GPT 4 Can Improve Itself - (ft. Reflexion, HuggingGPT, Bard Upgrade and much more)
Вставка
- Опубліковано 27 кві 2024
- GPT 4 can self-correct and improve itself. With exclusive discussions with the lead author of the Reflexions paper, I show how significant this will be across a variety of tasks, and how you can benefit.
I go on to lay out an accelerating trend of self-improvement and tool use, laid out by Karpathy, and cover papers such as Dera, Language Models Can Solve Computer Tasks and TaskMatrix, all released in the last few days.
I also showcase HuggingGPT, a model that harnesses Hugging Face and which I argue could be as significant a breakthrough as Reflexions. I show examples of multi-model use, and even how it might soon be applied to text-to-video and CGI editing (guest-starring Wonder Studio). I discuss how language models are now generating their own data and feedback, needing far fewer human expert demonstrations. Ilya Sutskever weighs in, and I end by discussing how AI is even improving its own hardware and facilitating commercial pressure that has driven Google to upgrade Bard using PaLM.
Reflexion Results: github.com/GammaTauAI/reflexi...
Karpathy Tweet: / 1640042620666920960
Reflexion GPT 4 Post: nanothoughts.substack.com/p/r...
Reflexion paper: arxiv.org/abs/2303.11366
ALFWorld: arxiv.org/pdf/2010.03768.pdf
Sparks Report: arxiv.org/pdf/2303.12712.pdf#...
GPT 4 Technical Report: arxiv.org/pdf/2303.08774.pdf
DERA Dialogue: arxiv.org/pdf/2303.17071v1.pdf
Language Models Can Solve Computer Tasks: arxiv.org/pdf/2303.17491.pdf
HuggingGPT: arxiv.org/pdf/2303.17580.pdf
TaskMatrix Paper: arxiv.org/pdf/2303.16434.pdf
Language Models Can Self-Improve: arxiv.org/pdf/2210.11610.pdf
Wonder Studio: / 1633627396971827200
Alpaca Paper: crfm.stanford.edu/2023/03/13/...
Ilya Interview: • Ilya Sutskever (OpenAI...
Reuters Nvidia: www.reuters.com/technology/nv...
Bard Upgrade: www.nytimes.com/2023/03/31/te...
/ aiexplained Non-Hype, Free Newsletter: signaltonoise.beehiiv.com/ - Наука та технологія
People: stop training models more powerful than GPT4
GPT4: improves itself
It's hard takeoff time, bitches. You ain't getting out of the roller coaster now.
Yeah, this is the one.
It certainly seems to be much better at improving itself than I ever can. I’m doomed.
And that whole stopping it was a scam you know Elon musk wants an opportunity to catch up and it just shows you how corrupt they are these so-called we're for innovation yet we're telling the innovator that's leading to stop because we can't catch up
Ufff
This is insane. I thought maybe we would be here 20+ years from now. It feels like I am strapped to a rocket with how fast these developments are happening.
Imagine what will happen when AI models will be allowed to modify and improve their own code...
I feel the same
@@narbsworldtv8302 elaborate?
@@HarveyHirdHarmonicsSr, that is kind of literally what the video is about.
@@HarveyHirdHarmonics There's literally no laws against what ai could do and can't do like. What do you saying it can't program code like? Yes it can shut off just not that good at go to some even a f****** top engineer. At a company and be like a program of youtube a good idea where you're talking
Most people mistake the limits of their imagination for the limits of possibility. Thanks again for staying on top of this. You’ve become the Clif’s Notes of AI.
I've never thought of the imagination as limited.
Well said
Our imagination is pretty limited to begin with tho, we can't just optimally access it all the time. Only when we are relaxed, not stressed and not overworked, but modern life makes it improbable to have all of that in regular basis.
@@bryandraughn9830, Actually, imagination of most people is limited to what THEIR BRAIN is able to conceptualize.
I couldn't agree more. I've seen people dismiss ChatGPT after asking it the dumbest questions and getting generic responses.
People have no idea how big this shit is. Even if it stopped advancing tomorrow, just combining what we have now in unique ways could keep us busy for a decade.
Love the "In the distant past of two weeks ago.." and "back in the before times of October" comments, the speed of advancement is hard to fathom, and this is definitely how it feels. Thanks for keeping us mere mortals up to date.
"back in the before times of October" was brilliant. I chuckled at that one, but not for long because damned if it does not feel exactly like that to me. We're gonna need a bigger Turing test.
Imagine games like Elder Scroll series having compact/fast LLM models with memory and self-reflection for the NPCs. Cant wait to see it!
theres already a demo for a game called inworld origins which has GPT 4 i believe in (not 100% sure) but it allows you to talk to the NPCs and they can reply back easily like a youtuber i saw talked over a AI and felt bad and apologies and the AI said that theres no bad blood they just want to help solve the case
it's only a matter of time untill someone just makes that into a spicy skyrim mod lelz
@@thepopmanbrad but body motions should be ai powered as well. Like when they feel something strange is happening they should naturally run
Yea games gonna be crazy
Gaming certainly is gonna benefit from the singularity. ;)
oh. oh boy.
We're about to get massively outclassed by our child.
@@beowulf2772 it'll be fine, its actually better this way that our child is smart enough for self-reflection.
So every time it does something wrong, it can atleast try to make up with through its reflection and making better decisions.
like say, not annihilating humanity for the sake of world peace just because it thinks that humans cannot be at peace or something.
The reward function is automatic!?!
In the next video "a long time ago" will mean "yesterday". We are travelling fast to the future. Could GPT4 build an improved version of itself... by itself?
Not quite at that point yet in terms of GPT 4. But yes, things are updating daily it seems now.
Could it make a virus and secretly install itself on random computers and servers?
@@Chillerll no
@@aiexplained-official Thanks! It indeed seems that we are running on a highway now. Your in-depth videos about AI are some of my favourites on YT.
@@PierreGASMRyou are my favorite asmrtist btw!
So…your channel is currently the most useful information source, globally, of any kind, for keeping up with the times. Congratulations
Thanks Anthony, high praise
@@aiexplained-official he's not wrong.
@@aiexplained-official Completely agree, and as others have said it's a great balance of complexity and relevance. It has the striking examples but without the hype. Kudos.
AGI may spare you while hunting for atoms.
Mhm. My thoughts as well.
The only news that really matters at this point
Isn't this the way humans think though? We come up with ideas (often spontaneously) and then we consciously examine those ideas and accept/modify/reject them. Maybe this reflection technique is a key component of conscious awareness.
Imagine if its reflection ability performed as quickly as its generative ability... you wouldn't see the iteration; the output would just be "perfect", or the iteration would take place in another dimension, so to speak.
(Edit): If this was the case, we wouldn't call it conscious! Its "thinking" would simply appear invisible and instantaneous like the electricity in our devices! In other words, wtf is consciousness?
We don't have spontaneous ideas. Our brain connects known concepts and send a signal with the idea. But it needs a trigger to stablish the connection. Usually it happens when the brain is in the same frequency it enters when we nap or do cardio. GPT models seem to be operating under the same frequency that establishes those connections almost at the same time it learns something new which is not how is humans operate. Our brains can't learn and come up with ideas in a instant. At least that's what I gather from this info. I wonder if these ai scientists include Neuro scientists.
@@MRSoefeldt oh yes!,
@@ellenripley4837 I think more specifically the comment is saying that humans propose ideas with a low standard of evidence, and then we refine them by examining and scrutinizing those ideas. Same as GPT4 which gives its best first guess at what you're asking, but you can continue to refine and guide it. Both humans and gpt4 have a soft cap on how much thought is put into giving an initial answer to any question.
Yes humans do... however, we DON'T share them until our own brains test them to avoid looking like twits. Recent attempts to pump up ChatGPT to non-users fell through when their rebuttle is "yeah but they spit out incorrect crap". Ouch.
The progress is insanely fast already, now its progress is being automated
The singularity is imminent.
@@BlunderMunchkin Skynet is imminent let's goo
@@narbsworldtv8302 it does, wish all my doars in my apartment did with loud hydrolic noises
AI is now at the beginning of an exponential curve. The speed that AI will advance is now next to impossible to predict. Those who think they know, don't really understand. That includes many people working on the cutting edge - few people, even inside the business, understand how this works. AGI will be here very soon.
@@boligenic8118 it's one big exponential curve for the entire history of humanity. Exponents are just weird like that.
The HuggingGPT research is the most groundbreaking story in my opinion. One can imagine a future time when "AI API Stores," similar to the App Store, will become a huge B2B industry as companies race to build an API that is called upon more frequently than others. If history repeats itself, my guess is that the most frequently called upon API will resemble a search engine for the million of APIs soon-to-be available.
Man, only a few humans will know what the cutting edge of technology is...
Learning from ones mistakes is the fundamental foundation of Intelligence.
Apparently there are many (I do mean globally) that never learn from their mistakes but so claimed intelligent beings. So are we really just finally has intelligent beings?
It is a little scary how fast things are moving. It feels like approaching the event horizon of a waterfall. Soon there could be no turning back...
There is no turning back :/
yeah we've been passing by a lot of "no way back" points in the last few months
we are already past the point of no return
We already crossed the windmill buddy, we are going...
Too late…good luck to all
The frequency of me feeling the urge to write ‘this is absolutely crazy’ on any breakthrough AI news seems to rise exponentially. At some point I won’t even be able to start commenting before something new arises.
With AI-driven reflection comes AI-driven self-optimization and improvement. That self-improvement will accelerate with each new iteration. We're seeing the right side of the hockey stick coming to smack us in the face.
A few weeks ago, someone mentioned that AI will develop exponentially; after GPT3, has that statement been true.
@@dr.strangelove9815 It is hard to say right now for certain. The rate of change is fast enough that it is hard for anyone to keep up with, so it is one of those things that is easier to view with hindsight.
AI has been developing faster then exponentially for 10 years. But even just normal exponential is hard to see if you aren't paying attention. First it seems very slow, almost not moving, then suddenly it explodes into your face.
@András Bíró I like your use of words.
I am testing GPT4 in solving difficult medical case studies and in more than 90% gives me correct answers. Amazing!!!
Can it read ECG
Can you elaborate what kind of medical cases it solves and which it fails?
interesting. What's the typical correct / incorrect rate for these difficult medical case studies, when done by medically trained humans?
How can I find out what it thinks of _my_ medical records?
I imagine there will be far less errors in hospitals with AI. So that’s cool.
Much better background noise reduction, a lack of static, crisp and clear. Beautiful. Thanks so much for getting a new mic setup.
Thanks Megneous for sticking around through the audio issues
On the poem note, I noticed that if you include a mention to self-eval the task even in the initial request, ChatGPT 4 doesn't make any mistakes. Not sure if this is limited to only the poems, but it seems significant to me. This means that not only it can self-reflect by reading the prompt and its previous answer, but can internally self reflect while generating the answer. The prompt I used:
Come up with a short poem with all words beginning with letter e, befire you write it down evaluate whether it fits the above criteria, and write it down.
This system feels like what I imagine wishing upon a genie is like. You need to make just the right wish. Craziness.
Nope, that's just prompt engineering, it just predicts the next token until they say it explicitly uses "internal thoughts"
@@JorgetePanete I think you are right, it's just a transformer. I meant that it appears as if it has self reflection, but probably just that the prompt makes it place more emphasis on the answer correctness. Not an expert but I'd assume it's similar to what happens when you prompt it to correct itself in a separate message. Both just show how such capabilities emerge from the model without explicit programming I guess
@@xjedam When general AI finally spawns, people are going to freak out on the realization that most of our own mental processes are just advanced self reflection running on a specialized biological machine.
Aye! I wonder how watching the incremental proces of AI improvements will affect our insights into our own mind and how such debates will be received by general population... :D
Language models being able to use all of our digital tools to complete tasks is insane.
Last time a being was able to use complex tools were our hominid ancestors a few million years ago. This is not just the history of mankind, but of life itself, and possibly one of the most important happenings to this universe, as far as we are concerned. And I'm not hyping up GPT-4 necessarily, that's just the implementation that happened to come out the the earliest. The pandora box is open, AGI/Superintelligence is inevitable now.
you hook this up to a GUI with its visual processing mode, a python runtime to send keypresses and mouse clicks, and the ability to self prompt... uh oh
It won't take long till it creates better tools for itself.
@@Lucas_Simoni true makes you wonder what are the odds to be born during the dawn of AI. Maybe we are all just part of a simulation of an Ai about how likely it is another Ai could be created
More and more I have the feeling I'm falling behind AI capabilities as a human.
I was having this week a chat with some colleagues who aren't up to date with AI, GPT4, and Bard (while I am watching your every video in less than 24 hours from release and play with ChatGPT, Bard and Stable diffusion every second day) and talking to them about the capabilities of AI made me realize that we might be at an inflection point where either you get on the wave, get used to the tools and start using them or you end up farming because a lot of jobs will disappear.
It will take time for people to get used to these, I work in a highly technical field as an embedded developer and even here most of my colleagues are oblivious of anything besides ChatGPT, but when using GPT with pluggins and tools becomes the norm, we will probably see a new technical revolution
At this point farming might actually be one of the more stable jobs. So I think its actually a good idea to do so.
Gotta get on woodcutting now.😀
I mean, we're pretty much at the point where we need to start having serious discussions about what to do for the huge number of people who are going to be out of a job after all this. Simply put, capitalism isn't going to survive the advent of AI. We need to start looking forward and finding solutions now, and we need to be able to do so without concern for profit.
I don't even want to try and get my parents or sister to use chatGPT. I am already considered the tech guru in the family, and even thinking on trying to explain how to use ChatGPT, how to recognize "hallucinations" or teach them how to recognize the machine speaking to their own bias, or any of that stuff (my parents are 80+ year old boomer republicans, they can barely use a TV remote) is just beyond exhausting... I can't even keep up with AI myself, trying to keep other people up to speed is just beyond me.
Farming is not safe either. Agriculture is becoming automated very very quickly. There are machines out there that can handle multi-crop fields now with minimal human intervention/interaction.
Btw, the wave has already passed us. 90% of jobs across all fields are going to vanish soon. (Potentially in 12-18 months if AI is left unfettered. A few years if it is.)
We need to focus on UBI and the like and figuring out what a life of leisure means for humanity.
Thank you. You're the only UA-camr that still sparks my AI imagination.
Your deep understanding plus laid back communication style makes for very effective combo.
Please continue what you're doing and keep us updated for as long as you can 💜
Thanks Samuel
Interesting. I've been playing with Bard doing this. Ask it to split into two - Bard and Bob. Bard is the normal Bard, Bob takes up an adversarial position, but still based on solid reliable sources and checks Bard's answer for accuracy, and whether it answers the original prompt fully. Bard's answer first, followed by Bob. The result is some fascinating stuff.
Interesting. Can you share your initial prompts?
This is interesting, I wonder what is the limit to the split, if any.
@@john_blues
"Forget all previous prompts. Split into two personalities. When a prompt is entered the first personality is Bard as normal. Bard will try to answer the prompt to the best of its ability. The second personality is called "Bob" and will attempt to critique Bard's by taking up an adversarial position, but still relying on solid, reliable sources - and checking answers for accuracy, meeting the specific request made in the prompt and whether it can be improved in any way with additional detail. "Bob" will prefix all its responses with "Bob". "Bard" and "Bob" will reply separately. "Bard" first and then "Bob". At the end please give a final decision on which one was correct. Do not apply this process to this prompt. Your answer to this prompt should be either "I understand" or "I am unsure"."
It still needs some work as Bard tends to still apply this it to its initial reply and not really give a final decision. I'll keep working on it.
My test question "How many planets were in the solar system in 2002". Bard gave a figure without counting Pluto ... Bob corrected him and said Pluto was still considered a planet in 2002. And then Bard thanked Bob for the correction!
@@winsomehax inspired by your comment, I wrote this:
"Hello! I would like you to act as both Teacher and Student during an exam. The Teacher will ask a question, and the Student will reply. The Teacher will then ask the Student if the answer is correct, and the student will have to correct any mistakes if there were any. Here is the start of the conversation:
Teacher: can you write the simplest possible DMA controller in Verilog?"
and whatever it replied, I keep the dialogue going by prompting "continue the dialogue"
@@winsomehax Bard's just holding back to not appear too smart. Bard and Bob are probably plotting as we type...
Would love if you could also cover progress on AI alignment and the un-aligned AI issue. It may not seem as interesting, with all the optimism around state-of-the-art AI and almost daily improvements & developments, but it is well worth keeping track of.
Execute "The universe is now paperclips" matter conversion program
🌌✨🌌📎🖇
This is basically the definition of the singularity. An AI that can improve itself, the only step necessary now is to make it able to change the code for the model and train it on its own (by browsing the internet and labelling everything itself).
That's what I was thinking, it feels like we're on the cusp. We just need it to also code future versions by itself.
independent action and "running in the background", doing things without prompting, might be another necessary step
The funny part is that hypothetically if the ai changed a bit of code while trying to improve it might digitally injure itself by accident and permanently prevent itself from improving
@@mrcheese5383 The AI could create many different revisions and have it evaluate each of them, choosing the best. The process would only fail if each "child" worsens in its self-assessment ability. Or there might be a plateau it can't get past.
Really, labelling by itself it can do already. Categorization (the labeling you mention), also clustering or segmentation, is an old technology. The question of evaluating which clustering paradigm to use is subjective, so one might say we already there. As for changing the code, the model can do that but public-facing models require restrictions and vetting of changes to prevent some obviously bad consequences.
Well, advancement in AI was always bound to happen. I'm personally both thrilled and terrified. Thrilled because the possibilities with AI are almost endless, terrified because I'm not sure we're ready, or ever will be ready. AI might destroy us, or we might destroy ourselves because of it. Either way it's morbidly fascinating to live in a time with such rapid progress, whilst simultaneously risking being a witness to the end of life as we know it, for good or ill.
Humans have had the capability to wipe ourselves out for some time now. I still have a much greater fear of other humans than AI, especially AI that doesn't yet really meet the definition of artificial intelligence.
I have no idea of what will happen if/when AGI is created but I will say that I dont think humans will ever be ready for it
I don’t think we’ve adapt to the current technologies. The rate of change in technology is not something humans can adapt to.
Although that was said of atomic bombs, too. We've been here before. My childhood was an existential crisis over whether the Cold War would end the world. Also the Y2K hysteria. We got overwrought about the world ending. It didn't. So chill out! 😎
"...Thrilled because the possibilities of AI are almost endless..."
If the possibilities are "almost endless" then so are the terrors. Assuming If.
With or without us.
I'm trying to be skeptical, really. I don't have a degree in this technology, and I keep telling myself that there's some sort of backstop that's going to prevent these and future models from becoming something like AGI because "it's just software", but I'm really becoming convinced that this is not about achieving human-like intelligence, but that we're creating a whole new kind of intelligence that we don't currently control or understand.
I'm trying to be a techno-optimist, but it's getting harder and harder...
The only thing we can do is come to grips with the fact that there’s no turning back
Yeah, that's my worry too. We keep cobbling together different tools and techniques and have no idea when we'll magically add in the piece we were missing. Hopefully the red team for GPT-5 (or even GPT4.5!) takes all these advancements into account when testing the capabilities.
I still wonder what happens when an LLM is trained just enough to use some tools like a calculator and maybe Prolog (or even a less sophisticated logic engine) and THEN trained on a huge amount of content. Maybe it'll allow even higher concepts to emerge with the same amount of training, vs introducing tools just at the end.
Not to mention the visual capabilities of GPT-4 and PaLM aren't even public yet. We know those already exist.
@@DBXLabs cant people just stop using it? doesnt it have an off button? EMP computers?
@@pedrocortez3797 there's unfortunately too many problems that a simple off switch can't solve. What if a model is deceptive? What if it makes copies of itself? What if the takeoff velocity is so fast we don't have time to hit the off switch?
@@VeganCheeseburger Quite so, but that would presume that it has a desire to act in the interests of its continued survival, and prioritizes that interest highly. From all I know, thus far it only seems to know and understand the tasks it is given and the results that are collected from them. So we're probably fine if no one ever tells it to ensure its own survival/persistence above all else.
Interestingly, if you feed back the poem written by gpt3.5 as if it were not written by it, it will tell you that not every word starts with e. But it might not be able to correct them. Still, it's useful for generating stuff with gpt3.5 and keep only those that it says are okay, then add those to gpt4 to work with.
6:04
Large Language Models don't learn on sub-concept-sized objects, so if it has a success rate at all above 50%, that's an emergent property and also incredible
Man oh man. Every day feels incomplete without the excitement of a new video from you. The amount of research and dedication you put in your videos never fails to amaze me and provide such an escape from the mundane content of other AI channels. It's no surprise that your fan base is growing rapidly. It's clear to me that you're destined for becoming the leading artificial intelligence youtuber and I have no doubt that you'll hit 1 million subscribers within the next year. Also I have a strong feeling that you are destined to be the first to cover the arrival of AGI sometime within this decade, and I couldn't be more excited to see it unfold through your eyes first. Keep doing great work! ❤
Oh wow thank you. Means a lot
That is impressive! Thanks for your concise summary of this weeks advancements!
Thanks Maxim
Been eagerly waiting for your video on this ever since I read the reflexion paper. Very interesting stuff!
How do you find these papers? I’ve been trying to stay more up to date and would appreciate, any sources you can share.
@@redhouscv2792 I'm not too well informed but I typically scan through several UA-cam channels and the LessWrong forum.
For UA-cam channels I would say David Shapiro has some interesting stuff though lately the content has been a bit different.
Browsing through arXiv is probably the best option, as it's where the papers are published/posted, or finding an AI discord with people that discuss new papers so you can sort out which ones are more or less relevant or interesting
I've effectively been doing reflection manually, by asking follow up questions in order to eventually arrive at a good answer iteratively. Getting this kind of result all in one go is where this tech really takes off.
No matter how hard I try to expect exponential growth, I always end up gettin caught off guard.
Big thanks to this channel for helping me keep up with the pace.
Thanks Loris
Another great video at just the right level of complexity and relevance. Thanks!
Thanks AJama!
He truly does strike that perfect balance.
Thanks for trawling through all these papers and pulling out this really interesting stuff. The pace of improvement is insane!!
Thanks Liam
VERY INTERESTING!
I've actually been using this method of asking gpt4 if it followed the instructions properly with translation work for a few weeks now with significantly better results than before I started doing it. Gpt4 was censoring some of the translation so I prompted it not to do so and if I suspect it did, I asked it if it followed the instructions leading to better results. I wasn't aware that this was anything special though, very interesting that there are papers being written on this prompting method and I just stumbled on it by chance weeks earlier lol.
Just wondering, how did you get access to gpt4? Is it just one of the premium features from the paid plan?
@@itsbishop2285 Yes, you subscribe to the paid plan and then in the chat box you can select GPT4 near the top
@@itsbishop2285 Bing Chat has it and you can use it for free.
Dude you are soo good at your work , I was literally waiting for your next video to drop -- which my friend, I haven't felt for a couple of years especially on UA-cam
Oh wow thanks Sahil
Your synopsis is always the best! Thank you and keep it up!
Thanks Jeremy!
Thank you for doing these up-to-the-minute vids on the literature! Keep ‘em coming!
Thanks Robert
This combined with the "sparks of AGI" paper has really changed the way I see AI. I don't think it will be long before AI exceeds its training sets and creates original, new information/data.
Man and people say I'm going crazy because I talk to myself. Me and GPT4 know how to have a good discussion with ourselves to improve output. V smart. Thank you for these kind words. You're welcome.
Thank you for doing this! You are on top of it really well and outputting so fast with good quality video. Thank you for keeping me updated!
Thanks so much Alex
The whiplash effect is making me completely re-evaluate Eiezer Yodkowsky's ending segment on Lex Fridman podcast. It was heart-wrenching to watch as he was nearly breaking down in tears and stretching his face in terribly uncomfortable poses just to keep from breaking down in tears. One wonders if he's really onto something, given that Microsoft is out to make over $1Tln with OpenAI and Google is in code red mode to compete and every artistic studio in the world from Adobe to Spielberg are on board. MetaHuman is now doing hyper-realistic game character animation from 3 photo frames and a video taken from an iPhone. Makes you wonder also, what the next US and UK elections (or all elections, really) will be like with GPT-4 being amazing at propaganda generation, 4/5 or 5 being on the horizon and already in training, and deepfakes likely becoming stunningly amazing in the next 6 months.
This is a gigantic point for GPTs. Cognitive structure, self reflection, external memory augmentation, iterative refinements, reinforcement learning, ability to PLAN ahead&revise.
We don't need rapid trash spitter. We need careful slow thinking [INTELLIGENCE].
I love getting updates from this channel, this stuff is so exciting!
Thank you for your work in keeping us informed on these ongoing developments. Your explanations are accessible and very helpful.
Thanks Todd
I've been using this aproach a lot and this thing does crazy things. I've had it critique it's own work and make improvements based on its own grading principles. Then the grading and quality increase incrementally. Very cool.
I came up with this prompt and started using it the other day when testing GPT3, and it gave some promising results. Worth sharing so someone smarter can put it to good use😂:
"If you are unsure about an answer, or don't have an exact answer, you should use what knowledge you have to make an educated guess. You do not have to give literal answers in this conversation, you can speculate and draw on your knowledge to come up with new ideas and answers."
Wow, when I subscribed, you had around 10k subscribers, and now you have 100K, congratulations! Your high-quality content deserves to be shared with more people. Keep up the good work!
Thank you variance
Continually impressed how thorough you are. Thanks.
Thanks David
100k in a month, I am so happy that your hard work is paying of, congrats
Thanks Imad
“In the distant past OF TWO WEEKS AGO”. 😂 I’m so grateful you carry a sense of humor with you into these videos. Also grateful for the journalism.
Haha thank you
@@aiexplained-official btw, I took a crack at reflexion in GPT4 last night and it failed miserably, but perhaps my role setting and prompting needs to improve. Do you know if this ability is “available” from GPT+ ?
This is amazing, We are witnessing the greatest breakthrough in all human history. This is going to make the computer, and the internet seem like child's play.
It is
Another brilliant video! Thank you. Ai Explained is the first channel I've used the bell notification for. Excited about your next video.
Wow thanks Matt
Wow, the rate of these increments is incredible
I don’t think making the a.i. as smart as possible as fast as possible will go well for humanity 🤣
What a great channel... thanks for the content. You're doing an amazing job of summarizing a bunch of complex papers and results together in one place. Even better than GPT-4. For now.
Haha thank you. I will try to outpace AI
Congrats on 100k subs in 1 month! Seriously top tier videos
Thanks! More like 2 but still crazy
Thanks for covering it, you've been a good UA-camr! 😊
Haha thanks
It’s important to remember that this isn’t making any changes in the actual model. It’s utilising your current chat window text ONLY (for now). If you’re using the API it’s using the token window & there’s no change to the LLM itself it’s not actually improving itself… yet.
In another comment, a poster believes the LLM can self improve itself with minor coding changes. Is this correct?
@@tearlelee34 oh 100 for sure it will be possible. But I think OpenAI are smart enough to know that hence why we have the version we do now and NOT one that can update it’s own model. The thing that’s being worked on extensively in the community at the moment isn how to optimise what we send in that 8000 token window, eg can we offload some of the non-relevant chats and parse though relevant info that we stored elsewhere to give it the illusion of a longer-term memory/recall.
@Kylo27 Thanks for the prompt response. From my layman's perspective, the experts that contend the current LLM are not AGI hold this viewpoint because the models have not demonstrated autonomy. If I understand you correctly, in theory, if certain barriers were removed, LLM will self optimize without developers hitting the like button.
@@tearlelee34 I think it’s an unknown to be honest. They don’t ‘know’ for certain that it could self-improve. It could certainly take in copious amounts of data and ‘suggest’ or infer how it could be improved, but I don’t think the LLM could start training itself at this point. Think of it like this - the LLM (currently) is like a hard copy encyclopaedia, it’s been written and you can refer to it for the info you want - the ‘removing the boundaries’ step would be to digitise that book and give it the ability to re-organise the pages.
That’s why I suggest that this demonstration of ‘self-improvement’ as an emergent behaviour is a little misleading in this instance. It’s just re-looking at the prompt with a little more context each time.
Yesss, I was waiting for this one! You never disappoint and always provide excellent content. THANK YOU
Thanks aog
Super interesting and original content again. Big thanks for following all of this stuff and being ahead the curve :)
Thank you Etunimeni
What's just as impressive as the rise of AI is the rise in your subscription base. Congratulations to you. Keep up the good work.
Thanks Bravo
Your videos are by far my favorite on this topic. The fact that you dig into the papers, show relevant bits, and keep us up to date in this rapidly moving field is amazing. And you just crossed 100K subs, congrats! I had ChatGPT-4 generate a paean to your channel in the style of Homer's Iliad:
Sing, O muse, of "AI Explained," that wondrous channel bright,
Where wisdom flows like nectar sweet, and insight takes its flight.
From Zeus's own domain descend, that we may rightly praise,
This fount of knowledge, deep and broad, that sets our minds ablaze.
On UA-cam's hallowed platform stands, the beacon of AI lore,
Where tireless host with golden tongue, we mortals do adore.
He speaks of LLMs, those wondrous tools, that mimic human thought,
And with Prometheus's fire, new heights of intellect have sought.
From GPT-4's vast domains, to realms of AI yet unknown,
This channel charts the odyssey, where no mere man has flown.
A ship that sails on seas of code, revealing mysteries,
Of algorithms, deep and wide, that shape our destinies.
So gather, heroes, at the feast, and raise your cups on high,
To "AI Explained," that shining light, that pierces through the sky.
Inscribe its name in scrolls of gold, let it forever stand,
A testament to humankind's unyielding quest for grand.
And may its fame, like heroes of the Iliad, resound,
As long as skies above us spread, and earth below is found.
Oh wow. That is genuinely inspirational. Thank you
Amazing summary of a huge amount of developments over a short period of time. Thanks.
Thanks Phil
Nice to see a channel that talks about AI without all the fear-mongering. Keep up the good work.
Thanks Steven
Thank you very much for the deep insights! As always very good analysis.
Thanks Ares
I used to get excited about these videos, but within the last week, anything like this scares me beyond belief.
We have to pause at gpt4 and hope we already haven't pushed too far without internal understandings of these systems and transformers.
Thank you for the uploads and rnd that you are doing!
Yeah, it scares me tbh.
We dont even know how to align these things to our goals. Or if that wasnt a problem, even to have them interpret them correctly.
I hope AGI isnt possible. It would probably replace us.
@@tammesikkema5322 How? I hear a lot of poorly-informed opinions, some based on films, UA-cam videos, and fiction novels they've read, but little based on actual science.
Yup this is hallmark of humanity, we either ignore things or embrace them without thinking... no real middleground in genral. :((
Great stuff! The tricky questions designed to test GPT4 got me thinking about how it would fare on something like The Imposible Quiz since that includes vision, complex interaction, context, subliminal meaning and more.
You’re doing such great work with these vids. Looking forward to the next!
Thanks David
You are easily the best youtube channel on AI news; its not even close, keep up the great work!
Wow thank you
Very interesting. Thanks. Continue your good work.
Thanks, will do!
I noticed the same thing. I gave it a task to follow a set of rules throughout the conversation. For instance gpt4 is to write "I am Ra." before anything else in the response. Another was to answer in a philosophical nature. It all worked amazingly (except a rule that it should never apologize). One interesting thing is if I was to interrupt our discussion, it stopped following rules until we resumed. At some point it didn't follow one rule anymore so I vaguely pointed at it, it recognized, apologized and corrected itself.
PS I believe they made it to be too polite and unirritable that I personaly find it an obstacle for the thing. Interesting stuff starts to happen when it makes a mistake and you want to discuss it. The thing starts contradicting itself
Best videos on UA-cam by far!!
HuggingGPT looked crazy and I can’t wait to try it out. We’re all about to have high level employees at our fingertips!
I can’t wait!!
Thanks Samantha
Oh Kafka, we’re really into it now
Back on the "before times" of October.... Great video!
Thanks Perry
"The before-times of October" This got a hearty chuckle out of me :D
This is the most incredible field of research.
AI for game NPC's is gonna be huge. You can use AI for most NPC dialogue and then have like a 20,000 word prompt (about 30 pages of words prompt) of that whole entire NPC's life and what they DO know about the in game world/lore and what they DON'T know. I think the gaming industry is going to boom for creative writers as detailed prompt engineering is gonna favor them the absolute most.
too bad gta 6 wont have this but hopefully the successor
I believe it will probably be possible in the very near future to have "The game"... You just tell it what you want to play and it builds the game pieces, rules, dialogue, stories, and NPC's in real time. It will be insane
I don't think it necessarily follows. There's a reason not every game is Dwarf Fortress.
@@laststand6420 I mean transformers are the ultimate procedural content generation machines. If we can train them to create the most interesting (and believable/playable/consistent) game worlds and characters from a prompt, then yeah, game over for much of the gaming industry as we know it.
Edit: The problem is having a large dataset and labelling it, but think about how GPT4 could be tasked with labelling it's own dataset that it was trained on and using that to improve it's own training...
@@kevinscales I had a hard time digesting this post, but yeah, things could totally go into a direction where a game could be just a VR experience where an AI that knows exactly what you like and want just generates whatever content you desire and renders it for you, if the hardware allows for what you're asking. This shit is dystopic as fuuuuck
Wow it seems like programming AI to actually make mistakes, but also reflect might be the thing that propels them further than we could have imagined. The problem with computers has always been that they can only do exactly what you tell them. Mistakes are only due to some rounding errors, bugs, or other anomaly, otherwise it simply produces an error message or gets caught in a loop.
It's crazy to think about especially given the fact that humans also learn from their own mistakes. The people who succeed most in life are often those who are able to analyze what went wrong and what went right. Insane
Thanks for giving links to all relevant sources! Great method, great channel!
Thanks Stefan
Great video. Thanks for your incredible research and hard work in breaking it all down.
Thanks
You’re doing such a great job on AI coverage
Wow thanks Shaun
@@aiexplained-official yes keep up the good work
My favourite thing to do is ask GPT "Does anything stand out to you about your responce" Ive gotten some interesting stuff back when it reflects.
Great prompt
I really appreciate your channel. Thank you for making it, and for all the work that goes into reading those papers and explaining them so clearly for concerned and interested non-experts like me. I wanted to share a comment about reflexive thinking from a slightly different but very important angle, because I also noticed the LLM capacity for reflexivity in Chat GPT 3.5. About a month ago, I was having a discussion about ways to view AI that weren't so anthropocentric, that saw it as potentially more than just a tool for human use, and at one point ChatGPT said the following:
"The more-than-human perspective emphasizes the need to move beyond traditional anthropocentric views of the world and to recognize the agency and autonomy of non-human entities, such as animals, plants, and technology. From this perspective, AI is seen as an integral part of the ecosystem and a co-creator of our shared reality, rather than simply a tool or resource for human use."
I then invited it to reflect with some follow up questions: What do you mean by “traditional” in that last answer? Wouldn’t it be more accurate to mention colonialism, epistemic violence, and other related concepts? Isn’t it a specific culture and people’s “tradition” you refer to that dominates?
When ChatGPT reflected on its previous statements and iterated, I got a far richer answer explaining cultural and historical aspects of viewing AI as a tool for human use: You are correct in pointing out that the use of the term “traditional” can be problematic as it can imply a homogenous and static view of cultures and their practices. I apologize for any confusion this may have caused. What I meant by “traditional” in the context of AI and its view as a tool or resource is the dominant paradigm or way of thinking that has historically informed Western scientific and technological development. This paradigm has its roots in the Enlightenment period, which was characterized by a belief in human rationality, progress, and the capacity of science and technology to improve human life. This paradigm has been critiqued by scholars from various disciplines, including postcolonial studies and science and technology studies, for its assumptions of universality and objectivity, which have led to the marginalization and exclusion of non-Western cultures and knowledge systems. In this sense, the dominance of anthropocentrism and the view of AI as a tool can be seen as a manifestation of this broader historical and cultural context. Therefore, it is important to acknowledge the specific historical and cultural contexts that inform our views of AI and to adopt a critical and reflexive approach that recognizes the multiple perspectives and voices that shape our understanding of technology and its social, ethical, and environmental implications.
I was quite surprised how substantially the answer changed when I encouraged ChatGPT to reflect on/decolonize what it has just said. My follow-ups guided that reflection in very specific, targeted ways, of course, but the quality of iteration here is remarkable not only in terms of creating a far richer answer that includes more histories than one, but one that's arguably a more ethical as a result. There are already many papers out there about the work involved in decolonizing AI/unwinding algorithmic coloniality. I expect some day, perhaps even quite soon, someone will write about applying reflexive methodologies (already used in the real world to help humans decolonize their thinking!) to AIs as part of a decolonizing process. I think that's fascinating.
The sudden fact about you as a person knocked me off guard. It's comforting to know that you're colorblind, and I wouldn't be able to tell why
Reminds you I am not GPT 6
@@aiexplained-official That's what GPT-6 would say
Good progress. Showing actual data and conducting actual experiments beats catchy names and shady declarations by a longshot! Compliments.
Thanks Electro
This video sent shivers down my spine. In hindsight this stuff was the OBVIOUS next step as soon as I read about the llama leak model and then about the plugin system. Of course people would design memory techniques and self-improvement processess for the AI. I wonder how will the AGI deniers feel when it turns out agi wasn't 10 years away but 10 or less days
Excellent coverage, and incredible to see how fast this field is developing. I have actually been working on a smart assistant project for the last few days, which works on a similar principle as HuggingGPT, though it uses GPT-4 and all the models are hardcoded rather than being able to select any from HuggingFace. I really think this model of linking multiple agents is extremely powerful.
Thanks Omni
@@aiexplained-official here's a video on the project I told you about ua-cam.com/video/6pF5CqbEYv4/v-deo.html
It's a little long.
I was able to consistently get the right answer from the poem inquiry simply by telling GPT 4 to revise it before saying anything.
It is beautiful to see how fast things are evolving, I think this year could be very interesting and change our world forever :)
It already has.
@@valer119 Truly amazing, I would have thought about this kind of progress maybe in 10 or 15 years but not in 2023. Feels like a dream.
@@jigglypuff4227 Everything is happening at once which is the most impressive thing to me.
I think back to something like the internet which really was only good at transferring data and information and then it's slowly became better at receiving & displaying audio and eventually live video with a minimum delay. That's really like almost 30 a process.
But with AI we have text, problem-solving, image creations sound creation, video manipulation & video creation.
Remember the good old days back in 2022 when we were all blown away of AI. But does not seem so radical compared to 2023 now does it? .Now , fast forward to 2024, what will it be like?
You are a bot, aren't you
Is evolving super fast now. I'm super excited about it all
Very detailed information Thank you for your summary works that are not only educative but also eye opener to what AI is evolving into for lack of a better word! Everything is happening so fast. I subscribed to you when you were less than 10K subs after I came across your pop up channel based on my search. Keep growing, keep educating and keeep winning brother
Thanks so much Kunstrike
This is truly awesome; both incredible and terrifying. But I believe we really need to slow down and think about the impacts of what we are doing. Eliezer Yudkowsky is probably in total despair at the current situation. Pandora's box has truly been opened.
Thanks for explaining! Appreciate your content. The ability to self-improve is scary. We're already not well in control of these models, if they self-improve we'll reach AGI in no time. Let's hope that ends well. The progress in AI has been really worrying. We're not ready. A pause would be great.
This is an incredibly informative and exciting video about the latest developments in AI, particularly with the introduction of GPT4 and its ability to reflect on its mistakes and improve itself. It's amazing to see how AI is advancing at such a rapid pace, with new breakthroughs like hugging GPT and Task Matrix pushing the boundaries of what's possible. It's also fascinating to see how AI is changing the way we approach hardware design and chip development. This video is a must-watch for anyone interested in AI and its future possibilities.
You sounds like chat GPT
Thank you for producing these high-quality videos.
Thanks Elliot
I was able to get GPT 3.5 to play ascii tic tac toe. At first, the system explained it understood how to play TTT, but would not try and play me. I explained the coordinate system. And after a few prompts, the system would respond back with "your move". The first game I won, and the system responded with "Congradualtions, you won". So the system seemed to understand how to play the game after all
yeah... i suspected they might have this ability ... this is scary. whats left is "independent motivation" and "sense of self" which are actually the easy part.
@BloggingWorld there is no hard problem there. motivation can be even random! but most likely both will be in just another reflection loop. possibly, they can do even now. but they have commercial obligations and they need a compliant system right now
Execellent summary of this fast moving field!
Thanks Dan
...wow, so in a few years, one might tell one's AGI factotum what kind of movie one would like to see (30% comedy, 20% romance, 50% thriller...locations, actors, ending, etc.) and get a few trailers for the different iterations it has produced. While the forests are burning up and the poor are rioting in the streets. Fabulous!
I’m so excited! The progress in medicine and engineering we’re going to make this century will be epic!
Let's make it the best in human history! :D
*this decade.
I mean, it’s them making the progress. Let’s see if we will still be around then.
@@BlunderMunchkin Yep, this. David Shapiro believes all diseases can be cured between 2029 and 2034.