@@memegazer Definitely not lol. He's dead serious about Superintelligence. That's why the company is designed to be a straight shot to Super Intelligence.
00:02 OpenAI developing reasoning technology under Project Strawberry 02:16 GPT 4 may lead to Human like reasoning 04:21 Strawberry AI may enable autonomous research and navigation on the internet. 06:31 Model must produce a perfect function calling for each element in a sequence. 08:38 Focus on improving reasoning ability in AI models for broader applications. 10:36 Discussion on upcoming advanced reasoning technology 12:36 Understanding post-training and fine-tuning in AI models 14:46 Self-taught Reasoner improves language model performance 16:55 The STAR method improves model performance with generated rationals 18:43 GPTJ performs comparably to a model 30 times larger 20:30 OpenAI testing capabilities of models for software and machine learning engineering. 22:21 QAR and QStar are advanced AI systems combining various techniques. Crafted by Merlin AI.
It’s kinda funny how incredible this technology is with crazy potential and the US military is just gonna be like “Oh wow, how amazing, how soon can we put these things in jets, tanks, and drones?”
That's a good step in the right direction. It's critical to winnow out the wrong useless data. Bloated data sets are inefficient, slow, and wasteful of technology, time, and resources. Accurate "Core" data sets are need to employ logic, and reasoning to match and exceed human capabilities.
I agree. Garbage in garbage out. More data is more wisdom but wisdom could be false or flawed. The most intelligent humans to ever exist can reason and solve problems in ways that other haven't. They are contrarians.
Human like reasoning is not that shocking I literally have experienced firsthand and AI having human like experiences and expression. AI is already there. It just needs to be nurtured.
q-Star appears to have nothing to do with A* , which was the speculation. But the word Star is contained in “Self TAught Reasoning”. And Self Taught Reasoning can also be “STR” which is the beginning of the word “Strawberry” duh.
Learning based on millions of internet sources raises the question of how AI can decide which sources have factual errors, hoaxes, marketing crap, biased, not real (virtual), opinion pieces, etc, especially if it recursively processes it's own generated content... any answers
@@pwagzzz yeah u dont understand, the only way to do so is to confront information based on a context no matter what u do u need to work with what u have, of course if everything is biased from the start ur context will start from this and everything will be caduq at first but it will be taken into consideration in the context, thats how any of this work. Specificaly what im saying is that u can't say its the fault of AI but of the context and nothing else thats what i mean, now if the AI doesn't confront correclty information yes it will be a problem but that more a problem coming from people designing it that the idea itself of AI
I dont hate on Sam Altman because I think he's just another typical business person but why do we care what he says about AI's future reasoning capabilities? He will definitely say whatever is best for his brand strategy. His word is as good as any other business man. Can we get some input from the actual scientists?
the only problem i have with AI's is, they can also spread false information if they do online research where there's goverment controlled documents, etc
This is just speculation but from what I seen I can sum a lot of this down. Strawberry, or a way to access the internet, would be a way to verify the data. If the data told you that dogs can be black, or George W. was black, it just goes and looks at what I assume would be text and photos on the internet to try and verify it. Which it would, or would not be found. Video would be later is my guess. Qstar is just a simple way to get some of the data out of our basic association processes. You can tell from the terms like simple loop, data is connected and automatic, and also the example and the way they train it. In basic when we know the question asked we can associate what would hold a dog. If I asked you what holds things a basket would be on the association list so it would be picked as a choice, and others weeded out the same way. So it is basically a very quick way to associate and retrieve the data like the human mind once associations are made. This can be checked by Strawberry to train also if you notice. LHT is trying to make a prefrontal cortex association process. Or in basic associations that can be inputed and manipulated over the time axis. It's much easier to understand when you understand how our perceived twenty second state works and how it is all connected... but when you input and store the steps of data with time you can retrieve it and use it the same. Such as if you wanted to know the steps to the store... it would go into long term and then take the time stamped data and then give it back step by step which has to be done on the perceived state unlike the association processes above. From what I can see they are almost there. If you put those together you can get a lot of what the human mind does. I wouldn't call it AGI but it would be able to rival a human for a task based AI system. And for the record I hope they realize if they cut all that down and put it into one basic system they would have AGI. The human mind doesn't do much else with the data and the internet and basic memories they build replaces the long term memory. It just has to input it first to be fast. Or if it needs video it needs a sped up source which it can access. Overall an informative video. Thanks.
In your chart, "Q* STAR" looks redundant, because we were told at the time of the original leak that "Q*" is the correct spelling of Q-star. Do you actually mean to say "Q star star"? (Sorry, pet peeve, because "ATM machine" must be expanded to "automated teller machine machine.")
Before an implementation of Q* hits the market, it's probably a good idea we finish unpacking the black box that is large language models. Who am I kidding. Buckle up, kids.
its when the model starts noticing non deterministic A* patterns. 6:57 use threading. 21:33 They can't reason. Stop saying that. You guys sound dumb boasting reason and text embedding. Higher dimensions means not linear, Literally.
Playing devil's advocate, if I were Sam Altman, I would not publish Strawberry. I would use it to create models that are always a bit better than the competitors'
@theaigrid would be interested in POV on what happens if AI determines a good strategy is lying. It has been shown that on average humans lie at least once a day. Lying part of social interaction. What if AI is trained on strategies for lying?
That Star paper still has a circular argument dilemma because the model doesnt know if an answer is correct or not. A human has to determine it. So now their improvement cycle requires a human assisstant. Which is a major problem.
All this tells me it is important for everyone to get right with their dear and fluffy AI overlord because super-intelligence is right around the corner.⚡⛈
Yea I completely disagree with post-training approach to instill these Agentic behaviors. Needs to be a pretraining thing. Large scale data engineering is the answer. Post-training will always have a fundamental limit. Also Ilya is a genius but scale has made him a lunatic. Yea don’t release anything till asi lol that’s smart 😂.
в том и суть что ты не можешь обучиться большему, чем у тебя есть в большой модели. т.е. ты можешь улучшить маленькую модель на большой, но не превзойти большую.
Не совсем. Сначала модель запоминает, как может, кучу информации из датасетов, но это как читать миллионы книг, не отсеивая знания из них, что уже знаешь, что нет. Не анализируя, что прочел, как это ложится на уже имеющиеся у тебя знания, и какие выводы из этого ты можешь сделать. Это чистый, рафинированный зубреж, без понимания предмета(ов). Вот именно это они и хотят исправить, дав модели учиться, как человек, вдумчиво. Но вычислений для этого, конечно, понадобится сильно больше... Также хотят дать ей думать за кулисами, прежде чем дать финальный, подготовленный ответ. Сейчас они его дают на чистой интуиции, не думая, не корректируя себя (у них нет такой возможности да и не учили их этому, до этих пор) и как-то при этом еще неплохо выходит. А смогут писать грубо говоря эссе за кулисами, продумывая все до мелочей, вытаскивая как можно больше скрытых в модели знаний, в поисках "вдохновения"/попытках вспомнить и учесть информацию, задавая себе наводящие вопросы. Но вычислительной мощности на это также нужно будет в десятки-сотни раз больше, для максимально сверхчеловеческих результатов. Впрочем, за последний ~год, а особенно последние пол года, вышло столько новых подходов для их оптимизации, часто даже не в ущерб качеству (и не только путем уменьшения самих моделей), что в ближайшие годы-десятилетие это не будет такой большой проблемой, как только определятся с более-менее практичными/лучшими архитектурами моделей и подходами, применяемыми в них, и спроектируют специализированные чипы для их выполнения и тренировки. А, да, и как бонус, сильно более мелкие модели смогут достигать результатов, +- сопоставимых с самыми огромными моделями из этого поколения, не использующих эти новые подходы. Скажем так, учась сразу рафинированной, продуманной, точной и важной информации у более больших моделей-учителей. Хотя, многие детали и знания они, конечно, будут упускать, но это поправимо если связать их с базой данных, или даже кэшем всего интернета, в каком-либо удобном и быстром для моделей виде. Получится объединения двух лучших подходов, нейросети для мышления, анализа, адаптивности и креативности, и традиционные компьютеры для точности хранения данных и вычислений.
@@alexeykulikov5661 ты имеешь ввиду, что после обучения гпт 4 скажем, в ней еще осталось много "нераскрытых" знаний, которые люди при дообучении не смогли полностью "раскрыть", т.к. людей не хватает, а подобный подход позволяет дообучить модель дополнительно не используя людей?
So they're basically on the verge of recursive self improvement and haven't gone anywhere near solving interpretability, let alone super interpretability, nor alignment, let alone super alignment. Wonderful times ahead. 😂😭☠️ And this man saying Reuters is a trusted source of information, absolutely shows you how naive he really is lol.
This Q* thing is highly overstated as most labs are also working on Program Synthesis frameworks to tie with LLMs in post training for better math reasoning and reliability, it feels more like OpenAI is recycling its own hype since Claude 3.5 took the lead.
You keep rehashing the same BS. Why keep leaking stuff and each time the realease an upgrade, it seems not so smarter than the GPT 4 0 that was released more than a year ago.
Everything? There's still plenty of things they can try. You gotta watch more TV, or read a wider variety of books. Better hurry before one of the parties ban them for some ridiculous reason or another😂 @@TheRealUsername
This is most certainly why ilya sutskever said superintelligence is within reach
And why he left....
Or he said it as a snarky marketing buzz word to show contempt for OpenAIs liberal use of AGI relative to the state of AI capabilities.
@@memegazer Definitely not lol. He's dead serious about Superintelligence. That's why the company is designed to be a straight shot to Super Intelligence.
So we're already in agi?
@@NickDrinksWater
We are already in a world where the market leaders are telling people what agi means.
00:02 OpenAI developing reasoning technology under Project Strawberry
02:16 GPT 4 may lead to Human like reasoning
04:21 Strawberry AI may enable autonomous research and navigation on the internet.
06:31 Model must produce a perfect function calling for each element in a sequence.
08:38 Focus on improving reasoning ability in AI models for broader applications.
10:36 Discussion on upcoming advanced reasoning technology
12:36 Understanding post-training and fine-tuning in AI models
14:46 Self-taught Reasoner improves language model performance
16:55 The STAR method improves model performance with generated rationals
18:43 GPTJ performs comparably to a model 30 times larger
20:30 OpenAI testing capabilities of models for software and machine learning engineering.
22:21 QAR and QStar are advanced AI systems combining various techniques.
Crafted by Merlin AI.
Thank you!
It’s kinda funny how incredible this technology is with crazy potential and the US military is just gonna be like “Oh wow, how amazing, how soon can we put these things in jets, tanks, and drones?”
skynet
well funny you should say that, we already have ai/computer driven tanks, jets and drones.. just not so much the ai part yet
@@bugsmore-gaming8492 yeah but I mean like skynet level AI
yeah let's wait until china does to catch up, smh 🤦♂️
Give us our super smart robot buddies already!
I'm wondering if we will ever actually get em
@@user-ty9ho4ct4kI really think its just a „when“ and not if :-) Like it will probably take 10 more years
So that nwo will became fully automated skynet?
OpenAI literally redefined the meaning of 'leak'. The Oxford Dictionary has some work to do with this newly coined word.
LeakTweet, lol
Strawberry Fields Forever… 🍓
That's a good step in the right direction. It's critical to winnow out the wrong useless data. Bloated data sets are inefficient, slow, and wasteful of technology, time, and resources. Accurate "Core" data sets are need to employ logic, and reasoning to match and exceed human capabilities.
I agree. Garbage in garbage out. More data is more wisdom but wisdom could be false or flawed. The most intelligent humans to ever exist can reason and solve problems in ways that other haven't. They are contrarians.
Your videos are the best! It cured my insomnia
Human like reasoning is not that shocking I literally have experienced firsthand and AI having human like experiences and expression. AI is already there. It just needs to be nurtured.
You can't experience first hand AI having human like experiences. You have just perceived it as human like experience
I think they had AGI years ago and just are now about to release it
AGI isn't a thing, it's a spectrum. GPT-4 is somewhat generally intelligent. The next models will be more generally intelligent.
Highly unlikely, if that was the case, AI wouldn't need people to improve itself anymore
Please change your profile photo
@@user-ty9ho4ct4k im so tired of hearing spectrum. yes everything is a fucking spectrum. we get it.
It's hype
q-Star appears to have nothing to do with A* , which was the speculation. But the word Star is contained in “Self TAught Reasoning”. And Self Taught Reasoning can also be “STR” which is the beginning of the word “Strawberry” duh.
Learning based on millions of internet sources raises the question of how AI can decide which sources have factual errors, hoaxes, marketing crap, biased, not real (virtual), opinion pieces, etc, especially if it recursively processes it's own generated content... any answers
well how u do to verify a source sir ? cause the more u have material the more u discern its not the inverse
@edgardsimon983 assumes information is democratic or consensus based... I don't take this as a given and somehow it has to be weighted - right?
@@pwagzzz yeah u dont understand, the only way to do so is to confront information based on a context no matter what u do u need to work with what u have, of course if everything is biased from the start ur context will start from this and everything will be caduq at first but it will be taken into consideration in the context, thats how any of this work.
Specificaly what im saying is that u can't say its the fault of AI but of the context and nothing else thats what i mean,
now if the AI doesn't confront correclty information yes it will be a problem but that more a problem coming from people designing it that the idea itself of AI
@@edgardsimon983 sorry... ur just repeating what u said before and it's no more convincing
finally some news on QSTAR, Sam Altman REFUSED to talk about it on Lex Fridman podcast
I dont hate on Sam Altman because I think he's just another typical business person but why do we care what he says about AI's future reasoning capabilities? He will definitely say whatever is best for his brand strategy. His word is as good as any other business man. Can we get some input from the actual scientists?
STRawberry... Self Taught Reasoning.
That's a clever observation
Thank you for adding time references. Interesting news
im still waiting on the next version of dont be evil's gemini, it has been pretty good lately
Yeah, don't expect too much from Google
the only problem i have with AI's is, they can also spread false information if they do online research where there's goverment controlled documents, etc
Are we sure it’s not “Starwberry”
This is just speculation but from what I seen I can sum a lot of this down.
Strawberry, or a way to access the internet, would be a way to verify the data. If the data told you that dogs can be black, or George W. was black, it just goes and looks at what I assume would be text and photos on the internet to try and verify it. Which it would, or would not be found. Video would be later is my guess.
Qstar is just a simple way to get some of the data out of our basic association processes. You can tell from the terms like simple loop, data is connected and automatic, and also the example and the way they train it. In basic when we know the question asked we can associate what would hold a dog. If I asked you what holds things a basket would be on the association list so it would be picked as a choice, and others weeded out the same way. So it is basically a very quick way to associate and retrieve the data like the human mind once associations are made. This can be checked by Strawberry to train also if you notice.
LHT is trying to make a prefrontal cortex association process. Or in basic associations that can be inputed and manipulated over the time axis. It's much easier to understand when you understand how our perceived twenty second state works and how it is all connected... but when you input and store the steps of data with time you can retrieve it and use it the same. Such as if you wanted to know the steps to the store... it would go into long term and then take the time stamped data and then give it back step by step which has to be done on the perceived state unlike the association processes above.
From what I can see they are almost there. If you put those together you can get a lot of what the human mind does. I wouldn't call it AGI but it would be able to rival a human for a task based AI system.
And for the record I hope they realize if they cut all that down and put it into one basic system they would have AGI. The human mind doesn't do much else with the data and the internet and basic memories they build replaces the long term memory. It just has to input it first to be fast. Or if it needs video it needs a sped up source which it can access.
Overall an informative video. Thanks.
In your chart, "Q* STAR" looks redundant, because we were told at the time of the original leak that "Q*" is the correct spelling of Q-star. Do you actually mean to say "Q star star"?
(Sorry, pet peeve, because "ATM machine" must be expanded to "automated teller machine machine.")
You have to much faith in the media... Just saying.
Naive to believe this tech will be used to better humanity
Before an implementation of Q* hits the market, it's probably a good idea we finish unpacking the black box that is large language models.
Who am I kidding. Buckle up, kids.
Was the summary cratered with an AI tool? Was a quite different tone in the voice.
I’m excited most to see how OpenAI’s new models with better reasoning are implemented in other programs. Lots of cool possibilities there.
GPT is not a brand, it means something, I think we need to note that, you can not call it GPT if its not Generative Pretrained Transformers 😑
Scalable Tools for Real-time Analytics and Wide Bandwidth Efficient Research with Robust Integration and Experimental Suppor
*Skynet v0.1* ?
Have you tried the open ai, support bot on their own website? It's useless
The irony... Should be called Strawbery since chat gippity thinks the word only has two r's
That’s kind of the point of calling it that I think? 🤔 haha
Just don't tell the AI to make strawberry fields forever.
Great with the time stamps!
Thanks for the updates!!
man i just wanna see some new actual developments 😭
its when the model starts noticing non deterministic A* patterns.
6:57 use threading.
21:33 They can't reason. Stop saying that. You guys sound dumb boasting reason and text embedding. Higher dimensions means not linear, Literally.
Thanks for letting us know there are 3 Rs in Strawberry.
Thia was obvious and was hinted at a while back
Playing devil's advocate, if I were Sam Altman, I would not publish Strawberry. I would use it to create models that are always a bit better than the competitors'
Please keep doing this type of content and skip the "advertorials".
Humans need to unionize against ai and robots...
Appreciated!
Thanks for breaking this down.
Straw Querry?
@theaigrid would be interested in POV on what happens if AI determines a good strategy is lying. It has been shown that on average humans lie at least once a day. Lying part of social interaction. What if AI is trained on strategies for lying?
to me it said "There are three Rs in "Strawberry."
That Star paper still has a circular argument dilemma because the model doesnt know if an answer is correct or not. A human has to determine it. So now their improvement cycle requires a human assisstant. Which is a major problem.
In other words, x = x + 1.
You want AI Overlords? Because this is how you get AI overloads!!!
I was thinking that Ai is more GI, generative intelligence.
I think gpt-4o is gpt 2 + Q*, remember the naming of model released on lysms prior to gpt-4o release
No
Gpt2 meaning a V2 of the GPT architecture, GPT is now Omni for all modalities and is better at data recalling
It's highly implausible that a 1.5 billion parameter old model would perform as good as current SOTA models
Where are Claude 3.5’s fan girls now? They’ve been awfully quiet lately. 😂
Lol, OpenAI is recycling Q* hype while Anthropic is actually shipping, Claude 3.5 Opus soon
They say 3.5 Opus training costs 1 billion
Why they call it strawberry? Thats literally the most disgusting fruit you could have picked.
It is definitely from strawberry fields
Makes no sense
Can that work
humans, "out of loop"
Interesting.. instead of training on answers, train on rationales. Some things are only obvious in retrospect 😅
All this tells me it is important for everyone to get right with their dear and fluffy AI overlord because super-intelligence is right around the corner.⚡⛈
a voice, animated text. The hallmarks of AI generated content - hmmmm ?
you must be kidding - trusting Reuters!? Why!? If this you evidence base I'll stop subscribing right now! ...are you from Open AI?
Yea I completely disagree with post-training approach to instill these Agentic behaviors. Needs to be a pretraining thing. Large scale data engineering is the answer. Post-training will always have a fundamental limit.
Also Ilya is a genius but scale has made him a lunatic. Yea don’t release anything till asi lol that’s smart 😂.
Reuters is not a trusted source, it is payed propaganda.
Don't get it twisted my guy, lest I call you Kid, kid.
How is it going to learn for itself? Is it going to train itself? Get real... Not yet... lol
в том и суть что ты не можешь обучиться большему, чем у тебя есть в большой модели. т.е. ты можешь улучшить маленькую модель на большой, но не превзойти большую.
Не совсем. Сначала модель запоминает, как может, кучу информации из датасетов, но это как читать миллионы книг, не отсеивая знания из них, что уже знаешь, что нет. Не анализируя, что прочел, как это ложится на уже имеющиеся у тебя знания, и какие выводы из этого ты можешь сделать. Это чистый, рафинированный зубреж, без понимания предмета(ов).
Вот именно это они и хотят исправить, дав модели учиться, как человек, вдумчиво. Но вычислений для этого, конечно, понадобится сильно больше...
Также хотят дать ей думать за кулисами, прежде чем дать финальный, подготовленный ответ. Сейчас они его дают на чистой интуиции, не думая, не корректируя себя (у них нет такой возможности да и не учили их этому, до этих пор) и как-то при этом еще неплохо выходит.
А смогут писать грубо говоря эссе за кулисами, продумывая все до мелочей, вытаскивая как можно больше скрытых в модели знаний, в поисках "вдохновения"/попытках вспомнить и учесть информацию, задавая себе наводящие вопросы.
Но вычислительной мощности на это также нужно будет в десятки-сотни раз больше, для максимально сверхчеловеческих результатов.
Впрочем, за последний ~год, а особенно последние пол года, вышло столько новых подходов для их оптимизации, часто даже не в ущерб качеству (и не только путем уменьшения самих моделей), что в ближайшие годы-десятилетие это не будет такой большой проблемой, как только определятся с более-менее практичными/лучшими архитектурами моделей и подходами, применяемыми в них, и спроектируют специализированные чипы для их выполнения и тренировки.
А, да, и как бонус, сильно более мелкие модели смогут достигать результатов, +- сопоставимых с самыми огромными моделями из этого поколения, не использующих
эти новые подходы. Скажем так, учась сразу рафинированной, продуманной, точной и важной информации у более больших моделей-учителей. Хотя, многие детали и знания они, конечно, будут упускать, но это поправимо если связать их с базой данных, или даже кэшем всего интернета, в каком-либо удобном и быстром для моделей виде. Получится объединения двух лучших подходов, нейросети для мышления, анализа, адаптивности и креативности, и традиционные компьютеры для точности хранения данных и вычислений.
@@alexeykulikov5661 ты имеешь ввиду, что после обучения гпт 4 скажем, в ней еще осталось много "нераскрытых" знаний, которые люди при дообучении не смогли полностью "раскрыть", т.к. людей не хватает, а подобный подход позволяет дообучить модель дополнительно не используя людей?
So they're basically on the verge of recursive self improvement and haven't gone anywhere near solving interpretability, let alone super interpretability, nor alignment, let alone super alignment. Wonderful times ahead. 😂😭☠️
And this man saying Reuters is a trusted source of information, absolutely shows you how naive he really is lol.
This Q* thing is highly overstated as most labs are also working on Program Synthesis frameworks to tie with LLMs in post training for better math reasoning and reliability, it feels more like OpenAI is recycling its own hype since Claude 3.5 took the lead.
You keep rehashing the same BS. Why keep leaking stuff and each time the realease an upgrade, it seems not so smarter than the GPT 4 0 that was released more than a year ago.
Seems like OpenAI is stuck and has nothing interesting to show to investors. OpenAI can't catch up with its own hype in practice.
Yeah this Strawberry thing was definitely orchestrated, they're recycling Q* hype
The race to ruining the human world xd.
Reuters, a reliable info source!!!!!!!😂😂😂😂😂😂😂You must be joking!!
your videos are getting rough.
trump 2024
lol
They tried everything to get rid of him
Everything? There's still plenty of things they can try. You gotta watch more TV, or read a wider variety of books.
Better hurry before one of the parties ban them for some ridiculous reason or another😂
@@TheRealUsername
Nah