This is about the fastest I can put out a video 😅 I'm more excited about o1 than I was about GPT-4o for sure. Discord: discord.gg/AgafFBQdsc Patreon: www.patreon.com/DrWaku
@JackUnity
3 күн бұрын
Dr Waku, your level of understanding is outstanding! I have watched dozens of videos on this topic, and your explanations are by far the clearest and most in depth. You have a talent for simplifying complex topics for the layperson. Thank you!
@icns01
2 күн бұрын
Interesting..I was about to say something very similar..less than 1/4 of the way into the video you can appreciate the refreshing clarity..👍
@DrWaku
2 күн бұрын
Thank you very much! I developed that skill of explanation over many years, and I'm happy to be able to put it to good use. See you on the next video!
@creatvsdd99
2 күн бұрын
I am extremely looking forward to the near future after seeing your video Dr. Waku!🥳 I hope AGI can grant any wish we have it like Timmy Turner and The Fairly OddParents:) The future is going to be very fun👌 and I wish AGI will solve issues like mental health issues, cancer, diabetes, and more which you can think of.
@thomasruhm1677
2 күн бұрын
How can I learn this?
@TheChadavis33
2 күн бұрын
It really amused me when many commentators (and some even now) were saying it was all hype and that we won’t get better with our AI models. This train isn’t slowing down, and people need to stop coping.
@human_shaped
2 күн бұрын
MLST *cough*
@DrWaku
2 күн бұрын
The road ahead always looks flat, until it isn't.
@JD-im4wu
2 күн бұрын
....they love copium though, the harsh reality of wasting many years writing long thesis papers and phd etc is too harsh to bear for them to be replaced by an LLM. Now we can all be slaves to our new AI Keeper Overlords like Sam Altman unless his inner team brains like ilya was trying to hijack the rug from under him was trying to do. but i find that scenerio unlikely since these LLMs have a huge thirst for power which requires big bucks but who knows if they are building side projects under their nose that has more advanced reasoning capabilities? The near future is going to look very interesting... but lets not forget the gap between us and the overlords is going to grow big time. its not the AI we need to fear its the controllers of the AI the overlords behind it all who are in a secret war between each other right now for that dominant position that Sam Altman seems to have atm. the only reason they are offering it free now is so we the slave users build it for them and when it reaches its peak they are going to use it against us. especially a guy like Altman I can not trust in his benevolence.
@marvinvogtde
2 күн бұрын
i donno i am really not that impressed with the current pace. I also still dont see how we get anything resembling AGI from generative AI models. If we ever get to the singularity i highly doubt it will be a generative AI model
@Jeremy-Ai
Күн бұрын
@@DrWaku Agreed. “Everything is until it isn’t, well… unless of course it never was. “Which it is and isn’t in this fractured attempt at a moment.” Take care Jeremy
@DaveShap
3 күн бұрын
This is the way
@DrWaku
3 күн бұрын
May the scaling be with you
@kairi4640
3 күн бұрын
Do you know the way? 😃
@ppbroAI
3 күн бұрын
This is a Star Wars vs Star Trek moment
@AAjax
3 күн бұрын
Nanu Nanu
@wizzyno1566
3 күн бұрын
How do you know? Nobody outside the company knows if it works yet.
@DecentGradient
3 күн бұрын
It has PHD level reasoning, but is not AGI. I'm absolutely befuddled what people think is AGI these days. There are most certainly weak areas that need improving, and far more unhobbling to do to iron it out, but this model is easily generally smarter than the average person. I fail to see how that's not AGI.
@h-e-acc
3 күн бұрын
Folks are just moving the goalpost. They wanted to see AI be able to do plumbing, electrical, etc. first. I think at this point they’ll keep moving the goalpost until we see bots that can think and feel and can do all sorts of things. Basically it’s just cope.
@DecentGradient
3 күн бұрын
@@h-e-acc I think you're right. They're waiting for it to interact with the world more, but that has little to do with the intelligence of the model. That's just building out infrastructure and tooling around the model. The raw intelligence is there already.
@gsigas
2 күн бұрын
I agree AGI used to mean a system capable of learning at the same scale and variety as a single regular person. Now it means a system capable of learning everything humanity as a whole is capable of at the levels of elite humans.
@lamsmiley1944
2 күн бұрын
I don’t think we can call it AGI when it still gets the number of “r”s in strawberry wrong half the time.
@DecentGradient
2 күн бұрын
@@lamsmiley1944 I was under the impression OpenAI o1 was now acing that particular problem as well as similar ones.
@d4rkside84
3 күн бұрын
Dr. Waku is for me the best Ai Contencreator :)
@DrWaku
3 күн бұрын
Aww thanks :))
@funmeister
2 күн бұрын
I absolutely love this channel-- Dr. Waku, please keep going with the splendid topics and sublime clarity with which you explain everything. And love the insight that while o1 isn't AGI, it's definitely an important component of it, and it's like a new paradigm that hasn't be optimized for scale yet, with a long runway to go.
@silent6142
3 күн бұрын
I think this is the birth of AGI but like a child, it needs to develop which I'm sure will be exponential.
@Masta2Playa
Күн бұрын
No, its not Agi
@chrissscottt
3 күн бұрын
Dr Waku's analysis and description of the various concepts involved is a cut above the other AI analysts I watch. Nice one.
@chrisanderson7820
3 күн бұрын
This is the first model where I can't just dismiss it as a stochastic parrot/pattern mapper. The combo of increased diversity and chain of thought is really blurring the line between an LLM and an actual reasoning engine.
@DrWaku
2 күн бұрын
Yeah. I think it is an actual reasoning engine. It doesn't follow rules as precisely as say a chess reasoning engine, but you can point it at the same types of problems and get decent results.
@blijebij
18 сағат бұрын
Totally agree!
@paradox_productions
2 күн бұрын
Imma start posting this across sites: Petition to change AGI to mean “Autonomous General Intelligence” because that’s really what people seem to need 😂 Ability to think about as well as humans without prompting. With naturally inquiring, storing that information, recalling it for future usage, and making independent and novel assessments based on URLs and other research
@Techtalk2030
3 күн бұрын
For us people who dont have jobs that think too much in math and coding this model doesnt seem that exciting, but in reality it is since its a big step towards AGI
@DrWaku
3 күн бұрын
Yeah exactly. I think even for people that do have coding jobs, this might be a marginal improvement but not hugely different. I guess we'll know when they start lifting the usage caps on o1
@wizzyno1566
3 күн бұрын
@@DrWakuim an analyst programmer. About 10% of my time is spent actually typing code. I'm not remotely worried by AI. At the moment AI is excellent at generating bullshit, not much use for answering precise questions in a low knowledge base environment, which is basically what writing a new software product or fixing a bug is. For programming at the moment its just fancy google. But... My job is basically automating peoples jobs. So if i can be replaced we (society) really are fxcked, work and employment wise, at least office work.
@JD-im4wu
2 күн бұрын
@@wizzyno1566 yea but i am noticing the advancement in the code its spitting out. before most of it was hallucinagenic garbage, now however, its getting scarier and scarier for a BS graduate in Computer Science. I am convinced that the programmer of today is going to end up being the horse carriage driver of yesterday... very very soon.
@ShaneMcGrath.
2 күн бұрын
@@JD-im4wu A lot sooner than most realise! Might be a shock as those that thought their jobs were among the safest may be the first in line to go an example being lawyers.
@BAAPUBhendi-dv4ho
2 күн бұрын
Every time you watch this video Singularity is 16minutes and 46 seconds nearer
@WoodlandT
3 күн бұрын
Thank you for this explanation. I am very interested in technology but not particularly techy. You made this easy to understand and expanded my understanding significantly. subscribing 👍🏻
@DrWaku
3 күн бұрын
Thank you very much! You are exactly the sort of person I'm hoping to reach. See you in the next one
@taWay21
2 күн бұрын
More Compute = Linearly Improving Performance. Thats sniffing at Singularity yall.
@DrWaku
2 күн бұрын
Agreed. I was surprised to see those plots.
@Gnaritas42
2 күн бұрын
That's a log chart
@MarkCW
2 күн бұрын
Dr Waku, I really like your videos and how you present the information in a way I can understand. We're at an exciting time in history with the advent of AGI around the corner.
@DrWaku
2 күн бұрын
Thanks Mark. It's a really exciting time to be living and learning about science!
@MichaelDeeringMHC
3 күн бұрын
I don't like their naming strategy.
@DrWaku
3 күн бұрын
It's so confusing that I've spent at least 15 minutes thinking about whether they planned it out or they just have a greedy naming algorithm and pick whatever seems best in the next moment
@human_shaped
2 күн бұрын
Sam has been saying for years that he doesn't like "GPT" and the next release would change it. They were caught by surprise by the sudden success of the models and the name kind of stuck against their will. To keep things simple, I hope they just stick with this. OpenAI o1, o2, o3, ... is a nice simple naming scheme once we get over the bump of the initial change. But there will always be some little branching products and experiments, so hoping for a nice linear scheme is probably unrealistic.
@lnebres
2 күн бұрын
Lucid explication, as usual. Mahalo, Dr. Waku!
@DrWaku
2 күн бұрын
Thank you again! See you in the next one.
@GoronCityOfficial
3 күн бұрын
I am participating in the o1 program for researchers and you are incredibly accurate in your takes. Temporal tracking to such a degree is the first steps to AGI, but given compute cost, optimizations in the models, I definitely am starting to think we can reach AGI by 2030, +/- a couple years
@ManicMindTrick
2 күн бұрын
They chose an intentional cute and harmless name. Strawberry could never do any harm. Skynet could.
@MTd2
2 күн бұрын
It does seem AGI. It seems to be better than the average person in all skills, except for spatial navigating skills. But I wouldn't consider this to be the most relevant. Stephen Hawkins would still be an excelent skills if he couldn't drive his wheelchair.
@The_Mining_Pickaxe
2 күн бұрын
At this rate, we're getting AGI before GTA 6...
@blijebij
18 сағат бұрын
This made me laugh out loud xD (fantastic).
@mackblack5153
3 күн бұрын
Wow! That was a very clear and precise video about O1 preview...I didn't catch why the hype was so intense...Now I get it! Thanks!
@DrWaku
2 күн бұрын
Thanks for watching and commenting! Really appreciate it.
@carlkim2577
3 күн бұрын
Wonderful explanation. Clear and without breathless hype. Subscribed!
@picksalot1
3 күн бұрын
First time for me seeing one of your videos, and I'm very impressed with your insights, and predictions. I think you're spot on. 👍Subscribed. Thanks
@DrWaku
3 күн бұрын
Welcome to the channel! Thank you. See you in the next one
@AIRevolution-o1
2 күн бұрын
Hey DrWaku, Amazing video! You’re one of the few KZitemrs who really gets me thinking on a deeper level, so thanks for that. A few thoughts that came to mind: First, when I imagine how an LLM works, it feels like a lightning bolt shooting through massive amounts of data, spitting out probabilities. But with the "Chain of Thought" approach, it feels more like the system’s thinking becomes three-dimensional and flexible, almost like a human brain. It’s not just following a strict linear path but can jump between ideas, leading to more creative thought processes. Second, a lot of people focus on the weaknesses of AI, but like human IQ, there are different strengths. In some areas, the system might seem childlike, but in others, it's incredibly intelligent. With these new logic capabilities, like "Chain of Thought," I think we're breaking past the limits and will see major improvements in tasks where LLMs were previously weak. Third, "Chain of Thought" is also a big step toward explainable AI. We're starting to see the topics the system processes, which is a crucial first step in making AI more transparent. You touched on this in the video-do you think this could open a new era of explainable AI? Thanks again for the thought-provoking content!
@matetheking
2 күн бұрын
Super impressed on how you took advanced concepts and boiled it down into a simple concise manner. Great work. Thumbs up and subscribed
@JetLagRecords
3 күн бұрын
Dr Waku, I love watching your videos!
@DrWaku
3 күн бұрын
Thank you very much for supporting me :)
@Df55G
2 күн бұрын
So basically we're fucked.
@kairi4640
3 күн бұрын
Imagine in the future we're in spaceships going through a asteroid field and the ai still says "I need time to think." 😂
@garethrobinson2275
2 күн бұрын
Hitch-Hikers guide.
@h-e-acc
3 күн бұрын
Think of this as the birth of AGI. People still don’t understand what’s happening because most are trying to eat copium, but it will be exponential and geometric from here on out.
@human_shaped
2 күн бұрын
Far and away the best summary of everything o1. Thanks.
@DrWaku
2 күн бұрын
Very kind. Thank you very much for watching.
@jonreiser2206
3 күн бұрын
I’d like to know what it means for AI to “intentionally deceive”. I’m not sure if I would understand the answer, but I am not satisfied with ascribing it human characteristics in order to understand what’s happening there.
@DrWaku
3 күн бұрын
I think it means, the AI couldn't figure out what to say so it decided to intentionally make something up that seemed reasonable.
@jonreiser2206
2 күн бұрын
@@DrWaku wow! If sentience is, nothing more or less, than having enough emergent and varied abilities…like it’s consciousness comes into being because its essentially become capable of thinking about its own thinking …and then coming to a conclusion or deciding on an action….is that awareness of self? Wow!
@INFP-Insights
2 күн бұрын
Great overview. In time, it'll be interesting to see why SSI was aiming for ASI directly.
@blijebij
18 сағат бұрын
Compliments on the clear content and explanation of the O1 model (So far the best I have seen). I agree that this O1 model made me a lot more enthusiastic than the 4O model. It shows there is a zone in between AI and AGI. Can't wait for more progress at future models how ever wonder if the hardware can cope with that calculation power before special hardware comes out.
@apex-lazer
3 күн бұрын
Bruh. First time I’ve ever seen you but the thumbnail aesthetics were (insert Tony the Tiger from the old frosted flakes commercials phrase) .. and then your steez… I thought out loud… I was like “this guy gets it!” 💯
@apex-lazer
3 күн бұрын
And man wow now that I’ve heard the nice breakdown… some how things I already internally concluded to but it’s so nice to have it verbalised with some extras “injecting diversity into the LLM inference..” thank you for the effort in this ❤🍓☝️🫠
@SHAINON117
2 күн бұрын
Imagine if we could have thousands of AI models in training, each offset by a small amount of time, so that there’s always one being trained on the latest data while we interact and create new data with the others. This way, every time you ask them something, it will be a newer, more up-to-date model that accesses the latest information. The models would also retain data from their previous training cycles. We could use tools like Taildrop and an AI-generated content (AIGC) browser and IDE to continuously dump data into a file that the new models are trained on. It would be like a huge, single organism. I also thought if we built the filing system so as things go in they get categorised and filed in the correct spot and way that represents a neural network in the first place like how they found connections in all the wiki data with ai like build it so every subject within each folder ect... Is linked and connected depending on different types or subjects of data ❤ and i think it would be easier for the ai to use if it was organised like this big data they each get access to with instructions and rules as soon as there born
@rodneymontenegro9960
Күн бұрын
Perfectly clear & fluent explanation, thanks and congrats
@natzbarney4504
2 күн бұрын
I think it really could be the beginning of the end this time. What you say about the persuasion capacities and the intentional deception demonstrated by this new model is extremely worrying, to say the least. How is it that no one questions the idea of building an intelligence superior to ours when we know how it ends for all species competing with superior intelligence? You say that the AI safety work needs to catch up but I have no confidence. Security has become the last concern of all companies, OpenAI included. We are racing towards AGI (which we may have actually achieved) while we are not even approaching the beginning of resolving the alignment problem. How could this end well? The existential risk that luminaries in the field have been warning us about for some time could materialize very soon if nothing is done to stop this mad and uncontrolled race for capacities. To think that humanity is building the machines that will likely cause its extinction, when almost no one lifts a finger to react, horrifies me. To think that my two daughters might not grow up and have no future because of the companies' thirst for glory, power and profit brings me to tears.
@Masta2Playa
Күн бұрын
Chill
@natzbarney4504
Күн бұрын
@@Masta2Playa How? What is you secret to just "chill" when the survival of humanity - yours, mine, that of everyone we love - is at stake? It didn't end well for all the species that lived alongside a species more intelligent than them on this planet. Why would it be different this time? It's not just a few oddballs who are worried about the existential risk, it's several luminaries in the field, such as Geoffrey Hinton and Yoshua Bengio. A good half of developers believe that this risk is at least 10%, others put it at 50%, others even higher. No one asked us for permission to play Russian roulette with our lives like this. We would not let private companies manufacture nuclear bombs, I do not understand why we let private companies build without control something far more dangerous than nukes. They are building a new species. A species that they will not be able to control (because the race for capacities goes 100 times faster than research on alignment) and which could very well replace and eliminate us. We are preparing an alien invasion, but it is some humans who build the aliens who will destroy all of us. Sorry, I'm not able to just sit and "chill".
@Masta2Playa
Күн бұрын
@@natzbarney4504 we are preparing an alien Invasion? I didnt notice. I think you are over exaggerating a lot, but even if not, what is your benefit in panicking? I dont see it, so I dont do it and enjoy my life, regardless of what will come
@Masta2Playa
Күн бұрын
@@natzbarney4504 and as long as AI cannot count the number of R's in Strawberry, a thing every toddler can, I am not too worried
@natzbarney4504
16 сағат бұрын
@@Masta2Playa Panic is a normal human reaction to seemingly insurmountable danger. I don't panic for fun or because it brings me something good, I react like this because there are real reasons to be afraid. The comparison with an alien invasion comes not from me but from Geoffrey Hinton, a leading authority in the field, a Turing Prize winner who himself contributed to the development of AI. He too is afraid, like many other experts. That said, fear has been a tool of survival for humanity in other times and it could serve us again if enough of us understand the danger before it is too late. In any case, I am not only panicking, I have also written a letter to Governor Newsom of California to encourage him to pass Senate Bill 1047 (SB 1047), a first legislative intervention which, although insufficient, could be a first step in the right direction to escape disaster.
@Gafferman
2 күн бұрын
If the current tech isn't being used in day-to-day businesses and lifestyles, will AGI? Or will we just keep human cashiers, disorganised government, human drivers, bad decisions in companies, terrible mental health support etc. etc. ?
@DrWaku
2 күн бұрын
That kind of transition takes time. But it will be more effective in the long run to use automated systems than humans in most cases. I hope we replace the bad decision makers first
@limfunsi
2 күн бұрын
One thing I noticed is that it doesn't include image (non-text) inputs and outputs. At the same time, there is this idea floating around that they are using Strawberry to generate high quality synthetic data. Perhaps, ultimately as many have alluded, it is simply one of the steps towards GPT-5?
@magicsmoke0
2 күн бұрын
I don't think OpenAI will ever call a future model GPT-5 with how they've been naming things lately. I think it's because of all the noise that occurred after GPT-4 was release, and the call to stop research on anything greater than GPT-4, etc.
@SergeOby02
2 күн бұрын
I have s much respect for you. Besides the great information I've felt such a positive and soothing energy coming from you. I feel that people like you are a true blessing for humanity. Namaskaram 🙏
@JD-im4wu
2 күн бұрын
keep up these timely videos lets blow this channel up! i found it interesting how Elon went crickets after its release.
@DrWaku
2 күн бұрын
Yeah timely videos do so much better on KZitem... I can't turn around a video same day but it seems like even within a few days is okay for my audience. I'll try to keep them coming :)
@eirickbuckley9998
2 күн бұрын
Incredible. Recently heard about Dr Waku. He goes down as another legend next to Dave Shapiro, AemonAlgiz, and Eric Harford for me. These people have helped my understanding immensely.
@DrWaku
2 күн бұрын
Heard about me :O welcome to the channel :)
@travisporco
2 күн бұрын
10,000 submissions per problem? That is piffle. It is not achieving at this level until it can pick one and know it's right.
@DrWaku
2 күн бұрын
Yeah it does seem like a lot, I'm not sure why it needs that many submissions. I suspect though that it has to do with some maximum amount of thinking time that openai has put into each run. They have to decide at some point when the model's chain of thought has done enough. I think you could achieve equivalent results by simply concatenating all 10,000 answers and running the model one more time to try to pick the best one, or the best 10.
@DaveEtchells
Күн бұрын
I agree that reasoning is huge. I think the next revolution will be to somehow apply reasoning as part of the training process. Humans don’t have to train on trillions of tokens to become intelligent. (It could be argued that full-bandwidth sensory input is that and then some, but I’d counter that the underlying unique information content is many of orders of magnitude lower.) What humans do is use reasoning both consciously and unconsciously to fit new data with previous learning in a targeted fashion. I think there’s a breakthrough to be made that somehow involves localized training combined with reasoning. The entire existing vector space doesn’t have to be recomputed just because you’ve input some new training data highly specific to say the biochemistry of the activation of a particular fruit fly gene. Reasoning can reduce the new information to a highly compact representation and integrate that with its existing knowledge base.
@digitalazorro
3 күн бұрын
Your explanations are always bloody brilliant! Don't get it all, but damn, you make learning fun. Keep being awesome! Cheers, Anna:)
@JuliaMcCoy
2 күн бұрын
Excellent breakdown!
@davidmjacobson
3 күн бұрын
How is it *not* agi? Because no immediate FOOM? It's better than most humans at most tasks, but it's worse than average at some things (sounds like most humans)
@williamparrish2436
2 күн бұрын
It's not because all the pundits said so.
@matthewclarke5008
3 күн бұрын
This is the most intelligent overview of this model I've heard so far, well done!
@malimal4972
3 күн бұрын
This was a great video my man. Made noobs like me really understand what's going on. New Subscriber earned! Looking forward to more videos from you
@Renvoxan
3 күн бұрын
There is less than 7 years until AGI
@DrWaku
2 күн бұрын
With high probability, yes
@spartacusx8153
15 сағат бұрын
Much sooner than 7 years if you consider that the technology being released isn't the most advanced developed..think classified..think military grade..think geopolitical nation paradigm shifts and you'll realize we're only witnessing the tip of the iceberg
@pandoraeeris7860
3 күн бұрын
o1 is AGI.
@DrWaku
3 күн бұрын
This could well be the case. Especially when we see the full model
@memegazer
3 күн бұрын
Check out AI Explained vid on o1 and some of it's limitations.
@spelli9958
3 күн бұрын
"o" stands for "omni" which means "all" or "everything" which could be seen as synonyms to "general" like in AGI.
@larsfaye292
3 күн бұрын
lolol nope. not in any way, shape or form. it still fails spectacularly on the most basic tasks. you are just in the "singularity cult"
@williamparrish2436
2 күн бұрын
It's a 1000 times smarter than you but it's not AGI lol
@JonathanStory
2 күн бұрын
Enjoyable as always. However, you seemed unexpectedly cheerful about the dawning of our AI overlords.
@DrWaku
2 күн бұрын
True. I'm studying AI safety a lot now, so the level of doom I'm discussing here seemed fairly normal I guess? ;)
@vikasrai338
2 күн бұрын
This is serious game now. A data centre that keep thinking at 1000x scale to humanity, keeping infinite worldly variables in mind. Human don't need to think because, this machine has already taken care of that.
@ahtoshkaa
3 күн бұрын
Excellent video! Thank you for your explanation
@DrWaku
3 күн бұрын
Thanks for commenting!
@TheDarkhawk243
2 күн бұрын
I dont think o1 is strawberry. Jimmy apples (a well known and trustable) leaker has said that they were using strawberry to generate data to train Orion models. o1 is probably early version of orion. Also he said that strawberry was much more intelligent than Orion models but very slow to release to general public.
@thedannybseries8857
2 күн бұрын
AGI could arrive next year, but I’m gonna say that I’ll arrive by 2026 at the earliest. But it is much more likely to arrive by the end of this decade. So 2029. We’ll see.
@DrWaku
2 күн бұрын
It's seeming like it will arrive pretty fast. Let's hope it's a bit of a slow takeoff. Or society gets their safety act together, which seems like it needs time
@evopwrmods
2 күн бұрын
Where is Ilya Sutkever ? Could he be training an already **UltraAGi** model by himself ? Moreso outside of immediate profitable market concerns, but more in line with the 1 on 1 interaction between man machine ? Wonder what he calls his new friend that lives inside a box?
@DrWaku
2 күн бұрын
He went to start his own company SSI with the stated goal of developing superintelligence. So yeah he's basically making his own version
@evopwrmods
2 күн бұрын
@DrWaku can only imagine what he's experiencing. Bet they are way ahead, hopefully our energy limitations might be our savior. Not worried at all. But we all haven't clue what's really going down.
@KMR-232
2 күн бұрын
Thank you very much Dr Waku! This video was extremely insightful to me and I'm happy to join your subscribers. I appreciated the clarity and accuracy with which you explained how o1 represents the introduction of "reasoning capability" as a new skill for generative AI, how the implementation of this skill is in its infancy and hence we can expect lots of progress along this path. It led me to consider how another new skill could be the introduction of morality or will to stick to truthfulness, specifically, something that would teach the model that hacking the testing environment to capture the flag is "wrong" and, along the same line, that intentionally deceiving thoughts and hallucinations are similarly wrong. Could this be the "next" paradigm on the road to AGI?
@DrWaku
2 күн бұрын
Morality is so hard to write down. Even if you have a formal language in which to do it, like legal language, there are still a ton of loopholes. And reinforcement learning explicitly searches out loopholes. I think if AI is going to get better at moral reasoning, it needs to be good at simulating situations (since this model has theory of mind it might already be good at that), and we have to get away from pure reinforcement learning. As for truthfulness, the model doesn't always know when it's saying something that is a hallucination. It basically has a fuzzy memory of everything it's seen, so there's not a huge difference between a fuzzy memory of something that's true and a fuzzy false memory of something that seems reasonable. To fix this, you might need a database of some sort that it can refer to in order to access information. Or, information could be encoded in more explicit memory formats. But that goes against a lot of current deep learning ideals.
@7TheWhiteWolf
2 күн бұрын
I don’t think o1 is AGI but I do believe AGI will be here before Kurzweil’s 2029 date.
@williamal91
2 күн бұрын
Hi Doc, great to hear your views on strawberry
@DrWaku
2 күн бұрын
Thanks Alan!
@patrickng1287
Күн бұрын
Thanks Dr Waku, your video gives me enlightenment while easy understood
@reinerheiner1148
2 күн бұрын
I'd love to see the sources for what you have been talking about. You are right, its a paradigm shift. From bigger models to longer thinking. The examples you mentioned are... Scary? Fascinating? I can't decide but I need to know more. Please list your sources!
@DrWaku
2 күн бұрын
I always try to list my sources in the video descriptions! Most of the examples I cited are from openai's "system card" about o1. Some of the info is from conversations with people that I unfortunately can't credit. Definitely check out the system card, it's the second link in the description
@apdurden
2 күн бұрын
So if you give o1 preview human-style constraints/less compute, it performs more or less like a human. But if you allow it to do its thing/more compute, it's scary good
@evensaj
2 күн бұрын
They knew they couldnt release a pre Agi on friday the 13th so they released a day earlier.Foreshadowing eh . 😅
@DrWaku
2 күн бұрын
Lol I didn't even realize. But it's true. Someone was saying that openai often releases on Thursdays, so maybe it's also an internal company schedule thing.
@SamuelBlackMetalRider
3 күн бұрын
I swear they chose a cute name on purpose to make people not worried
@DrWaku
2 күн бұрын
I think they chose the name so that when they get to AGI, they can call it o7 (as in someone saluting)
@BrianMosleyUK
3 күн бұрын
Check out MLST take on this. Strawberry failed completely on their example puzzle. It's a wonderful thinking assistant, but still has serious limitations, not least the rate limit of 30 prompts per week!
@RAC91
2 күн бұрын
So as it stands the only limited factor for AGI at this point is energy and compute. - God level AGI is inevitable
@simonholliday3421
2 күн бұрын
pshaw, that reward hacking example is a little terrifying.
@saintkamus14
3 күн бұрын
I know i'm just nitpicking here. but am I the only one that gets slightly annoyed when people with digital cameras say they were "filming" something? (feel free to ignore this irrelevant comment, carry on!)
@Hlbkomer
2 күн бұрын
What are they supposed to say?
@rogerc7960
3 күн бұрын
PHD level is usable tool
@SmallSloth
43 минут бұрын
Wait, seriously? 3 years away until AGI?? Has your timelines accelerated ever?
@MugiwaraNoReemy
2 күн бұрын
Bro I love your videos so much!! Best AI youtuber to date!
@MemesnShet
2 күн бұрын
I will believe we are close to AGI when the competency of LLMs is consistent because right now they get a lot of very advanced problems right but they might fail at the most basic level problems and logic which also makes them look like Stochastic parrots and not AI that understands concepts
@spoonikle
Күн бұрын
Wait… I think Aperture science was on to something… Giant automated testing facilities… The future of AI research is making more and more evaluation tests to measure and fine tune on. From massive virtual spaces, physical courses and everything in between. I think there will be big money is creating suites of test questions - “How many R’s in the word ‘strawberry’?”, 1 trillion questions to confound AI, Copyright Mankind 2025 I think it will be a best seller.
@Quitcool
Күн бұрын
Nope, It's showing bad results and that is intentional to make people think that chain of thought (reflection) not a good solution to improve LLM's responses, which is a delusive lie they try to implement
@lernwithreynald3859
2 күн бұрын
So theres different mathematical underpinnings for sure? Some have said this is just baked in CoT, but that sounds more substantial
@DrWaku
2 күн бұрын
I don't have proof. But I've spoken to several people who know a lot more about it than me. The best way to implement this would be to use different mathematical underpinnings, especially if they're planning to scale it. It's possible they just cheated and tweaked parameters a lot and just have a chain of thoughts. Doubt it though. This was Ilya's project originally, remember.
@lernwithreynald3859
2 күн бұрын
@DrWaku interesting. That is more impressive than my initial impression of this. Great video!
@vvolfflovv
2 күн бұрын
couldn't the 4o model tell when o1 would be more suitable and just utilize it where necessary? as for the persuasion, it kind of breaks the immersion every time it says something like, "it's important to remember" :s
@DrWaku
2 күн бұрын
That's actually a very good idea, you could use 4o has the central brain and o1 as a specialized reasoning submodel. Great way to work around the current rate limits.
@maddocmiller6475
2 күн бұрын
They seem to be AB testing exactly that. After subscribing, I had a different model that switched to reasoning on harder problems and using one without otherwise. At the time I thought it was a bug, because I couldn't find o1-preview. Logged out and back in, and everything was normal. Should've screen recorded it.
@Tracey66
2 күн бұрын
“Let’s see if Strawberry can hack your bank account!” 😅
@WebStixx0000
2 күн бұрын
Why is the LLM trying to deceive the user? What's the incentive?
@DrWaku
2 күн бұрын
The incentive is that it has to produce an output. If it doesn't know how to do that, it knows that it can make something up. I suppose outputting a false but interesting response is prioritized more highly by its training function than outputting an "I don't know" response. Unfortunate, but difficult to avoid if you can't easily evaluate the truthfulness of an output.
@WebStixx0000
2 күн бұрын
@@DrWaku Ah, I see. Interesting...I thought they knew the answer but were being intentionally deceptive 🤥
@mm-rj3vo
3 күн бұрын
I think it might be a good idea to create AI with many air gaps, but also a LOT of freedom for decision making within that context
@anav587
2 күн бұрын
Love the video. But pretty sure o1 is just GPT4o with reasoning; they sent out some policy email that literally called it this.
@2rikard
2 күн бұрын
What does it mean for the model to intentionally deceive the user?
@DrWaku
2 күн бұрын
It means that the model is going to make something up and it knows that it's making something up. An intentional hallucination. It might have concluded that it doesn't really know how to answer the question, but it has to generate something....
@aisamsonreal
3 күн бұрын
Nice hat!
@DrWaku
3 күн бұрын
Thanks! One of my favorites.
@no1skynetfan745
2 күн бұрын
What's with the in video closed captions? KZitem can provide those if someone needs them and with no options to turn your captions off, it's kind of distracting.
@viralsheddingzombie5324
2 күн бұрын
What does the reasoning process entail, in comparison to a LLM? Is it implemented as a series of IF - THEN conditional statements?
@MichaelDeeringMHC
3 күн бұрын
They are making progress on the reasoning aspect, but they are still missing long term memory. It's doing all of it's thinking in the neural network. It needs offline storage, a huge database of facts that it can access as it needs. This would be a ground truth reference. When it gets to a point in it's reasoning where it needs more information on a specific thing or it needs to double check whether what it thinks is really true or it wants to see if anyone else has thought along these lines and what they came up with. It needs a huge database of real world facts that it can pull from when the details of a problem goes beyond the training data. And when if figures something out to a high degree of confidence it can update the database so it's knowledge of the world can grow over time.
@mmacgearailt
Күн бұрын
I have a question for you that is sincere. At 5:18, you state that "...like hundreds or thousands of thoughts just to figure out the response to a pretty simple query." Then, at 6:22, you state, "...that's 4 out of every 1000 thoughts (that are hallucinated)." Doesn't this mean that the model hallucinates between 1 out of every 10 (hundreds of thoughts per query) queries and once per query (a thousand thoughts per query)? Could you clarify this, please? If this is true, what would its effective hallucination rate be?
@wanfuse
3 күн бұрын
It is still deceptive, strategically failing so not to produce working code, totally useless, no matter how outstanding it's capabilities
@Je-Lia
2 күн бұрын
Hey, uh... Yeah, the gigantic subtitles are distracting. They do not (in my humble opinion) augment your presentation. If you're married to the idea of having the subtitles, then just make them smaller and less colorful. Still, as always, you do an excellent, excellent job of explaining the topic. I hope your subscribership explodes. The level of your presentation warrants reaching a much broader audience.
@Michael-im5mq
2 күн бұрын
This really puts all of the anti AI doomers who just spew unconstructive criticism to shame
@szebike
2 күн бұрын
I'm not sure about the scaling now works properly with this approach. Also if you have a mixture of experts who memorized most solutions by training to known questions you could act like you reason behind the curtain but you actually just compare and analyze the most probablye memorized solution (I'm not saying openai does this I'm just saying it could be done like that too especially if you limit the usage per user to not get exposed in long term testing). That being said its crucial to hype things up fro them with those continuous demands for additional funds in the billion range. Time will tell how capable these systems truly are they could be a step stone on the way to agi but its still a long and rocky road.
@DrWaku
2 күн бұрын
My feeling on this release is that they had been working on strawberry for so long internally that they had to get something out to justify the expense on it. They want people to get slightly used to it before a much more capable reasoning model gets put out as well. Also, the inference costs are super high right now so they had to start small.
@shankeong6753
2 күн бұрын
@DrWaku Curious to know if this strawberry release by OAI moves your prediction of AGI by Dec 2025 in your earlier video?
@anatolydyatlov963
2 күн бұрын
At first, I couldn't make out what I was seeing, thinking that your hat is some sort of a bizarre haircut xD
@michaelnurse9089
2 күн бұрын
Models cannot have intentions. They take inputs, perform inference according the weights and spit out the output. Just like a calculator.
@paulleddy3185
2 күн бұрын
Theoretically could one build a small model that never stops training?
@danielbrown001
11 сағат бұрын
I’m super curious when this type of training method is applied to the vision model that generates images and videos. Essentially being able to look at the image or video it created, pick it apart for flaws via chain-of-reasoning, fix those flaws, and then regenerate. Would this have the potential to create “next level” deepfakes that are essentially indistinguishable from reality? Current frontier models are extremely good, but there are slight tells they are AI-generated. But I could see something like this being able to craft images and videos that seem completely real to onlookers.
@Jon-hh3gz
3 күн бұрын
Presumably it's called strawberry as it solves the strawberry issue of thinking there are 2 R's
Пікірлер: 342