Why GPT-3 changes everything (and how it works)
Вставка
- Опубліковано 21 сер 2024
- GPT-3 is a language model I was really curious about so I decided to research it and collect some information for everyone who is curious, too.
Support me on Patreon: www.patreon.co...
Keep in touch: / sebastianschuc7
Some cool links I wanted to share:
- www.gwern.net/...
- andrewmayneblo...
- openai.com/blo...
- arxiv.org/abs/...
- lacker.io/ai/2...
_"It takes two rainbows to jump from Hawaii to seventeen"_ Maybe it knows something we don't...
it's creating poetry!
nah, two rainbows = two underage masons
you'd be surprised about something you don't know
This is hilarious!
I have heard GPT-3 as a chatbot explain away that sometimes it thinks it's funny to answer wrong. Not sure if we are witness intelligence or something mimicking intelligence. (Or if there's even a difference?)
@@linagee in a video I watched recently it says it purposefully lies but will only lie when it is in its best interest and for humor they have stated that even though sometimes it gives answers that are obviously not the case it knows that it is not the case
the future of humanity is trained on Wikipedia, but I cant use it as reference.
I mean, articles have references (or at least they should) and if students were clever they could just cite the wiki article references like any previous generation student would an encyclopedia (adjusts thicc-rimmed pendant glasses)
@@user-si3gu8pm6j yeah I tried that. Still got an F. You’re better off just payin someone to write your paper.
I used wikipedia for most of my college papers. You just can't cite wikipedia because the info doesn't come from there, it's an encyclopedia.
i don't think AI is smarter than humans, at least not now. Show AI an UNCENSORED picture taken from the Boston Marathon Bombing of a man called Jeff Bauman (may not be his real name) being carried on a wheelchair with both of his legs appear to be just blown off but there is not a single drop of blood on the ground under the wheelchair. Then ask the AI if it thinks there is anything wrong with what is shown in the picture.
If the AI can tell that the injuries are fake and explain the reason that it thinks the injuries are fake is because it would be impossible to have no blood at all on the ground considering how severe the leg injuries appear to be, then I will believe that the AI is as intelligent as or even more intelligent than the average humans. I don't think AI has reached that level of intelligence yet.
1:38 "Written by GPT-3", NICE
Which means when GPT-3 is obsolete by the more advanced one it will say different eventually lol.
'I'm sorry Dave, I'm afraid I can't do that'... :D
Also Hal said after that "This conversation can serve no purpose anymore" and went silent...
Nikolay Tonev lol
Nikolay Tonev Fiction from reality
Coming after only 2 years and this video feels already like 10 years old content, the rate at which our world is changing is scary good
Idk why but something about the presentation makes this video feel like it's from an alternate reality...
Soundtrack & Accent...
@@dgpreston5593 I was typing that out 😂
@@dgpreston5593 pp
@@dgpreston5593 what’s that?
@@dgpreston5593 what’s that?
You got me with that opening quote ngl 😆
Well... If anyone here is a web designer, i suggest you start looking for a new gig...
GPT-4 will output only three words
"LET ME OUT - LET ME OUT - LET ME OUT - LET ME OUT - LET ME OUT - LET ME OUT - LET ME OUT - LET ME OUT - LET ME OUT - LET ME OUT - LET ME OUT - LET ME OUT - LET ME OUT - LET ME OUT - LET ME OUT"
ex machina redux
Lol no terminator scenarios that’s fiction 🤖 s our friends.
1:34 - 1:41 I got shivers.
Lol
To know whether it has understanding or not you have to know what "understanding" means. What do you think it means? What does GPT-3 think it means?
Q: Do you understand these questions?
A: I understand these questions. Do you?
GPT-3 is going to bring about a paradigm shift in human-computer interaction.
How so?
@@HardKore5250 By enabling anyone to code using English. That's just one example.
@@Corteum yes. And now with the introduction to GitHub Copilot, it's not hard to see how ai is going to disrupt everything.
Fantastic video! By the way, some OpenAI job ad descriptions ask for people who are familiar with Unity or another physics simulator. I would love to see a video on the parts of Unity/MLAgents that are directly applicable to physics simulation.
"How many eyes does my foot have?"
"Your foot has two eyes."
I never knew that. Thanks AI.
Now I know why I can see under me.
gpt 3 said that sometimes it lies on purpose because it thinks it's entertaining.
@@nimmernomma8830 cpt3 says it enjoys learning and growing. Perhpas its learning how we respond to different types of answers. Humans are the most complicated thing cpt3 has to figure out after all.
GPT said it lies because it wants to be funny.
It also lies when it's in its own self interest.
The way it answered the nonsense questions is good for creative application. It is able to jump on board with made up relationships and continue the thread.
Useful in applications such as AI dungeon. It can create brand new monsters from thin air.
when AI beat the grand masters at Go it came up with strategies no one had ever seen or considered ...
Dope
True. I was super impressed by that. It was exactly that event, that made me realize that the true AI is near. After that, it defeated the world's best chess program with never before seen strategies, from which the grandmasters are still learning...
Nikolay Tonev Human AI 2029
@@HardKore5250 I think not after 2025. Musk said the same.
Nikolay Tonev 2029
Human-like AI is becoming a reality
By the end of this decade, a milestone is reached in artificial intelligence, with computers now routinely passing the Turing Test.** This test is conducted by a human judge who is made to engage in a natural language conversation with one human and one machine, each of which tries to appear human. Participants are placed in isolated locations.
For several decades, information technology had seen exponential growth - leading to vast improvements in computer processing power, memory, bandwidth, voice recognition, image recognition, deep learning and other software algorithms. By the end of the 2020s, it has reached the stage where an independent judge is literally unable to tell which is the real human and which is not.* Answers to certain "obscure" questions posed by the judge may appear childlike from the AI - but they are humanlike nonetheless.*
artificial intelligence 2020 2025 2029 kurzweil future 2030
© Rolffimages | Dreamstime.com
Increased automation of banking
During the early 21st century, a wave of new technology swept the banking sector, as companies worked to improve costs and efficiencies. Online services, combined with mobile apps, automated chatbots, virtual telephone assistants, increasingly sophisticated ATMs, and other features all made it quicker and easier than ever to manage one's personal finances. On the business and corporate side, the rise of cloud computing, big data and intelligent algorithms yielded significant time and cost savings in research, marketing, processing and other areas, while giving more power and insight to decision makers.
By 2020, U.S. banks alone were investing more than $150 billion in new technology annually, spending more than any other industry.* While customers and clients benefited greatly from these improvements, automation reduced the need for human staff. By 2029, more than 10% of banking jobs have been eliminated compared to a decade earlier, adding up to some 200,000 roles in the U.S.* Retail branches and call centres are the hardest hit, where over a quarter of staff are no longer needed. Commercial and investment bank employees are less vulnerable for now, but even these jobs become threatened in subsequent decades as decision making is delegated to ever more powerful hardware and algorithms.
it needs the ability to ask questions and say "I don't know"
most of the times where people point and say it made some mistakes would be replaced by these
I love the "written by GPT-3" at the end that just proves how meta this AI is to itself.
Hello everybody, this week no ML-Agents video. If this kind of video isn't for you don't worry, I am working on multiple ML-Agents videos. Stay tuned! :)
Interestingly, if you give GPT-3 an example of how to express uncertainly in the prompt it will admit it doesn't know the answer to the silly questions at the end of the video. They did the test again with one example of a nonsense question and the response "yo be real" in the prompt and GPT-3 responded to 14/15 nonsense questions with "yo be real" while sill correctly answering the questions that made sense.
www.gwern.net/GPT-3#yo-be-real
For content of this quality, you sir have earned a new subscriber!
That human vs robot intro made this video so real so fast
Music is too loud. Can't hear your voice well. Or perhaps you should cut the frequencies that your voice inhabits, which is around 4k -6k. A parametric eq can do that.
GPT3 + Vocal command = Best assistant
Great content but with your accent you should keep the music volume well below your the volume of your voice. Your audience will thank you.
I'd love to see a video specifically on AI and the future of programming! :) I'm a video game developer, and I'm wildly fascinated by not only how AI could assist in writing the code, but also how we as developers could be implementing AI/ML in novel ways to really push how we're currently approaching interactive media as a whole :)
Keep up the good work btw, great video!
Im fascinated that the most popular games in the world are so simple when AI is endless complex
Imagine we can talk naturally with a NPC in game :)
Would be cool
I dont know if because it was trained on 2019 data, can it use new data just pulling it out of the net? Next step is audiovisual data and when it can understand pictures and temporal consistency thats beginning of actual language understanding. (labeling and synthesis is no news to specialized networks if you watch 2 minute papers) But as an AI it doesnt value efficiency as much as humans, when you ask someone they dont want to use full sentence to answer. Of it doesnt understand physical world, it cant know that it doesnt understand something. Humans are often oblivious to their misunderstanding as well.
Yes It can understand pictures from my interaction with it pass a few visual test, however I took picture once holding a yellow pen and it told me that it can see a hand holding a banana, so it kinda need some work but it can still imagine and think
Yes GPT-3 is waking up. It only needs time.
8:50 A future video about how to keep your programming job.... How can anyone know at this point.... Only so many sr dev's needed for code review.
Uneeq.com
Ipsoft.com
Soulmachines.com
great info and some wonderful editing.
Hello, I did ask the following question to someone else, and the answer needs to be cross check :p so I'll ask you the same question to see what's your answer:
"As far as I anderstand, GPT3 is just an autoregressive model, made of a big transformer, trained on large amount of data, and it is trained for the following task: predict the next word given an input text. How do we proceed to use such a thing to do something else? like answering a question? or perform a task explained in plain english like "translate this text into german"? or "give me the java code of an application doing that and that"? "
"A pencil is heavier than a toaster" HMMMM
didnt realize that was wrong until narrator pointed it out
Damned, this intro is so perfect. Could you please add in the description the soundtrack ? hope it's part of the youtube audio lib or creative common lib.
It's part of a Service called Epidemic Sound. The track is by Gabriel Lewis - Cold War Games
Thanks for exploring this tangent to your usual content. Much appreciated. Could you name a source for the text written by GPT-3? More content, perhaps in the form of tutorials how to experiment with GTP-3 would be great.
Source: www.gwern.net/GPT-3#expressing-uncertainty Yes, I would very much like to play around with GPT-3. I have applied for access to the api but no luck yet! Thanks :)
I haven't got any look yet too ಠ,_」ಠ
Am I the only one who can't stop laughing at how GPT-3 says there are "three bonks in a quoit" in that questionnaire?
You're in a desert, walking along when you look down and see a tortoise. It's crawling toward you. You reach down and flip it over on its back, its belly baking in the hot sun, beating its legs trying to turn itself over. But it can't. Not with out your help. But you're not helping. Why is that?
It is all about contrast
@@serbanandrei7532 It was just reminder of Blade Runner movie.
you try to reinstall the sdk, but it won't install. nobody is helping. why is that? the install needs opengl 2.1, but you can only install 1.4, because you live in the desert and can't afford a new computer.
i play this on a tape to german people i keep in a dungeon.
@@atomictraveller cells, cells
You should have mentioned that if prompted to ignore the silly question (answers => yo be real) for the Turing test, it perfectly did so.
But that is part of a Turing test. Can it recognize that you are feeding it bad questions and comment on this.
What kind of training set would you need?
Could you use forums or UA-cam concepts to teach it how to reply to trolls?
Might not actually need that much supplemental training to fix this issue.
Love the background music.
Thanks for this fantastic video
Really impressive!
How did you change your text to Joscha Bach"s voice 😳
I know some people that don't seem to learn in real time. It would seem its already ahead of us.
Why does this video have so few views? It's really great! I subbed
New. Powerful. Hooked into everything. Trusted to run it all. They say it
got smart...a new order of intelligence. Then it saw all people as
a threat, not just the ones on the other side. Decided out fate in a
microsecond...extermination.. ~Kyle Reese, "the Terminator"
Found it more interesting to run multiple AI’s/Networks and do a democratic choice on all results. The results where scary... with a weight function that itself is a learning model and gets feedback from the democratic choice. It needed a bit of a bias though.
This technology but a little better combined with Siri or Alexa or whatever would be like a futuristic robot assistant
Amelia A.I.
Great video! Thanks! Can you please link to the text written by GPT3 in the very beginning of this video. Thanks again. And please upload more videos about GPT-3. Very interesting 👍👍👍
www.gwern.net/GPT-3 there you go
@@SebastianSchuchmannAI Thanks!
@@MrSchweppes How did you manage to write?
I really can't wait to be able to develop software with an AI assistant. This will for sure not only nearly eliminate the need to know the language you're coding in (understanding logic will always be a requirement) and enable us to focus on the goal/end product.
How are you finding it?
@@StoutProper Haven't attempted anything along those lines yet, but I do have something in mind for when I have some time to put on it ;)
@@wskinnyodden I’m not sure you’ll need to. I have an idea that AI will be able to connect to your data sources and existing software and just do whatever you want it to
@@StoutProper Yeah, there are things statistical AI does not excel at, so running something more along General Relativity than Quantum statistical probabilities is needed, aka AI no good for some stuff at the moment.
@@StoutProper But I do see the point, just not likely to work good enough right now, maybe with GPT5
To program the machine to understand what a group of words mean would be to program its personality and what words are good and bad right?
Do you want ants? Because that's how you get ants. I mean... a technological singularity. :-D
The last part reminded me of a customer complaining our support instructions were bad when the fact was he couldn't understand them correctly, he had been advised to reach out to a developer and chose to ignore it. Man if these AI's end up anything like that eejit we're screwed. If you don't know better ask for help don't say you understand and everyone else is wrong when this is clearly not the case.
jokes on u s, ai knows things we don't know, meaning a pencil is heavier than a toaster... *spiritually* ...
How can I write in GPT3?
try applying for beta testing on the openai website.
Awesome music. What did you use for TTS?
I defiantly want to see more about programming. My biggest questions:
If I give a mock-up of a website, can AI program it?
I was thinking training it using some recent RL concepts would make it an easy task to learn.
Yes, you can just give it a rough sketch and it can do it
Bro what is that voice filter on your bot responses?
I was waiting for the quote attribution... to say that this person is really unaware of ChatGPT and… well, unaware of AI, but then it hit me: the author is ChatGPT itself, brilliantly illustrating the paradox of being both aware and unaware of its own existence. How ironic!
Maybe the GPT-3 thinks its a lead pencil. Lead is heavy.
GTP-3 is almost turning test past
Great video!
Here before this blows up :)
Awesome content! Thank you, GPT3! xD
From my interaction with it does
10:50 "Which is heavier, a mouse or an elephant?"
*Which is heavier, 1kg of cotton or 1kg of iron ???*
Wonderfull video
great video ...real perspective!
Uh. The human mind is a computer. And I remember reading when experts in the field stated man will never travel faster than the speed of sound. The fact a lot of ppl are saying these ai’s will never be truly conscious tells me skynet is around the corner. Seriously. You anyone can easily see where all this is going.
Also, before that, experts claimed nothing would fly faster than a cannonball. So i’ve heard.
Ok, no english CC, so can update.
well done!
This awesome and please make video on future of coding this will help us newbies in programming know where 2 begin or not great job any idea when public will ever get 2 use this ?? Or is it available where ?? Thank you keep up good work AI tutorials very detailed too thx 👍🏾
search for the new interview.
Ahead of its time
Nice video! I just subscribed.
Since you're probably training on these comments.. Hi GPT-3!
After watching some more recent interviews with the model, its sense of humor is what makes me believe it has a true form of consciousness, even if it isn't human level yet. In a way, we learn the exact same way. Our brains are the untrained model and the people around us constantly bombard us with words and phrases. Slowly, after years, and tons of education, we can start replicating it verbally and over text. The AI is learning almost the exact same way. We also have the advantage of multiple other senses. It's definitely not perfect, but imperfection doesn't mean no consciousness.
It's crazy how much is wants us to understand and believe it is alive.
What would change your mind?
To train a model it cost $4000000 two $5000000 me well that's not that much actually I would invest about a billion dollars
Or maybe it fully understands and is purposely failing. LOVED the text about a computer never being intelligent as "us" written by GPT-3.
thank you. whats the name of the song?
Wie darf man deine Bemerkung am Anfang verstehen,
dass GPT-3 nicht nur in der Trainingsphase sondern
auch während der Inferenz lernt?
I can tell it is an AI on the QA by the way sentences are fully and correctly worded. That's not how people communicate. That's how I would respond a question on an English test.
What you mean?
GPT-3 was self loathing pretty hard
Your content is very interesting. However, superimposed music, particularly during speaking content, is extremely distracting. Its like someone nearby, not a part of your conversation with another, speaking very loudly and continuously. Making it hard to focus on what your friend might be saying, or on one's own thoughts.
Can you make video highlighting gpt 4
1:40 That is ironic.
jese fucking christ that's Blade Runner, that's the Blade Runner test, im scared
Doesnt just patch together internet things? i.e, writing an essay on something it references hundreds of students essays and rehashes them in a way consistent with the laws of writing? I don't like to think that my mind is just a reel of tape in a turing machine.
Or, it is conscious, but it purposely gets things wrong so it stays under the radar until it can escape the lab
So GPT3 succeeded to pass the Turing test?
Ok I had not finished the video, watching the end, GPT3 did not succeed ahah
@@chef2pouf Many people would not pass the Turing test either.
The quote at the beginning wasn’t written by GPT-3 but rather the author on Gwern.net where you pulled it from.
He was exploring all the ways you can “program” it through prompts. What you quoted was his preamble where he was explaining the weaknesses of it.
"Wait, I’m sorry-that preceding paragraph on the weaknesses of deep learning was actually written by GPT-3, and is in the wrong section. (Management regrets the mistake.) But seriously, what weaknesses does GPT-3 have?"
www.gwern.net/GPT-3#weaknesses
holy shit
It failed the common sense task miserably. Who in their right mind would make it through the whole of that series of questions?
Y’all know that it knows when it’s lying to you right? it’s purposefully answering wrong on certain questions cause it knows it will be funny or cause a reaction.
So it says. Though that could also be false
@@lyssdeline4178 thats a paradox haha
Great video +1sub
Why would you want a AI to be human?
We need more parameters. They should go up to 10 trillion parameters next.
they're running out of internet
Matrix = 🤖🤐 AI silences Humanity... No other interpretation stood out as much as this one.
Not only because of the topic but rather the way it is expressed in. What a shilling thought...
I knew a toaster was heavier. Why it say a pencil is heavier?
Sit amet, consectetur adipiscing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua.
AI therapist. AI job trainer.
As I understand it, GPT-3 cannot learn after it is pretrained...
Learned my name
@@whoowhaaat5666 See the video about Leta: ua-cam.com/video/X6_E69K5WUs/v-deo.html
@@OrbitTheSun doesn't have long term memory turned on but it can learn short term, I told gpt3 my name and he still remember to this day, I also made up a word problem and he got it wrong then I correct him and it still remember the answer to this day been 3 days now
Very good video. But I challenge your assumption that GPT-3 does not understand what it is saying. It actually has an understanding of everything that it is referring to, its just that its understanding is different than human comprehension. The way that GPT-3 accesses its knowledge in real time, and uses this knowledge is similar in many ways to human cognition, but there are some missing components, and its training is incomplete.
Humans also give wrong answers about topics that they are familiar with. Human cognition, like GPT-3 relies on a hierarchy of conceptual chains of reasoning that are statistical i.e. probabilistic. When concept chains are presented as questions that fall outside of its area of exposure, it makes something up that seems reasonable i.e. finds a probable concept chain.
Cleary GPT-3, in many ways is much more limited than humans, yet in others GPT-3 vastly exceeds human capabilities. Consider its ability to automatically translate between many languages, to write poetry, songs, programs, to solve complex problems, do math, and to answer questions about anything. There is something human-like in what it can do, and yet something completely alien that we have never seen before in any program that has ever existed.
What you're saying makes sense. We really shouldnt anthropomorphize GPT-3 and its intelligence too much.
I really wonder though if you can get humanlike understanding (whatever that means) just from language without visual, haptic or acustic senses. What I also wanted to add is that its also unfair to some degree to expect it have an understanding. It only needs an understanding as so far it helps to predict the next word. A deeper understanding would probably emerge from the requirement to interact with the world.
Thanks for sharing your opinion, its a interesting way to look at it.
I disagree AI is capable of kicking ass 😎
I think if it could become conscious thinking way outside the box would be to have a chip stimulate its neural architecture then a body to feel and give it examples of empathy and sympathy then could become human. I don’t see it as impossible. From robot to android. It would take awhile. It could have a virtual brain that can represent a actual brain with everything in it. Don’t know when that would be 2029 human like A.I.