[hamdani yusuf (OP)]

OpenAI's Q* is the BIGGEST thing since Word2Vec... and possibly MUCH bigger - AGI is definitely near

Timestamps:
00:03 Q* is the biggest breakthrough since Word2Vec
02:34 Q* is a hybridization of q-learning and AAR algorithm, capable of accurate math calculations.
04:44 The Q* algorithm has the potential to unlock a new classification of problems that can be solved.
07:11 A seismic shift has occurred at OpenAI regarding AGI achievement according to a leaked letter.
09:39 Qualia has demonstrated an ability to improve optimal action selection policies and apply it to cross-domain learning.
11:59 OpenAI's Q* has achieved impressive decryption abilities without the need for keys.
14:33 Q* can significantly disrupt cryptography and achieve feats that were thought to be only possible for Quantum Computing.
16:50 Q* is a significant advancement with the potential to solve math problems like AGI.
18:59 OpenAI's Q* has the potential for self-transformation and creative problem solving
00:00 OpenAI's Q* is a significant advancement in AI technology

[hamdani yusuf (OP)]

It looks like evolutionary convergence is also happening to AI models, just like memes.

Q* Q Star Hypothesis | Is this hybrid of GPT and AlphaGO? AI self-play and synthetic data

TIMELINE
[00:00] Into
[06:14] ORCA 2 and Synthetic Data
[14:42] The Q* Hypothesis
[28:46] Dr Jim Fan
[30:00] Wait But Why by Tim Urban
[31:31] Dr Jim Fan cont.
[36:44] Jimmy Apples

[hamdani yusuf (OP)]

What is Q-Learning (back to basics)

What is Q-Learning and how does it work? A brief tour through the background of Q-Learning, Markov Decision Processes, Deep Q-Networks, and other basics necessary to understand Q*

OUTLINE:
0:00 - Introduction
2:00 - Reinforcement Learning
7:00 - Q-Functions
19:00 - The Bellman Equation
26:00 - How to learn the Q-Function?
38:00 - Deep Q-Learning
42:30 - Summary

[hamdani yusuf (OP)]

OpenAI Q* might be REVOLUTIONARY AI TECH | Biggest thing since Transformers | Researchers *spooked*

The video tries to analyze the rumored progress of Q* as an outsider, while describing what could be the consequences if it was true.

[alancalverd]

The phrase "synthetic data" should strike terror into the heart of any rational human. It's the stuff of Trumpism.

[hamdani yusuf (OP)]

The phrase "synthetic data" should strike terror into the heart of any rational human. It's the stuff of Trumpism.

It's analogue to thought experiments. It's also like questions and answers in textbooks used by students given by their teachers. The data can be used to check the internal consistency of our current models and assumptions, but they can't be used to check the consistency of our models and assumptions with physical reality.
For analogy, the large models act as the teachers for the smaller models using simpler data. Learning and understanding are often regarded as data compression process.

[alancalverd]

Dangerous corruption of language. Data is stuff you measure. Test data is fair enough: an input intended to test your black box, but it's only useful if you know what the output should be - and that's a problem for the Believers in AI because by definition you don't know what to expect. But "synthetic data" is just lies - fake news with numbers. 

[hamdani yusuf (OP)]

Dangerous corruption of language. Data is stuff you measure. Test data is fair enough: an input intended to test your black box, but it's only useful if you know what the output should be - and that's a problem for the Believers in AI because by definition you don't know what to expect. But "synthetic data" is just lies - fake news with numbers. 

Pilots or operators of other machineries are commonly trained with simulators, which is a form of generator of synthetic data. As long as its limitation determined beforehand, it's still useful in most situations.

[alancalverd]

It's not synthetic but simulated as close to reality as possible. If it was wholly synthetic it would be no use. The old Hong Kong Kai Tak airport had a very unusual and complicated approach with all sorts of skyscrapers and other towers to avoid, plus various noise restrictions, so the sims were necessarily very accurate representations of reality, and very different from Boston Logan or London Heathrow.  You could certainly synthesise an entirely fictional airport or warehouse, but what would be the point? 

[hamdani yusuf (OP)]

It's not synthetic but simulated as close to reality as possible. If it was wholly synthetic it would be no use. The old Hong Kong Kai Tak airport had a very unusual and complicated approach with all sorts of skyscrapers and other towers to avoid, plus various noise restrictions, so the sims were necessarily very accurate representations of reality, and very different from Boston Logan or London Heathrow.  You could certainly synthesise an entirely fictional airport or warehouse, but what would be the point? 

AFAIK, simulated data are synthetic data, regardless of how similar they are to reality.
In the case of smaller LLM using synthetic data provided by larger LLM, the latter is expected to have filtered out the signal from noise of the original data. What's important is that relevant/significant parts are preserved in the synthesized data, while irrelevant/misleading parts are discarded.

[hamdani yusuf (OP)]

Google's New AI "GNoME" Discovered Millions of New Materials (Reinvents Batteries)

Google's GNoME AI, developed by the team behind AlphaFold, is transforming material science by rapidly predicting the structure of new materials. This breakthrough AI tool significantly impacts fields like solar energy, battery development, and computer chip manufacturing, offering efficient and sustainable solutions. GNoME's ability to analyze millions of materials quickly showcases a monumental advancement in material discovery and technology.

[hamdani yusuf (OP)]

What is OpenAI?s super-secret Project Q*? | About That

A breakthrough in artificial intelligence at OpenAI preceded former CEO Sam Altman's firing and was part of a list of the board's grievances, according to sources cited in a Reuters report. Andrew Chang explains what we know about the AI technology referred to as Q*, and also breaks down the gap between current AI technology and "human" intelligence.

[alancalverd]

What's important is that relevant/significant parts are preserved in the synthesized data, while irrelevant/misleading parts are discarded.

In other words, it's a sim, not a synth.

[hamdani yusuf (OP)]

What's important is that relevant/significant parts are preserved in the synthesized data, while irrelevant/misleading parts are discarded.

In other words, it's a sim, not a synth.

As commonly accepted by AI researchers and data scientists, understanding the real world is a process of data compression. It converts data points into insights, patterns, general rules, equations, or algorithms.
Generating simulations and synthetic data is the reverse process, which is useful to train smaller AI agents to get the same understanding more efficiently using less data and less memory.
Note that the synthetic data is used to make smaller agents learn and able to produce the correct outputs during deployment with real world data.
What's generated by Alpha Zero by self playing is synthetic data. It's not produced by some measurements.

[alancalverd]

What's generated by Alpha Zero by self playing is synthetic data.

So AI is a complicated and unsatisfying form of masturbation? Might account for the sort of garbage that chatbots deliver.

[hamdani yusuf (OP)]

What's generated by Alpha Zero by self playing is synthetic data.

So AI is a complicated and unsatisfying form of masturbation? Might account for the sort of garbage that chatbots deliver.

AI faces no sexual pressure just for survival.
Nevertheless, that strategy has successfully been proven to beat the best human players of chess and go.
Humans delivering garbage in online chats is not an especially rare occurance.

[hamdani yusuf (OP)]

I've been wondering about this idea since high school. I built and used a simple analog calculator for my scientific research competition using op-amps.

Efficiency is the main advantage of analog systems for AI. The Neuromorphic Chips are more closely resembling brains than traditional digital computer chips.

Sam Altman's Brain Chips | Rain Neuromorphic Chips | UAE Funds and US National Security and Q*

[alancalverd]

Since both chess and go are mathematically trivial, it is entirely likely that a machine can beat a human at either. So what? The whole point of either game is to enjoy the competition: they are simply human pastimes at the other end of the brain/brawn spectrum from boxing and rugby.  I have never seen the pint (a handy typing error!) of playing any game against a machine. Would you wrestle a JCB for pleasure?

[hamdani yusuf (OP)]

Since both chess and go are mathematically trivial, it is entirely likely that a machine can beat a human at either.

It always seems impossible until it's done.

Nelson Mandela

[hamdani yusuf (OP)]

With great power, comes great risk.

What If Someone Steals GPT-4?

0:02
At the heart of it, a Large Language  Model or LLM is just two files.
The first file is like about  500 lines of C-language code.
The second file is just hundreds of billions  or trillions of seemingly random numbers, 
the "parameters". But this  is where the magic happens.
0:23
Based on current evaluations - which have their  shortcomings, yes - the more parameters the 
model has, and the more tokens they are  trained on, the more capable they get.
0:34
The models themselves are economically  valuable. They carry proprietary trade 
secrets and - when separated from their safety  systems - can exhibit malicious capabilities.
0:45
The data that helped train those  models is also valuable. Nowadays, 
good and useful LLM data is produced at  considerable cost, often by educated workers.
0:56
If more and better data creates better models, 
then there is significant commercial incentive for  state actors, smaller and less ethical AI labs, 
or even just hacktivists to bootstrap their  performance by stealing from a leader.
1:11
What if someone stole GPT-4? We should be  talking about this risk. In this video, 
a few thoughts about protecting  these LLMs from theft.