@@JorgetePaneteI prompted: “say it like an infant”, and I swear to God it rephrased it’s paragraph into literally: “ga goo ga, goo goo ga ga” for like 2 paragraphs lol
It's like with teaching professors, there are smart ones that can't transfer that knowledge to the pupils and then there are good ones. I prefer the second bunch.
Progress is velocity, so it cannot _slow down,_ it can only _decrease._ Likewise "the rate of progress" is a pleonasm - unless you literally mean exponential growth. 🤓
@@brexitgreens Your interpretation of progress as a rate is not the only valid interpretation. Lots of contexts call for progress to be interpreted as a distance, independent of time. If someone asks you "what is your progress on your homework?" you wouldn't answer "about 10% per hour". That would be obtuse.
Those people are really coping, thinking AI is a trend that will come to a complete stop once it stops being popular. Once companies catch onto the next big buzzword thing and people stop trying to use it for everything (its too early for that), AI's reputation should improve.
The generated Einsten and the kid's pictures are creepy. Look so bad and artificial. The Internet is full of generated pictures now. I am sick of them already!
The idea is you can still ask for the concise final answer, but if you are interested in learning the concepts yourself then the AI used to be bad at that and introduce new mistakes. For AIs designed to teach this is a huge improvement and on the other hand overall correctness is still improving.
@@Khofax but if you wish to have it over explain every minute detail, why not explicitly ask for that? why must the default assume that we want to have over done explanations?
@@acters124I agree! Most people who are asking an AI something, already know what they are talking about, they just need a push in the right direction…
4:35 i don't get it i would argue is terrible for 99% of people. you are choosing to handicap your model so it can speak a bit clearer but what's the point? in the rare cases where an ai chatbot spits out something very complex you can just literally ask it to dumb it down and it just does. not to mention that i nor anyone else i know have never felt like an output from a chatbot was too profound for me to figure out what it said(it's not that i am super smart, chatbots just produce very understandable outputs) 1:55 think to yourself, do we really need to specially train the ai for such a usecase? what's wrong with having it use powers? if you are confused you can then ask it to clarify. am i missing something?
They aren't making it dumber for legibility, they are making it more legible at the same intelligence I think The answer probably shouldn't have powers cause the question says 3 times and has nothing to do with the
Bruh, as a kid i hated showing my work but understood why but still didn’t want to. This ai might get even better but also resources might be used more to show work or something.
It is very useful to learn what types of things the person you are talking to already knows and what types of things they don't know. Frontier models have been becoming less useful to me as they keep adding more and more information I don't need to their answers.
4:28 exactly like people, you don’t show your work then it’ll be hard for others to understand. It’s like trying to talk when you don’t know English but wanna share, gets difficult. That’s why explaining to a kid helps us simplify.
It seems this could be a way to demonstrate "P not NP", or is it just me hallucinating again? What a time to hallucinate! ☺ Thanks for the video, fellow scholar Károly.
It would be refreshing to hear your critique of symbolic regression for the retrieval of novel concepts from Ai (activation targeting and distillation). Especially Cranmer's techniques for deriving proofs.
Hope it isn't too long before we get a video on Flux, cause it's honestly such a huge step up compared to all it's competitors that it's still hard for me to believe
I wouldn't say that those examples all follow common rules. This is like teaching an AI to be that one remarkable professor who could actually teach well-the one who could answer your questions perfectly without going over your head or oversimplifying things to the point of being useless.
bc they are usually talk shit. Yesturday example: «The number 42 has several interesting mathematical properties. First, it is the only number that is both the sum of the cubes of its digits and their product. That is, $4^3 + 2^3 = 64 + 8 = 72$, and $4 \times 2 = 8$.»
My guess is that if the AI asks "do you want a more thorough explanation?" and if so creates the more accurate but less legible answer, then you have the best of both.
In a strange way, maybe a U-net style autoencoder could leverage something like this for image reconstruction. Like if there was some way to logically convert the latent layer's outputs to something more semantic, we could then use this to force the encoder to "dumb down" those embeddings and give more interpretable latent embeddings? Why you'd want that, and how you'd accomplish it, I have no idea.
I like the idea of GAN or model coupling or integrating a lot of specialized models and classic tools. It is on the surface, but modern LLMs can`t use memory (with ideas and thesises, accumulated from knowledge and previous talk)
How could it work on images? Segmentation maybe? Say for example a varifier knows what a hand should look like as a segmentation while the einstien could know what overall images should look like?
I really don't see any legibility problem with the solution shown at 1:54. To me, the explanations at 2:04 and 2:11 just seem like more verbose versions of the solution, repeating parts of the question and adding filler phrases in between to increase length. From the paper, I noticed two things that seem like big limitations for now: 1. They have only tested this method on grade school math problems, since it makes the dataset easy to generate. It seems like a stretch to assume that their observations will transfer to something like the sorting algorithm shown at 1:19. 2. The authors mention "checkability doesn’t necessarily capture everything we intuitively want from legibility" and that seems to be the main drawback. They have a definition for checkability which you can see in the paper, and they also tested a small subset of the results on paid humans to compare legibility scores with their own checkability metric, but they had a 45 second time limit there - and it seems quite probable that mistakes in longer solutions will often be harder to spot.
Legibility Tax is in direction of the horizontal Accuracy axis, which makes sense, because for a more legible output you DO have to pay the Legibility Tax. 4:32
It's similar, but different. The similarity is that this approach uses two separate models focusing on their roles. But the roles are different, specifically for the secondary model. In GAN it's role is simply to guess if something was produced by model or not.
I wonder if this would work for image generation, a dumbed-down verification each step. I use detailed prompts to generate anime waifu images, but the multiple objects can be messy, and usually not that reconizable compared to handdrawn/real world. A bathtub too small for the character in a bathroom, for example. AIs making sure the generation result is dumbed down sounds like a great possibility for handling such cases in image generation.
This reminds me of a professor I had who was supposedly very smart but not good at teaching. I believe that being able to teach something is just as important as understanding something, so I would weigh both in determining how smart someone really is in a subject.
Love it. I use AI as a user, and I’m actually replacing not having colleagues with AI (working alone). So I just want advices and good explanations to confort some choices. More pedagogy also helps in proofreading what the AI wrote manually
I do vastly prefer Mistral to the open AI models. This video is helping me understand why. Possibly helping to find a better model in the future too. LLama too, it's pleasant. I only need small things troubleshooted and it gets it correct enough. Programming aid is great since i can just implement the suggestions and instantly see if it works or not. So correctness is not that big of a deal. One shot correctness even less. It's good to have, but not that necessary. I'm a novice programmer. Just learning. And oh boy those AI are a great learning tool. I'm going to do a small rant about them as learning tools. People say it's a bad thing to use AI to cheat on your homework or whatever but i really think it's more like a personal teacher. You can't just copypaste the answers. The teacher will know. Unless you can specially train the model to output your style. And that's not easy. If the kid can pull that off, yea they got enough skills to get a job in the future even if their history knowledge or whatever is lacking.
Google's "AI overview" is uniformly incorrect in every even slightly technical question I ask it. I asked it "give me a list of materials by relative permittivity" and one of the answers was "animal organs and human blood: about the same as other liquids" also "bakelite: 1-100" (these are both wrong, btw). Wake me up when any of this works for a question more technical than "what color is a green apple?". What a time to be disappointed.
AI legibility, AI hallucination, AI alignment-these terms sounded like science fiction just a few years ago. Now, they feel mundane. I think history will remember the 2020s as the decade of the AI revolution.
Smart != Intelligent. Intelligent actions are those which benefit both the actor, and others. This is the definition of intelligence, according to Deitrich Bonhoeffer, and his Theory Of Stupidity.
Why would you need a machine to think for you, when dealing with politics? You should think for yourself, as the political results have an effect on you.
@@Felipe3001miranda I asked AI to tell me the relation between pedophilia and homosexuality and I received a warning saying they cant talk about that. It is biased.
whatever in short they just made a verifier model to boost the accuracy of their models on questions they are capable of solving and minimize overall hallucinations. Nothing new. I'd be more interested in seeing how small models work together to achieve much better results than a bigger model
0:14 am I tripping ? the loop will never end bc j starts off being larger than 0 and it'll never be less than 0... is this why the human is going "???"
Correction at 4:33 - you ARE paying the legibility tax, which is why the blue model is not as smart as the green. That's the tax you pay for making a smarter model human understandable.
Some people call this artificial intelligence, but the reality is this technology will enhance us. So instead of artificial intelligence, I think we’ll augment our intelligence.
Been doing this the whole time. I use custom ai's with my own data and test the answers on other ai's amd different sessions in a way so i can make complex topics like existence more clear
I don't think people are rating the current state of the AI but rather the possibility of what if this thing becomes 5 or 10 times better within the decade
If AI are getting so smart that we now need them to dumb down their answers enough for us to understand them, then that really shows just how much they're outpacing us.
It's about the understandability of answers, not how complex they are. Some people have a knack of giving an answer to a question that is clear and concise, even if the answer is complicated. Some are awful by being convoluted even when giving a simple answer.
More like "being a midwit is overrated." Topwits are categorically different; they're not Masters of Their Field as FA Hayek talked about but rather Hayek's Puzzlers/Muddlers, the kind that do fully independent thinking, not riding the tails of existing social or academic notions.
It's not even conscious yet and we're already struggling to understand our AI. Humans won't be able to claim they're intelligent in a few more years...
This is absolutely the holy grail (besides the so called (artificial general intelligence") , because the more complex problems AI solves the less we humans can reproduce them ourselves. So.. unless we don't fuse ourselves with the future AI, we have to be able to at least understand the chunks of its doing.
And yeah about your question if it was capable as some kind of diffusion model.. I think my first guess would be like vector based graphics. Maybe also in the animations and game industry where mathematics is a key part of understanding a visual representation. Whereas pixel-per-pixel based it would require an integrated structure with two or more levels of (let's call it) design. For example: First.. ground structure (define a general pixel distribution by means of either an array of orthogonal vectors or introduce chaos with randomized vectors (seeds will determine the outcome and shall be handled cautiously) ), second... Rearrange the pixels by vectorizing a few of them at a time and put them to a specific test (condition) and by conditioning them and introducing more vectors one can achieve controlled diffusion (at small (diffusion) steps, in this case case by introducing a specific integration time), third.. the expression it has to fulfill by applying a mathematical filter (gaussian, e.t.c)
fascinating paper. has implications for epistemology and moral problems that society is increasingly going to face because of the increase in AI and robots, but also because of the worsening climate and environment situation. two things i would love to be able to improve which this paper may help with: coding (AI still not good at many real-world coding problems) and verification of facts, especially medical facts, where lives are on the line.
there were so many outstanding mathematicians to choose from, but they had to name it Einstein after the physicist because he is more popular in pop culture, terrible
I couldn't find a single Ai which provided me a completely good code (c# python). I spent more time fixing the problem for Ai rather that write it down by myself. Naughty Naughty....
On one side this is cool, on the other, this will mostly only help the dumb people and waste time of smarter people. Whenever you increase the number of lines 2 or 3x to better explain something, you are also wasting tokens aka time and money. There are prompts like "explain like I am 5" which basically does the same thing already, so now we will have to add "explain like I am Einstein" I guess..
How do we know that the calculator is really smart? We give it a bunch of math equations to solve, and see how well it does. Spoiler, a calculator is not smart, neither is AI, it’s just an even bigger calculator.
Nope. I tested with actual prompts. Directly (Claude, GPT, LLama) and in blind competition I asked to make reverted list of toys. With a little mistake (Enumerate 4-points list from 5 to 1). Target result: 5 lego parts 4 juggling balls 3-head dragon 2 wooden cubes Only one LLM (Gigachat) did it after 3 prompts, with examples, in a row, but not perfectly. Most of them failed on reverted list. They are still statistic parrots without any sense of structure. And they can repeat only structures, that they mostly have seen. Narrow-specialized networks are great and amazing. Alpha-fold is revolution. But not LLMs.
@@bzikarius You know benchmarks already exist and are for more robust than your ad hoc one, right? And you're aware than an AI got silver (nearly gold) in the IMO?
@@Dave-rd6sp Look up "Alice in Wonderland: Simple Tasks Showing Complete Reasoning Breakdown in State-Of-the-Art Large Language Models", you will see AI get stupid when the problem is not in the training data (akka those gold benchmarks you mention)
@@Dave-rd6sp Yes, let`s teach model patterns and then ask about SAME patterns. It solves (mostly)! WHOA! JFYI: Chat GPT do not solve math tasks itself, it uses Wolfram Alpha. Stop be amazed newbie, stop trust cherry-picked results, and dive deeper. NN are still amazing, they works kind of intuition or insanely big filter. Best AI (now, with current architecture) should be kinda «autistic-savant» - it is safe and effective. Do you remember «pizza with glue» case? My little brother (he was kinda 5 yo) once walked back home with bread and suddenly dropped it. So he was afraid, and decided to wash it. Should I say, he never washed bread again? But LLMs «wash bread» over and over. Despite gigawatts of energy and billion dollars. They still can`t remember any ideas, rules, structures. The answers are checked with filter (kind «do not say GINGER» and «Celtic men can be black asian even if it is 15000 years ago») Yesterday Gemini said, that «42 is prime number BECAUSE it divides ONLY on 2 and itself».Nuff said. And the sad part. In my country kids should pass common complex test after school. So they are mostly learn how to pass test. And they usually pass with high score. Test contains errors, and to pass it, they reply wrong too. But mostly they are still extremely dumb, can`t build and solve real tasks, can`t think critically or creatively… This is the issue for natural intellect as you see.
What a time to be alive!
Not sure if that's not a threat 😅
AI: What a time to be artificially "alive"!
What a time to be AI!
🔫 _2024: No Time To Be Unalived_
@@brexitgreens 😎 Agent 000, Licensed to Chill
It reminds me of that classic prompt: "Explain it to me like if I am a 10 years old"
ELI5
I sometimes do "ELI5, then summarize it using caveman speak, then summarize the summary using caveman speak".
@@JorgetePaneteI prompted: “say it like an infant”, and I swear to God it rephrased it’s paragraph into literally: “ga goo ga, goo goo ga ga” for like 2 paragraphs lol
Me explica como se eu fosse uma criança de 10 anos de idade
This is hilarious @@virtualgrowhouse
It's like with teaching professors, there are smart ones that can't transfer that knowledge to the pupils and then there are good ones. I prefer the second bunch.
What time to be AI
Makes sense. It's easier to verify than to prove.
en.wikipedia.org/wiki/P_versus_NP_problem
that's not a general rule. True for NP style problems.
but for example easier to write a perl program/huge regexp/..., than understand one.
@@eleklink8406 Is writing a program really the same as proving something?
people talking about how the rate of ai progress is slowing down, and then you post something like this LOL. What a time to be alive fr.
Progress is velocity, so it cannot _slow down,_ it can only _decrease._ Likewise "the rate of progress" is a pleonasm - unless you literally mean exponential growth. 🤓
@@brexitgreens Your interpretation of progress as a rate is not the only valid interpretation. Lots of contexts call for progress to be interpreted as a distance, independent of time. If someone asks you "what is your progress on your homework?" you wouldn't answer "about 10% per hour". That would be obtuse.
Those people are really coping, thinking AI is a trend that will come to a complete stop once it stops being popular. Once companies catch onto the next big buzzword thing and people stop trying to use it for everything (its too early for that), AI's reputation should improve.
@@michaelleue7594 😆
@@michaelleue7594 "About 10% per hour." 😆
The generated Einsten and the kid's pictures are creepy. Look so bad and artificial. The Internet is full of generated pictures now. I am sick of them already!
Ugh. The first answer was very clear and then the training made it way worse. I dont want my answers hidden in a wall of text.
The idea is you can still ask for the concise final answer, but if you are interested in learning the concepts yourself then the AI used to be bad at that and introduce new mistakes. For AIs designed to teach this is a huge improvement and on the other hand overall correctness is still improving.
@@Khofax but if you wish to have it over explain every minute detail, why not explicitly ask for that? why must the default assume that we want to have over done explanations?
@@acters124I agree! Most people who are asking an AI something, already know what they are talking about, they just need a push in the right direction…
@@acters124because eventually it will know what YOU know and will be able to answer accordingly
I think the problem is trying to keep humans (children) in the loop. The AI is smarter than us now. We need AI to verify AI and work with AI.
4:35 i don't get it
i would argue is terrible for 99% of people. you are choosing to handicap your model so it can speak a bit clearer but what's the point? in the rare cases where an ai chatbot spits out something very complex you can just literally ask it to dumb it down and it just does. not to mention that i nor anyone else i know have never felt like an output from a chatbot was too profound for me to figure out what it said(it's not that i am super smart, chatbots just produce very understandable outputs)
1:55 think to yourself, do we really need to specially train the ai for such a usecase? what's wrong with having it use powers? if you are confused you can then ask it to clarify.
am i missing something?
What do you mean
They aren't making it dumber for legibility, they are making it more legible at the same intelligence I think
The answer probably shouldn't have powers cause the question says 3 times and has nothing to do with the
@@D---3 If you go to the first timestamp I think it becomes clear, same level of understandability at the cost of intelligence
Oh wait nvm they are making a model smarter without paying a legibility tax
So the accuracy is higher without reducing legibility
@@D---3 No, that's a mistaken in the narration. They _are_ paying the "legibility tax" in intelligence.
I think Dr. Zsolnai is an AI. He literally sounds like a soundboard on all these videos.
Yep, he is. At least the voice😁
Perhaps is it true. His live presentation was far different.
May be it is AI-filter. Compare with videos 5 years ago.
When released this will be like the beginning of Agents being an intrinsic part of mainstream LLMs
Will be rolling out in the coming weeks.
My belly?
Bruh, as a kid i hated showing my work but understood why but still didn’t want to. This ai might get even better but also resources might be used more to show work or something.
It is very useful to learn what types of things the person you are talking to already knows and what types of things they don't know. Frontier models have been becoming less useful to me as they keep adding more and more information I don't need to their answers.
I have the same feeling. It's like everything you asks it spits a wikipedia article... and we already have wikipedia for that
@@marcosfraguela yup, and as soon as it thinks to create a list, you know you are about to get an entire page of useless crap.
4:28 exactly like people, you don’t show your work then it’ll be hard for others to understand. It’s like trying to talk when you don’t know English but wanna share, gets difficult. That’s why explaining to a kid helps us simplify.
I'm going to send that thumbnail in the boys groupchat whenever they push another squad in game.
😂
It seems this could be a way to demonstrate "P not NP", or is it just me hallucinating again? What a time to hallucinate! ☺ Thanks for the video, fellow scholar Károly.
It would be refreshing to hear your critique of symbolic regression for the retrieval of novel concepts from Ai (activation targeting and distillation). Especially Cranmer's techniques for deriving proofs.
Hope it isn't too long before we get a video on Flux, cause it's honestly such a huge step up compared to all it's competitors that it's still hard for me to believe
What is Flux?
@@smyk1975open source image generator. very high quality results and truly open source.
@@smyk1975 new sota-level image model dropped out of nowhere
A grade school student is usually required to show their work.
A PHD student is required to defend their thesis.
Why not force AI to explain as well?
I wouldn't say that those examples all follow common rules. This is like teaching an AI to be that one remarkable professor who could actually teach well-the one who could answer your questions perfectly without going over your head or oversimplifying things to the point of being useless.
bc they are usually talk shit.
Yesturday example:
«The number 42 has several interesting mathematical properties. First, it is the only number that is both the sum of the cubes of its digits and their product. That is, $4^3 + 2^3 = 64 + 8 = 72$, and $4 \times 2 = 8$.»
I love that you use a.i to clone your voice! Just needs a better microphone for best recording quality.
My guess is that if the AI asks "do you want a more thorough explanation?" and if so creates the more accurate but less legible answer, then you have the best of both.
In a strange way, maybe a U-net style autoencoder could leverage something like this for image reconstruction. Like if there was some way to logically convert the latent layer's outputs to something more semantic, we could then use this to force the encoder to "dumb down" those embeddings and give more interpretable latent embeddings? Why you'd want that, and how you'd accomplish it, I have no idea.
I like the idea of GAN or model coupling or integrating a lot of specialized models and classic tools. It is on the surface, but modern LLMs can`t use memory (with ideas and thesises, accumulated from knowledge and previous talk)
"So, how do we make them smarter?"
"That's the neat part! You DON'T"
Truly, what a time to be alive!
How could it work on images?
Segmentation maybe?
Say for example a varifier knows what a hand should look like as a segmentation while the einstien could know what overall images should look like?
I really don't see any legibility problem with the solution shown at 1:54. To me, the explanations at 2:04 and 2:11 just seem like more verbose versions of the solution, repeating parts of the question and adding filler phrases in between to increase length. From the paper, I noticed two things that seem like big limitations for now:
1. They have only tested this method on grade school math problems, since it makes the dataset easy to generate. It seems like a stretch to assume that their observations will transfer to something like the sorting algorithm shown at 1:19.
2. The authors mention "checkability doesn’t necessarily capture everything we intuitively want from legibility" and that seems to be the main drawback. They have a definition for checkability which you can see in the paper, and they also tested a small subset of the results on paid humans to compare legibility scores with their own checkability metric, but they had a 45 second time limit there - and it seems quite probable that mistakes in longer solutions will often be harder to spot.
Legibility Tax is in direction of the horizontal Accuracy axis, which makes sense, because for a more legible output you DO have to pay the Legibility Tax. 4:32
Isn't this architecture same as GAN?
Oh no, totally different. They used the words Einstein and child, not generator and discriminator! /s
It's similar, but different. The similarity is that this approach uses two separate models focusing on their roles. But the roles are different, specifically for the secondary model. In GAN it's role is simply to guess if something was produced by model or not.
I wonder if this would work for image generation, a dumbed-down verification each step. I use detailed prompts to generate anime waifu images, but the multiple objects can be messy, and usually not that reconizable compared to handdrawn/real world. A bathtub too small for the character in a bathroom, for example. AIs making sure the generation result is dumbed down sounds like a great possibility for handling such cases in image generation.
Research paper is available for free? What a time to be alive!
This reminds me of a professor I had who was supposedly very smart but not good at teaching. I believe that being able to teach something is just as important as understanding something, so I would weigh both in determining how smart someone really is in a subject.
The legibility tax was along the X axis not along the Y, and the chart showed you are paying it by having a legible model that’s less accurate.
Your Two Minute Papers are absolutely fabulous - especially this one
Me when my grades start slipping:
"Being smart is overrated!"
📝 Dr. Alan Thompson: *"On the new irrelevance of intelligence".* An old paper which has aged extremely well. (Citing from memory.)
Love it. I use AI as a user, and I’m actually replacing not having colleagues with AI (working alone). So I just want advices and good explanations to confort some choices.
More pedagogy also helps in proofreading what the AI wrote manually
I do vastly prefer Mistral to the open AI models. This video is helping me understand why. Possibly helping to find a better model in the future too. LLama too, it's pleasant. I only need small things troubleshooted and it gets it correct enough. Programming aid is great since i can just implement the suggestions and instantly see if it works or not. So correctness is not that big of a deal. One shot correctness even less. It's good to have, but not that necessary.
I'm a novice programmer. Just learning. And oh boy those AI are a great learning tool.
I'm going to do a small rant about them as learning tools.
People say it's a bad thing to use AI to cheat on your homework or whatever but i really think it's more like a personal teacher. You can't just copypaste the answers. The teacher will know. Unless you can specially train the model to output your style. And that's not easy. If the kid can pull that off, yea they got enough skills to get a job in the future even if their history knowledge or whatever is lacking.
Einstein AI: "How do you do fellow kids?"
Google's "AI overview" is uniformly incorrect in every even slightly technical question I ask it.
I asked it "give me a list of materials by relative permittivity" and one of the answers was "animal organs and human blood: about the same as other liquids" also "bakelite: 1-100" (these are both wrong, btw).
Wake me up when any of this works for a question more technical than "what color is a green apple?".
What a time to be disappointed.
AI legibility, AI hallucination, AI alignment-these terms sounded like science fiction just a few years ago. Now, they feel mundane. I think history will remember the 2020s as the decade of the AI revolution.
Hey this is almost approaching my triumverate model. So funny that it took OpenAI until here to get to where I was at during GPT 3.0
Very cool. You should work for them. Show them you did what they have only accomplished now 2 years ago and they might offer you a good salary
@@chasingdaydreams2788 I'm not willing to move my family from my country. Otherwise I would.
DM [0:33-2:15]
Accuracy vs. Legibility
I want GPT5 now
Smart != Intelligent.
Intelligent actions are those which benefit both the actor, and others.
This is the definition of intelligence, according to Deitrich Bonhoeffer, and his Theory Of Stupidity.
Now for them to lift the "I can't talk about that" political blockages and we're getting somewhere
Yeah the social guardrails are pretty lame. Especially given that what's politically correct now will be outdated and offensive in 10 years.
Why would you need a machine to think for you, when dealing with politics? You should think for yourself, as the political results have an effect on you.
This is such a "I want to use it for immoral/morally gray things but the AI doesn't let me" comment
@@Felipe3001miranda I asked AI to tell me the relation between pedophilia and homosexuality and I received a warning saying they cant talk about that. It is biased.
What, salty that OpenAI won't let you generate anti-EU slop for your russian smurf account?
whatever in short they just made a verifier model to boost the accuracy of their models on questions they are capable of solving and minimize overall hallucinations.
Nothing new. I'd be more interested in seeing how small models work together to achieve much better results than a bigger model
astounding work 2mins on the job of verifier mode
Dang it my papers just blew off my desk!
0:14 am I tripping ? the loop will never end bc j starts off being larger than 0 and it'll never be less than 0... is this why the human is going "???"
Will be great to apply to all publications. All publications should be more understandable.
They should use it to prevent hallucinations and overcomplexity
What alive to be a time !
Thank you for the quick heads-up !
Nice. One thing, the language used in the video does (at least for me) raise one question: is this an OpenAI sponsored video? Cheers
"If you can't explain it to a 4 year old, you don't understand it yourself" -things Einstein never said
Great breakdown of AI’s complexity vs. understandability trade-off! Fascinating research from OpenAI.
Correction at 4:33 - you ARE paying the legibility tax, which is why the blue model is not as smart as the green. That's the tax you pay for making a smarter model human understandable.
Some people call this artificial intelligence, but the reality is this technology will enhance us. So instead of artificial intelligence, I think we’ll augment our intelligence.
Been doing this the whole time. I use custom ai's with my own data and test the answers on other ai's amd different sessions in a way so i can make complex topics like existence more clear
A kid that's corrects the expert, also known as Asperger's. 💀💪
That's like complaining about a female boxer for being too strong 😏
I love the creative thinking by those authors.
technically, the kid could approach log size of the Einstein, right? At least in terms of the problem difficulty the Einstein model can solve.
I don't think people are rating the current state of the AI but rather the possibility of what if this thing becomes 5 or 10 times better within the decade
Its kind of like the magi from Evangelion except without real brains
Spooky, open ai needed a new technique to increase legibility of smarter models
If AI are getting so smart that we now need them to dumb down their answers enough for us to understand them, then that really shows just how much they're outpacing us.
It's about the understandability of answers, not how complex they are. Some people have a knack of giving an answer to a question that is clear and concise, even if the answer is complicated. Some are awful by being convoluted even when giving a simple answer.
Hmm, if doctors can not see through the Einstein's lies, how we can expect kids to see through?
I’m not surprised.
Humans evolve intelligence this way too.
More like "being a midwit is overrated." Topwits are categorically different; they're not Masters of Their Field as FA Hayek talked about but rather Hayek's Puzzlers/Muddlers, the kind that do fully independent thinking, not riding the tails of existing social or academic notions.
ELI5 for AI
Objective maximizer
It's not even conscious yet and we're already struggling to understand our AI.
Humans won't be able to claim they're intelligent in a few more years...
So the "Kid" knows that "Einstein" is wrong, despite not knowing itself what the right answer is? How the heck does that work?
Same way you can read and criticize a novel, while being far from capable of writing it yourself
Some kind of training to let it spot hallucinations?
it doesnt. the accuracy is significantly lowered using this method which means the verifier hallucinates
Reading old mathematical, astronomical, and historical scripts with unintelligible writing.
This is absolutely the holy grail (besides the so called (artificial general intelligence") , because the more complex problems AI solves the less we humans can reproduce them ourselves. So.. unless we don't fuse ourselves with the future AI, we have to be able to at least understand the chunks of its doing.
And yeah about your question if it was capable as some kind of diffusion model.. I think my first guess would be like vector based graphics. Maybe also in the animations and game industry where mathematics is a key part of understanding a visual representation. Whereas pixel-per-pixel based it would require an integrated structure with two or more levels of (let's call it) design. For example: First.. ground structure (define a general pixel distribution by means of either an array of orthogonal vectors or introduce chaos with randomized vectors (seeds will determine the outcome and shall be handled cautiously) ), second... Rearrange the pixels by vectorizing a few of them at a time and put them to a specific test (condition) and by conditioning them and introducing more vectors one can achieve controlled diffusion (at small (diffusion) steps, in this case case by introducing a specific integration time), third.. the expression it has to fulfill by applying a mathematical filter (gaussian, e.t.c)
42
fascinating paper. has implications for epistemology and moral problems that society is increasingly going to face because of the increase in AI and robots, but also because of the worsening climate and environment situation. two things i would love to be able to improve which this paper may help with: coding (AI still not good at many real-world coding problems) and verification of facts, especially medical facts, where lives are on the line.
Video uploaded 12 sec ago
Ur comment 7hr ago 😵💫
@@systembreaker4864 Just another Matrix bug 🤷♂
@@IlyaIlya-ef4nz it's happening a lot just yesterday i saw same thing
there were so many outstanding mathematicians to choose from, but they had to name it Einstein after the physicist because he is more popular in pop culture, terrible
Hello random person on the internet 👋
I DON'T KNOW YOU! THAT'S MY PURSE!
Hello, why are you stealing their pourse?
Quit stealing peoples purses bro. Quick, someone catch this guy!
what a time to be alive [2]
Can they reason from first principles?
What a time to be alive!
as a smart person: title is VERY TRUE. fml
I couldn't find a single Ai which provided me a completely good code (c# python). I spent more time fixing the problem for Ai rather that write it down by myself. Naughty Naughty....
Resistance is futile!
What a time to simulate being alive!🤖😛
gpt-4o is sometimes worse than gpt-4 legacy.
On one side this is cool, on the other, this will mostly only help the dumb people and waste time of smarter people. Whenever you increase the number of lines 2 or 3x to better explain something, you are also wasting tokens aka time and money. There are prompts like "explain like I am 5" which basically does the same thing already, so now we will have to add "explain like I am Einstein" I guess..
How do we know that the calculator is really smart? We give it a bunch of math equations to solve, and see how well it does. Spoiler, a calculator is not smart, neither is AI, it’s just an even bigger calculator.
What a time to be alive
Best channel I've ever joined.
Everything until AGI will be overrated
Man I love u 2 minutes paper u are god for curious students like me
You are too kind, thank you so much! 🙏
@@TwoMinutePapers thanks so much fro your time to reply u made my day
What a time to be a lie!
When statistic parrot can pretend to be so smart as PHD, but can`t solve simpliest task, as my 9-year old nephew can
The irony of this post being a complete lie.
Nope. I tested with actual prompts. Directly (Claude, GPT, LLama) and in blind competition
I asked to make reverted list of toys. With a little mistake (Enumerate 4-points list from 5 to 1). Target result:
5 lego parts
4 juggling balls
3-head dragon
2 wooden cubes
Only one LLM (Gigachat) did it after 3 prompts, with examples, in a row, but not perfectly.
Most of them failed on reverted list.
They are still statistic parrots without any sense of structure. And they can repeat only structures, that they mostly have seen.
Narrow-specialized networks are great and amazing. Alpha-fold is revolution. But not LLMs.
@@bzikarius You know benchmarks already exist and are for more robust than your ad hoc one, right? And you're aware than an AI got silver (nearly gold) in the IMO?
@@Dave-rd6sp Look up "Alice in Wonderland: Simple Tasks Showing Complete Reasoning Breakdown in State-Of-the-Art Large Language Models", you will see AI get stupid when the problem is not in the training data (akka those gold benchmarks you mention)
@@Dave-rd6sp Yes, let`s teach model patterns and then ask about SAME patterns. It solves (mostly)! WHOA!
JFYI: Chat GPT do not solve math tasks itself, it uses Wolfram Alpha.
Stop be amazed newbie, stop trust cherry-picked results, and dive deeper. NN are still amazing, they works kind of intuition or insanely big filter.
Best AI (now, with current architecture) should be kinda «autistic-savant» - it is safe and effective.
Do you remember «pizza with glue» case? My little brother (he was kinda 5 yo) once walked back home with bread and suddenly dropped it. So he was afraid, and decided to wash it. Should I say, he never washed bread again? But LLMs «wash bread» over and over. Despite gigawatts of energy and billion dollars. They still can`t remember any ideas, rules, structures. The answers are checked with filter (kind «do not say GINGER» and «Celtic men can be black asian even if it is 15000 years ago»)
Yesterday Gemini said, that «42 is prime number BECAUSE it divides ONLY on 2 and itself».Nuff said.
And the sad part. In my country kids should pass common complex test after school.
So they are mostly learn how to pass test. And they usually pass with high score. Test contains errors, and to pass it, they reply wrong too.
But mostly they are still extremely dumb, can`t build and solve real tasks, can`t think critically or creatively… This is the issue for natural intellect as you see.
How do we use this?
yaaay, making sure AI will still be able to talk to humans in 5 years :D
Every time i hear Open AI i remember that the Open isn't Open
Reminds me of GAN
okay. After allowing ai being connected to internet, human now starts teaching ai how to lie. Congratulations!
Are we still having war over human origin or something...?