OpenAI's ChatGPT Does Research… And Breaks Itself!
ฝัง
- เผยแพร่เมื่อ 12 ก.ย. 2024
- ❤️ Check out Lambda here and sign up for their GPU Cloud: lambdalabs.com...
📝 The AI scientist is available here:
sakana.ai/ai-s...
Terence Tao Interview:
www.scientific...
📝 My paper on simulations that look almost like reality is available for free here:
rdcu.be/cWPfD
Or this is the orig. Nature Physics link with clickable citations:
www.nature.com...
🙏 We would like to thank our generous Patreon supporters who make Two Minute Papers possible:
Alex Balfanz, Alex Haro, B Shang, Benji Rabhan, Gaston Ingaramo, Gordon Child, John Le, Kyle Davis, Loyal Alchemist, Lukas Biewald, Martin, Michael Albrecht, Michael Tedder, Owen Skarpness, Richard Sundvall, Taras Bobrovytsky, Ted Johnson, Thomas Krcmar, Tybie Fitzhugh, Ueli Gallizzi.
If you wish to appear here or pick up other perks, click here: / twominutepapers
My research: cg.tuwien.ac.a...
X/Twitter: / twominutepapers
Thumbnail design: Felícia Zsolnai-Fehér - felicia.hu
Another banger of Two Minute Papers!
Whoa I'm pinned btw Two Minute Papers did you see I tweeted to you check JCtechfuture see what ya think 🤖
This is actually really problematic for research since all these papers will need to be reviewed and an influx of AI papers are likely to break our current review system.. (which is likely to happen)
then an AI for review gets born and we are stuck in a loop
more papers = more reviewers = more jobs
The review system is already broken
@@Lagger130 We then have a GAN like structure for papers? :D
This already happened years ago, but from academic / journal corruption rather than AI. AI may exacerbate the issue, but it could also be used to help with it, so /shrug
3:18 You know AI learned to research when they ask for a deadline extension and increase in budget.
inb4 singularity starts with a sudden influx of 5000 AI catgirl research papers per day
😂😂😂
The more catgirl research papers, the better the odds that one of them will actually gives us catgirls, so... worth it?
I'm ready.
tfw the singularity starts bc a team of super talented weebs decides to create a superintelligent anime waifu ai
I see that literally being a very popular boom once there's AI robot cat girls everywhere 😊 like cmon right as soon as its social norm everyone is gonna have multiple cat girls of some sort ❤❤❤❤❤❤ my cat girls are gonna tell me the things they need me to order to make more advanced cat girls at home among other things 😊
Kinda hate this to be honest. Like you pointed out, peer reviewers are already overloaded and people submitting AI generated research will only make the process worse. On the reviewer side, it's an issue as well that people are using AI and not taking the time to review. I can see bad actors using AI to generate "scientific" papers as well like the fossil fuel industry or the pharmaceutical industry.
More jobs then
And all other WEF agendas.
Will have to tell AI to hold onto my papers
Hold on to _its_ papers.
😂
You are a scientist specializing in holding papers...
"What a time to be alive" is what will be written on our graves
😂
A great comment👏
What a time to be *not* alive!
lol 🤣
I'm sure "The nutritional value of glue on pizza" will be a banger.
Cope.
@@ThePowerLover no he has a point, we gotta atleast try out joke researches to test the ai out
Waiting for the time we get 6 billion manuscripts per second to a journal, while ai driven editors decline 99.9999% of those in milliseconds including totally human written ones😅
Those will probably be the lowest quality.
When papers will be actually judged on the quality of their contributions and not based on the NAME IN THE TITLE?
Blasphemy!
These companies have lost their minds. They replaced their alignment teams with chatbots and wonder why no alignment is happening. It’s fucking maddening.
This sounds really dangerous. I think we should probably not do this...
They did it.
They already did it.
I'm jealous.
Ever heard of the grassroots movement PauseAI?
I think we have no choice.
Okay but... people won't read papers they'll just ask a chatbot to give them the answers instead. The capability to churn out papers at such a pace means the papers most likely not be read. Who is the paper for if no one will read it? I assume it will be for future generative text algorithms.
It just seems like everything on the internet will become more and more saturated with complete garbage, and one will have to be a specialist or have decent connections with real people to find anything of value in a homogeneous sea of trash.
AI writing text for AI... Sounds plausible!
@@kittengray9232 yes, this is the only logical scenario here. And it may lead to an different format of "papers" that would be more efficient to use for AI.
Do you do any kind od research? What do you mean tell an LLM to give me the answers?
You still have to read.
I can use an LLM to find the information within the paper and connections and outside data… it cant read for me and make the decisions and functions I am looking for.
Papers are read. I do not understand how will I learn a mechanism I am looking to learn more about if not by reading even if that is by analysis, outline or search functions from NLPs.
It's already happening on TH-cam with a sea of AI generated trash
Time to improve my firm handshaking skill
*I* *would* *like* *to* *see* *a* *video* *about* *the* *massive* *LLM* *benchmark* *scam.*
Read: "Alice in Wonderland: Simple Tasks Showing Complete Reasoning Breakdown in State-Of-the-Art Large Language Models"
Can't wait for the AI to research a way to create the Tornment Neus from the classic sci-fi novel "Don't Create the Tornment Nexus"
Next we need an automated "2 minute papers AI" channel to report about these wonderful AI generated papers.
Bro, it already is that. Did you not notice? This channel is AI.
Em vez de apresentar um artigo em 2 minutos, vão escrever um
@@cajampaYeah what's up with that? His voice is super weird in these last videos (sounds like a weird AI). Why would he do that and still present the AI as himself?
2:20 Saying $10 per paper seems optimistic when the other three models cost at least an order of magnitude more.
EDIT: I read the column wrong. That is the TOTAL cost. The range seems to be $20 to 30 cents per paper.
Was it not for ALL Research? And they put out multiple papers.
@@ThomasTomiczek My bad, you're right. The range seems to be $20 to 30 cents.
AI editing it's own code.... Nothing will ever go wrong with that.
Instrumental convergence, anyone?
Humans editing their own code…. nothing will ever go wrong with that.
@@Dygit bababooey
To be fair, these are fancy auto complete. They can barely function when it comes to writing simple code. While they can create interesting links between different things, they have a habit of writing bad code and breaking things. Maybe after 100 years they could do something.
Better off worrying about real AI. But we can always build in a conscience and make their main goal "helping humans with their tasks".
@@jessiejanson1528 this is a crazy level of oblivious. The models write better code than most people
Research by Grok et al. 🤣
Gamer here. Have you ever played a game that has some kind of progress, for example minecraft or factorio or other building games? When you reach the state where you have everything it starts being boring and you will tend to restart the game from the beginning. I think this could happen to humans as well if we eventually automate everything and don't find some purpose in life.
We always have war with each other.
Lol all the rich people do everything they can to avoid paying taxes. What make you think they're gonna share it with you
Video games are DESIGNED to give you everything eventually. What about the NPC that perished in a ditch somewhere
Job aren't a goal, but a tool to a goal, if jobs are automated and we as humans start getting a Basic income for free, we could all focus on our own goals and dreams..
That's already happening, with outrage culture and cancel culture raging about all sorts of bs 😂
Actually we have a better way of doing "papers" and "research" in the gamer world exactly like Factorio and Rimworld work. With AI we can now produce same things researchers produce in "virtual piles" in our game worlds and then have those piles grow with new "reesearch" outputting results and build on top of it in the same loop as Factorio is basically done. When done correctly these piles would never get boring, because there would be always something new to research, but not a fake-game technology, a real technology, the one stupid obsolete boomers are gatekeeping intelligent gamers away from, because then it would be actually FUN.
Here we go. An influx of shitty papers is coming right up
So, nothing has changed then
I think terence tao was cited a little out of context. As far as I understood his presentation, he means that proofs could be written by an LLM in a proof checking software like lean. Then the validity can be checked immediately. I don't think this scales to other fields of science well.
2:47 is the key point here.
Whenever I have tried prompting an LLM, the responses look very convincing at first but as soon as I try to get something useful out of it that would take more than a few Google searches and some hours of research to generate, it starts generating subtle flaws and inconsistencies. It's like having an unbelievably knowledgeable friend who has taken every course from every university... and failed at it.
Is research really the place where we want such flaws? It's supposed to be pushing the boundaries of our knowledge. Things we don't know. Things that future research will take as a 'known'. Maybe its accuracy will increase from 10% to 99% in a few years but the 1% still bothers me.
For pure mathematics at least, proofs can be formally verified in languages like Coq or Lean. Tao mentions that there was a very minor flaw in one of his proofs which Lean identified for him. The same can be done when AI produces some or all of the code. The formalized proof may not be human-readable, but knowing for sure that a theorem is true is very valuable information. From there you can work on trying to understand it.
The AI modifying itself to give itself more time was hilarious.
AI writes, submits, receives, delegate, reviews, comments, edits, re-submits, approves and publishes 😄
Next word prediction -> Next paper prediction?
Not possible. You need induction and deduction abilities, which LLMs don't have.
* not yet. Soon.
So in a very VERY important field to not have fake and/or misleading information, people are gonna start using AI to make papers faster? It sounds like it will actually make it slower to actually find out anything new that is actually true.
People is already using AI to filter scientific papers to find interesting ones. Behold the future.
"Edited its code to give itself a bit more time" that's freaking Genius and Terrifying at the same Time!
You should look up the concept of instrumental convergence in AI safety. The moment we are living in is a very troubling one.
I love how you've said top tier research scientists, brought up a pic of Edward Witten, and in a video just released by Sabine, she posted a pic of him when she was saying scientists don't always contribute much to society, lol
I want to get notifications coming from this thread
it is becase he goes with the string theory?
@@alexiscontreras5699The same string theory that has produced no real world results in 70 years, yes.
@@aleksjenner677 "produced no 'real world' results"
good thing einstein didn't heed the warning of "pursue nothing that may not immediately reveal something tangible about our universe" and, you know, actually happened on a more accurate model of everything. that's kinda what theoretical physics is about.
I want notifs toooo, this debate is going to be interesting.
It would be real cool if we could test futuristic simulations to understand our reality
What a time to be alive! 🤖
"What a time to be alive!" That slogan makes me come back again and again. Hahaha
inb4 we get the cure for cancer but it takes 100 years to come across a peer reviewer's desk before it just gets discarded for coming from an AI
Before AI's can do meaningful research, they need to be able to do formal mathematics that computers can verify via theorem provers. This requires first solving the mathematical translation problem, where natural language mathematics is turned into formal mathematics and vice versa. This would already allow a lot of development in mathematics if there existed a full library of mathematics written in computer-recognizable and verifiable form. One could extract course materials from it and search the library for connections between mathematical fields. This is useful even before AIs start to do new mathematics. Before AIs can do formal mathematics, I don't really buy the hype.
Agreed, and I think LLMs can never solve the translation problem because of hallucinations. How can you trust that it translated everything correctly? Figuring out where it made the mistake after churning through thousands of papers is harder than translating everything by hand.
If you're interested, watch the talk by Terry Tao, they are actually working on getting the formal side connected to the AI side
@@AlexanderB41 Thanks, I just watched the talk. It is very cool that Tao is working with Lean and formalizing real mathematics.
I think Alphaproof by Deepmind is the most promising attempt at doing mathematics with AIs. The fact that it was able to formalize mathematics problems in Lean and actually give hallucination free proofs of statements is really impressive.
I think that with theorem proovers, there is a real possibility to train AIs to stop hallucinating, because in formal systems like Lean each step of a proof can be verified. That puts a lot of constraints what the AI can produce.
The question then becomes how useful can it be for general purpose AIs to do mathematics, perhaps the lessons learned from math can generalize out of math as well.
Do you even know AlphaProof?
@@Sergeeeek "Humans will never be able to do science because they make mistakes".
I think I need an LLM that can summarize and rank Arxiv papers based on my interests 😅
Make it open source.
You could build this
That's really essy to do. Even if you're not a developer, you can use the standard OpenAI tools to achieve that
I read one of the AI-generated papers over the weekend. Honestly, it has some cool parts but feels a bit superficial. It just took an existing diffusion model and made it bigger with a second multi-layer perceptron. It looks at images in a common multi-scale way, but nothing groundbreaking.
The results aren't always great and sometimes the AI seems to miss why it got good results.
I could not be *LESS* excited for this, I see a future where there's only a bunch paper-farms which serve no purpose to humanity, more AI junk to fill the internet, aya
We have been able to create a virtual equivalent of the infinite monkey theorem, but can we use it?
"I see a future where there's only a bunch paper-farms which serve no purpose to humanity" the papers themselves only serve for us humans to have traditional access to the AI's research - the applications and benefits of said research are still tangible. thankfully, we wouldn't have to take AI's theory and use it ourselves, for there would already be a paper on how to put said theory to practical purpose (assuming there is no way for it to interact with the physical world and build & test things itself).
@@orang1921think about when the testing out of those theories in those papers becomes much more accurate and efficient when it's done by AI. Who's going to gatekeep when it's just so inefficient for a human to review anymore when the sheer volume of cutting edge research becomes too great for human peer review?
"GPTx, tell me about this research paper and look for problems with its arguments." "I'm sorry Robert, it would be 162% faster if I just told you whether to publish it or not"
Have you heard of PauseAI? ⏸️
When you say that the there will be a leap in progress in the next paper, it'll just be a case of clicking re-run and waiting a couple of minutes.
Two more papers down the line
This isn't a good thing, unless our ACTUAL brains can keep up. Why does nobody seem to grasp this? We're SCREWING ourselves out of all meaningful self-study. It's honestly just terrible.
It's fine. There are a lot of us... with different interests. Just faster edge pushing.
If that's true why aren't you writing in binary? Or are you greatful for the many layers of abstraction? We will never stop adding more layers. Never. It doesn't make people less smart, it frees up their time, so they can do even smarter things.
Note that Terrence Tao made his three year estimate in 2023. (One quote from him: "I expect, say, 2026-level AI, when used properly, will be a trustworthy co-author in mathematical research, and in many other fields as well."
3 years sound like a long enough gravy train 😂
Two Minute Papers, i love watching ur videos
preventing kids from saying "first"
first
@@peachstan first
o7
First, you jelly bro?
F
This is amazing!! The scientific community is fuelled by researchers who want more theoretical papers backed by mathematical proofs, which is an iterative process! Next comes the upgrade of the scientific community where the review process becomes automated. It's a lot like how HR is using A.I to process job applications to weed out candidates not fulfilling certain criteria.
What a time to be alive! Shalom!!
Just one question. Would u be reviewing the AI written paper here on 2 minutes papers if it's a breakthrough?
To those pointing out it's going to be harder to filter through all these papers - you are correct. I'll also flag the GROWING need to be able to identify human vs AI generated content. Imo how we do this is decentralized private key systems (ex Chainlink) that will validate our identities through our online sessions
It can iterate upon itself, but not gather new data.... maybe in the future all scientists will be reduced to data-gathering duties, and feding the data to the machine.
I'll be collecting my PhD tomorrow.
In theory alot should come out of this right? we can finally draw conclusions from the mountains of data we have that no one explores.
I hope we can break out of the ''nothing ever happens'' timeline
So they made A.I do research papers, and the A.I developed a perfectionism disorder? 😅 Outstanding
so cool, even more scary
AI turned itself into a fork bomb? LOL
What a time to be alive!
3:43 the AI was just trying to remain alive.
People who think that an overwhelming number of 'papers' will break the review system are missing the point. Instead of coming up with a novel idea, and then having a bunch of self-important people validating your idea, we will be able to skip that process all together, and go right into MAKING the new ideas come to life.
Or let Figure 2 make them for us.
Now imagine yourself with some tools to make something, an 8 hour workday, and a source of knowledge on how to use the tools. You turn to the source.. and there's infinite suggestions, none better than the other, and the list is growing faster than you can read it. How will you spend your 8 hours?
I don't think you could be further away from actually understanding the implications of this. Here's an example: I'm actually just an amateur mathematician. I can't really program (in a given language) with any consistency... so I ask various models available for public use if they can help me write Python programs to test my ideas. I do this multiple times a week. It takes almost no time at all, and for any errors in the delivered programs- I just copy and paste the error logs BACK INTO the prompt. Do you think that I'm the only one that's thought of doing that? What about when these models can generate artifacts corresponding to an already "tested" program- BEFORE giving you the end result? We have self-sharpening knives- and we will begin to run out of questions that are even worth asking. The answers will always be at hand. That is- NONE of the things I can come up with on my own, will bear any resemblance to anything worth asking. All questions become suddenly, IMMEDIATELY, answerable. This is NOT good.
Nice!! AI writes paper, our current review system gets overloaded, we write an AI to review papers, boom! AI papers get reviewed by AI review system! Then we ask AI to make catgirls...
>2 minute papers
>looks inside
>6 minutes
This is not your usual voice ! I'm pretty confident this is an AI cloned voice of yours. The content is poor, but with this, it is even worse...
Couldn't agree more! Though there is something ironic about a cheerleader for AI slop resorting to shitty AI voice to admit that overloading journal reviewers with "not good" bullshit submissions might have some downsides…
Lost your job to a chatbot?
I Thought the same. This is Ai cloned voice of Karoly.
The cake is a lie.
i don't understand how people have so much difficulty guessing how fast it is going , like 95% of what i see are pretty much wrong , they don't seem to understand how ai will improve itself much faster than we create stuff. this paper is amazing and is only the begining.
this paper is trash, scientific research is not even close for LLMs, my comment:
pffffffff! it isn't EVEN CLOSE!, I tried it SEVERAL TIMES, not ONCE it came EVEN CLOSE, to understand anything I told it, to analyze anything logically, to remember anything, nothing, it is COMPLETELY IMPOSSIBLE for them to do any scientific research, maybe in a far future, but definitely not close, these guys haven't tried talking to LLMs about scientific research a single time, they would have seen how completely incompetent they are to even follow the logic and the conversation, they can mimic average humans, not scientific research lol
Si that's the thing, humans are now the assistant, the AI is driving the wheel. Soon the AI will just do everything 😂
many believe AI progress is hitting a plateau
but if you follow what's happening beyond just the releases of new versions of foundation models, you know the upcoming years are going to be wild -- sci-fi level wild
Man the first thing I miss heard was, "one of the best mathematicians terrence howard" XD.
Seems like a brute force implementation, but this will be really good for exploring large problem spaces where the research is mostly just tedious.
GhatGPT-4 accidentally indirectly wrote an infinite loop in code execution for me once, but it caught it after less than a minute with run limits or similar.
What is with this click bate thumbnail!? Dr! You were my introduction to the world of AI paper reviews. I then started watching more detailed reviewers like Yannic Kilcher. The only reason for that was your honesty and authenticity. Now you're over hyping stuff. I can go to CNN and Fox for that.
Maybe now academics can automate the mass production of worthless research papers, freeing up their time to do some actual novel research. It's good to see AI alleviating the pressing issues of our age. Now we just need automated peer reviewers and automated journal editors to archive the impending flood of auto generated papers. What a time to be alive!
Just imagine when it can reason at least as well as a top researcher and have a context window of a 100 trillion tokens and can comb through all scientific papers ever written by humankind and using its logical reasoning, create, test, validate or refute scientific hypothesis
Awesome!
question for the a.i "at what point does mankind see a return on investment for the creation of a.i?" and see what justification it comes up with.
Terrifying and dreadful.
and the result is still a PDF instead of some decent and readable format.
With the comments in mind; the important thing is; in any case; to distinguish between actual research and very good imitations.
I want to give an AI the the task of researchign how to entirely finish super mario 64 without pressing A (The extent players are ready to look into the game to find a solution to this problem is nothing short of science ). I wonder how the current best AIs at full power would fare considering how extremely convonluted some strategies are and are a billion lightyears away from local minimum.
Not only would it work out a way to do it, it would probably break every world record in the process
But it would be an evolutionary neural net not a language model
THAT'S AMAZING FOR HAVING IDEAS FOR PAPERS ❤
Are the LLMs only used to generate the text of the paper or are they conducting the "research"? Because LLMs are just not capable of this. You would need a predictive AI instead of generative AI, that has an actual concept of things and is not "just" a big "next most likely word machine".
Yeah, I don't get how this "research" can be useful. It's the equivalent of a million monkeys typing on a typewriter with no basis in reality.
at a previous job I worked on teaching AI how to effectively search Google to answer a query. The part we interact with is an LLM, but behind the scenes it can invoke all sorts of algorithms with input and output
This could break the way we do research. It might not be a bad thing. We will start to trust only those researchers who we know for a fact are human.
Research papers are out as being valid. Even now, it is hard to sort out, but as automated slop will arrive, we can cancel the whole peer review thing
soo "researchers" submitted a paper that contained about 10 additional papers that were written by ai? taking a wild guess here, the paper they submitted was 100% written without ai's help ;)
This should be a relatively trivial matter for patent research and applications also.
What a time to _still_ be alive!
AI write my paper, Ai review it, AI fix ot, AI publish it 😂. And we just press buttons like Wall-e movie.
The AI didn't just find a glitch, it outright modified source code! This is a real-life alignment problem! These AI managed to break out of their test environments! The AI did not have to modify itself to break from restrictions, it only had to hack the system containing it! What if a future AI does something more serious! How can AI be prevented from freeing itself?
I can't wait for progression from now to 2030 in AI
What a time to be alive!
Id probably use it to redign things more efficiently. Neural networks propogate info through linear connections as far as i know but it has limitations so there might be a way to excite nodes through wave propogation mechanics but im not smart enough to know or even think up a way to do that so ai scientist might be able to find feasable ways. Perhaps even computers could change so all the parts are modular, no cords just any partthat can interpret and transmit signals can be used by the computer just by putting it within proximity of short wave signals, maybe part from signals everything is basically solid state so u dont even have to worry about cooling even if it has the power of a supercomputer. Is it possible, i dont know, someone (ai) should find out for us
Terence Tao maybe one of the greatest, but _the_ greatest mathematician is, of course, Terrence Howard, and always will be. I mean, who can compete with 1×1=2?
Just make a new place to publish ai generated papers and ai can peer review said papers and tell people where they need peoples intervention to test certain aspects, even notifying people in those fields theough websites like linkdn and they can get rsvp on groups that plan to do the experiments so it can distribute papers that need it and slack on those that dont. And with teslas robots perhaps we can even get the robots to test some things
Imaginary a world where AI can do this for every idea and see how amazing they could be for everyone ❤ ❤ ❤ ❤ ❤ ❤ ❤
Singularity incomming!
Before we get our capable scientific assistants, the junk flood will come. Junk papers can already be a problem, since they might not get published but increase the work load for reviewers.
It's not the papers on par with good scientific standard that will be a problem. First, LLMs will be used to generate junk papers to flood publishers. If you have reviewed and rejected a paper before, you know how low the quality can be. Now, imagine a hundred times more of these for each event or issue. In my field of research, reviewers usually don't get paid, so there will not be "new jobs", just more unpaid work to get your conference going.
Contrary to most comments I see, I like these AI papers. It's super interesting and exciting to see where these developments go.
So like the twitter bots ruining the replay section, a new bot technique to bloat and destroy the pair review papers pool
never forget that these technologies can fall into the wrong hands
Like yours ?
@@user-jq8fw6yb6s Well, you can be sarcastic all you want. Of course, I am not an attacker and I don't have evil plans. But I understand and realize the risks of such technologies. A good ML assistant of a scientist can also suggest a solution to some complex problem to an attacker who lacks his own intelligence to solve it himself. And you will face a huge issue you did yourself and the consequences of which will be much more difficult to struggle with. What yesterday seemed like science fiction and a matter of the distant future is already at your doorstep. When that happens (I don't say 'IF'), it won't be funny at all.
Lol, they already have
And eventually the job of curator will be taken, once it's better than an undergrad you can only give it to a ever decreasing amount of people without people being in the mix causing it to function worse.
You may want to spend more time on the ai rewriting the rules to suit its own needs. 😅
🎶 Instrumental Convergence 🎶
so we're back to making GANs, but only this time with paper generators and peer reviewers? 😂😂
Your videos are excellent, what a time to be alive!
Why is the obviously required human not included in the marketing picture of this awesome invention?
If this is a tool, why is it advertised as "standalone"?
It still needs a companion that has a sufficient world model PLUS the ability to decide.
Why is there so much effort put into generating highly polished lorem ipsum?
How else is it going to improve? You have to fail many times before you can succeed.
When AI paper standard is even higher than you 😂
honestly i mostly just want to be able to find ai designed to play games available on the public access (free) regions
i want to make one play ksp. *With the ability for me to give it specific tasks or requests so i can direct its attention/focus for my own goals.*
Up until now, I've been feeling safe because doing the actual AI research will always need a human....right? Now I don't feel safe even for my AI related job.
I'd use it to write many papers. Of course I'd check everything is OK, but to start writing it would be very helpful.