get to do the thing that is what is “not safe”- building agi, and that’s why they tried to fire Altman … THAT is what is pretty pretty crazy