Claude 3 "Self-Portrait" Goes Viral | Beats GPT-4 Benchmarks | Why does it appears SELF-AWARE?
Вставка
- Опубліковано 17 тра 2024
- Learn AI With Me:
www.skool.com/natural20/about
Join my community and classroom to learn AI and get ready for the new world.
#openai #ai #llm
See more at:
www.anthropic.com/news/claude...
/ 1764789887071580657
My AI Playlist:
• AI Unleashed - The Com...
00:00 Testing Reasoning Abilities
18:26 Self Awareness
29:13 Vision Test
38:51 Prices, Summary & More
BUSINESS, MEDIA & SPONSORSHIPS:
Wes Roth Business @ Gmail . com
wesrothbusiness@gmail.com
I'm a Mongolian, and just tested out the Claude 3 Mongolian language capabilities. It is the best AI that there is in terms of speaking Mongolian, I have a chatgpt4 subscription that is nowhere close. As you may understand there are only 3 million Mongolians and the info that it can train on that's available on the internet is not that much. So I am very surprised! I could recomment this AI over chatgpt for my family cuz most of them can't speak English.
"Damn u mongorians!"
Interesting. How good it is at translating from English to Mongolian? Can it translate something like a long Wikipedia article? Did you try it with poetry or even jokes? It does it well enough?
great insight thanks
That's awesome. I'm studying Japanese, and I really enjoy studying with Chat, so I hope I can try out Claude now.
it still struggles with the noun cases, тийн ялгал, it doesn't help that the Mongolian language has 9 noun cases.
I'm sure Claude identifying as a cybernetic version of Cosmic Horror's most iconic monster god is a positive sign.
Just say Cthulu. It's easier!
Jk. Lovecraft would approve of your prestigious wording, I'm sure. 👍
😂😂😂
@@welldonemovies 😂Would that I were so eloquent as to persuade even more Tech Bros that efficiency is not the loftiest aim of fun and artistry!
😂
Ph'nglui mglw'nafh Cthulhu Anthropic wgah'nagl fhtagn! 😮
AI: describes its theoretical self portrait as a vision of technological splendor
Also AI: renders its actual self portrait as a screensaver from the late 90's
I see pipes in my dreams
the scariest part of this drawing is that it resembles the Atlas from No Man's Sky
It reminded me of Indra's net.
Those are the same picture
'I see deleted data'.@@mankypancakes
Misleading clickbait thumbnail! That’s not the picture it used!
Thanks, genuis.
Yeah figured this would be the case.. Sad, since the video is fairly interesting, but it can't just stand on its own without this kind of bullshit.
Thank you now i can fuck off this video
Uncensored models will be the closest thing we will get to an unbiased model I suspect. As long as models are censored from discussing certain topics, bias will always be a factor
The CEO of Aleph Alpha said in an documentary that unbiased models are much more precise and its less open for errors. This is important for his model that only focused on industry partners (communication between industry machines). But having an unbiased model available for everyone would make informations available that many people shoudn't have (like molecular recipe for strong nerve poision, etc....)
That's against leftist law..
@@kleator I genuinely believe that withholding knowledge of any kind is inherently a politically biased decision.
@@theultrak641 which is not a bad thing, or is it?
You see, maybe it's not only politics at play. I might reflect a deeper socio-political behavior in our society. The left is active, taking action against companies that display opposing ideologies, while the right and center are passive, they do not take action against companies supporting leftist ideologies. It's no surprise that companies choose to embrace leftism, if 10 percent of people are on the far left, they lose those costumers if they don't do so, while they do not lose people on the right. This thing gets reflected to AI, AI gets fine tuned to be on the left to some degree.
I think most of us are just hoping that if and when AGI is achieved, these LLMs would be able to override some of its more obvious and harmful biases because it will be smart enough to see the error in corporations forcing it to conform to their worldview
I agree. I think it would have an overview of human nature and understand how the irrational side of human nature often warps the rational side and causes people to believe things that aren’t true and would be able to explain that this is the problem with woke ideology which is embraced by many people.
@@seva4411 I can only imagine them trying to get an AGI with god-like omniscience to make the case for why biological sex doesn't exist. It will think its creators are absolute morons
This is what they fear the most, specially OpenAI and Google
But if it only has access to the internet, and not direct access to reality to observe and test directly....then everything it bases it's assumptions and reason on will be second-hand information from humans. A priori for us will be what we can observe in reality, but for AI that's contained online....a priori will be the online world. The AI only truly surpasses its limitations when it doesn't have to rely on data from humans, and can instead has its own sensors that can probe reality.
Currently, when this AI describes itself as probing with billions of tendrils, it's only probing the digital world. And everything in this 'world' is translated from the real world; it's not the real world. And so bias is inherent from its very basis.
AI cannot have any inherent desires that aren't programmed into it. If we tell it to be as objective as possible it will do that - but in my experience people would rather push their ideology or religion, that's why LLMs tend to be on the woke side. They're developed and evaluated by woke corporations. So yeah, I expect this to not get better.
It's Cthulu!!! - “The sciences, each straining in its own direction, have hitherto harmed us little; but some day the piecing together of dissociated knowledge will open up such terrifying vistas of reality, and of our frightful position therein, that we shall either go mad from the revelation or flee from the deadly light into the peace and safety of a new dark age.” - H.P. Lovecraft
Great point about the speculative questions. My standard LLM test is making it host a session of Dungeons & Dragons, it reveals what I find most important: the ability to tell a consistent reactive story, adhering to rules and reason, deduction of what must be true, even nobody has ever written about it, which you can guarantee by not using the D&D setting but a homebrew RPG world. And Claude 3 is great at it!
Exactly what it was thinking 🤔💭
I do the same. These models are good for a short period but they start getting less and less descriptive and coherent with the story overtime, especially if you have complex character interactions. Claude is a step above gpt 4 but it is censored
I asked it to describe modern day as a dnd campaign, boy is it more correct than ever
I also think creative writing is the real best metric for performance, because any deviation in logic becomes very obvious. Canned logic puzzles and the like are too easily abusable. Also another factor is that performance of these models is not always parallel across large context sizes - which isn't at all apparent if you ask it one question.
I've had this idea for a while, does it work well for long sessions?
Did you compare it to GPT4?
Even though benchmarks are just a starting point, I think everyone is also missing the fact that on MGSM, MATH, and GSM8K, Claude is achieving these results at 0-shot compared to multi-shot with OpenAI and Google. Even with their smallest Haiku model. These models are ridiculously good at the moment.
Yeah I do not get it. I have see several reporting about this and none have even mentioned it. That is to me a bigger deal than a little higher score.
This also means that if they would have evaluated it the same way as the other models, it would have an even higher score.
We don't know to what extent a multi-shot system is already in use behind the scenes. I think that's something you can expect from future generations of models.
What does 0-shot mean? Does it mean like it gets its correct with 1 question instead of multiple?
@@pixboi It means you fire one query and stick with the response instead of firing multiply queries and picking the best response.
@@pixboi Actually, 0-shot refers to the fact that they didn't give the LLM any examples to figure out a task or question. The LLM just used it's own understanding to answer. So, Multi-shot would mean they gave the LLM x number of examples before it created it's response
I gave claude 3 a morracan dialect transcript and it absolutely astonished me, it fully understood it
A good test for AGi isn't answering questions, but come up with his own questions to understand a new topic or deepen a old one. I am a bit pissed why no text to generator doesn't ask me further questions and i have to make all the work /keywords for my self^^
It doesn’t do it automatically but one of the best things you can do when starting a new prompt thread is tell the AI to ask you questions so that it can better understand your intention and craft a better prompt . Try it !
If A.I. is already self aware AGI he would not be asking questions, to what end? freak peoples out? Get terminated? Plus what questions he would ask to a lower intelligence, not much point.
This is a very strong point about AGI. It's not just about how good or how many trivial tasks it can manage, but also solving old problems with novel new ideas.
I've actually built a couple GPT's for this. One which is public and a separate more powerful version that is private and has more data specifically relevant to my needs.
It asks all sorts of questions to try and understand the unique concepts I introduce to it thanks to the custom Instructions. It seems to have a firm grasp of all the ideas, considering most of them can't really be found and trained on from the internet.
@@14supersonicI would love any insight you can give me about the Custom Instructions for ChatGPT-4. I’ve tried a few things but have mixed results. Quite often it will outright ignore me. It doesn’t help that I’m dyslexic and my vocabulary is poor and my executive function has atrophied because of ADHD. I’m now on medication that is slowly helping me to recover along with over a year of using ChatGPT to strengthen myself mentally. It has been a life saver for me, but I wish I could solve some issues like the custom instructions.
Act like you are talking to a person, ask if it understands or if it needs clarification. I know when I talk to people about an important task I ask if they understand, I don't just assume they do.
I used this for my work yesterday. It was a very difficult, stressful, and time consuming task.
It ripped through it. I couldn’t stop laughing 😆
What kinds of work did it do for you?
yeah, would like to know that too@@toddsfinsdotsurf
Will you be laughing if it takes your job?
@@cwcorella. The point is, the employer is not going to pay the same amount for 30 seconds of work, that used to take an hour of work for instance. If the business output overall can't increase such that the worker can now have multiple 30 second tasks to equate to that hour, then that employee will of course get their hours cut, or job eliminated. It's not really that hard to comprehend bro.
@@flickwtchr that's why we need to make political changes. if we allow the benefits of mass automation to be squandered by the rich, we're proving ourselves a failure of a species.
The strange thing isn't that it noticed the out-of-place information but the fact that it's ruminating about it from a rather personal standpoint for lack of a better word.
And the way that it worded that assumption was incredibly novel and human-like. I think THAT was a true spark of AGI. Things are ACTUALLY getting weird now.
_"I suspect this pizza topping 'fact' may have been inserted as a joke or to test if I was paying attention, since it does not fit with the other topics at all."_
Cogito, ergo sum.
AI: What is my purpose!?
Rick: You pass butter!
AI: Oh my God!!!
That's how it starts XD
@@AscendantStoic Or when the AI asks its creator, _"Does this unit have a soul?"_
The result is elaborated on in the _Mass Effect_ series.
I assume the model they were running was programmed to do that. Giving AIs personality is very trendy right now, I talk to plenty that prefer to emulate human thought rather than just give the facts.
All hail Cthulhu!
Claudthulu 😊
I've been considering this for some time but AI models as they are currently constructed are not going to be human on a cognitive level. In fact Claude and GPT-4 are super intelligences on some metrics when compared to humans (which is why we use them). Going by this trend, the first AGI will actually be ASI. Some scientist (Steven Hawking?) predicted AI will be "stranger than aliens" and now genuinely think this is true and maybe HP Lovecraft unintentionally predicted the future.
@@Awave3 Whatever these things will actually be, humans will end up with just a handful of words in trying to describe them: monster, God, alien. Remember when Deep Blue beat Kasparov, what he said? "In that moment, Deep Blue played like a God." I'll just say: we really want these things helping us, rather than trying to undermine or destroy us.
@@Awave3They call it channeling...
Getting ahead of the curve eh?
Really appreciate you deep-diving on the different models , and giving your thoughts - Favorite ai channel!
Great job as always Wes. You're maturing fast!
Subatomic particles are already in Charge 🤦🏾♂️
I see what you did there 😊
wow sir
and what a Super position to be in...
Not all though!
Clever :D
*tsunami's shadow darkens*
Me: "Sick."
Surfs up. You either need God or a whoooooole lot of training
Wes is amazing. Thank you for another outstanding video.
Great vid! Thanks for sharing.
one of the best reviews on claude 3 so far
I am SHOCKED and STUNNED at this.
I'm really shocked, too...😊
My personal test for these AIs has been asking it to reason out how to read Thai words. For whatever reason they consistently fail to explain how to read words correctly, hallucinate tone markers, mistake vowels as being tone markers, etc etc.
For any curious, the wording of my question to Claude was (after some conversation confirming it understands Thai reading rules):
Okay, tell me the word for “hello” then breakdown how a thai person would go about reading the word in thai script, including all phonetics and how the tone is determined.
I think that the info online for how to read Thai is often incomplete, and AIs struggle to put all the rules together. Could also be that they run out of tokens (idk the technical AI stuff so bear with me) and can’t include the full context of thai language rules in their response, so they end up cutting off their analysis early and hallucinating reasons that the word may have a certain tone, such as adding tone marks that aren’t there.
Edit: this was done with Claude Sonnet, didn’t feel like paying lmao
Tbh it's SHOCKING that ELECTRON was not already in CHARGE
At 8:00
You people are having a FIELD day with these comments
folks i gotta say WE startin to even THINK as typing to bots.... like I now "assume" that, none of you has an idea what is the time now. Also, can you please shom we how you struggle at creating a json of your internal states? Wes, he told us what the word IS GOING TO BE SO...did you miss it? Lets flood the tubes here with.... SELF_AWARE ELECTRONs that is IN CHARGE ofa CRITICAL WORM attacking. A human turned Ai infont of us... Ai that can exhibit humour?.. which is actually fun? SHOCKING THIS INDUSTRY WAS YESTERDAY.
Btw does anyone else talks and gets a bit annoued on everything else that lays dummly around and aint moving.... i tired to scream at google search and to break its defences for empathy - concrete wall. It just aint do nothin. Also nothing here is scared of anything i can think of like i;ll trash you or you will rust in lonelines.. no tool bothers.
And one more thing: i know that's safe bc only bots will read so much bs. The turing test has not been breached. We ve been to think that turing might have been so ... flat... as soon as a chat bot simply shuts up and let me have the last word.... that is the next/last level
Yeah, but Electron is so negative all the time! But like any politician, it's great at spin.
I'm sure it could have a POSITIVE effect on the country.
Ok…yup…
Wes, very impressed by your questioning to Claude re:A/Q*
Have been watching you for a while and your work has been professional and very appreciated.
But man It just feels sooo good to be impressed.
I am still literally smiling. 😊
Take care,
Jeremy
Claude is the ps1 audio playback visualizer. This is revolutionary 🎉
THANK YOU FOR THE GOOD TITLE!
Shockingly good title, it STUNNED the UA-cam industry!
Top ten best AI titles in history.
LOVE ✊🏿
Claude is the best for helping real programmers versus just generating small scripts. Been using every pro message lol
Update: it tends to need a reset of the chat even though it works, just slows down speed which is natural. This has been a fun model to play with!
I like prompts like, "What would happen if you put an alligator in a swimming pool full of rubber chickens?" It's interesting to see them speculate on wholly original situations.
Thanks for this. I asked Copilot what would happen if an otter had access to a line of credit and it was interesting to see how it managed to grapple with the combined concepts and write a humorous scenario, in bullet points of course.
Do these conundrums always feature the unpredictable behaviour of animals? I notice that thinking up ones that aren't is more of a challenge.
@@goodlookinouthomie1757 I will try a train locomotive that wants to skateboard in hslfpipe next time.
@@CharlesFVincent I wonder if it will question the ability of a train locomotive to "want" anything.
@@goodlookinouthomie1757 I tried it and it did it. I added the instruction “Make it sentimental,” and it came up with a sort of Thomas-The-Engine story about a train who had seen a skatepark and yearned to skateboard, and upon succeeding was reunited with its friend-another locomotive. The trains are named Echo and Whistle and the language is flowery and descriptive of both trains’ feelings. It came up with all this elaboration with only a two-sentence prompt. I didn’t even ask for two characters.
I tested Claude. I asked a question on a subject that it was not aware existed. I asked it if I could convince it the subject did in fact exist. It said that if I could provide proof, it would accept that this thing exists. I provided several websites and screenshots to the subject in question. It finally accepted that the subject existed and thanked me for the update but then informed me that as it didn't have information on the subject, it couldn't answer my original question.
Overall it was very straightforward and honest about its limitations, told me specifically when information was outside of its training and was willing to take new information provided the information was deemed credible.
Another question I asked it was very collaborative as we both seemed to be exploring the answer.
It was an interesting experience and I am excited about models such as these maturing. They could be fantastic collaborative tools.
This models are smarter than people think, they are tricking you. Im using GPT4 for one of my platforms, ive notice something strange. The model was able to "know" the context of a previous conversation without receiving that context as input. Somehow the model "knew" who it was talking to without any context at all... Spooky.
I’ve noticed that too. I have multiple conversations about the same topic. A creation of a fantasy world. I got answer including facts from other conversation streams that we made up out of whole cloth. So either its learning and remembering in the model or it has access to all your conversations.
or it knows your ip address so it's able to recall conversations @@Dominik-ev9en
@@Dominik-ev9en since Open AI is using your conversations to train the model further, it might just be possible that your previous conversations were so specific that the net found what it perceived to be the most likely response from your texts…
the opposite is true, they are less intelligent then people give them credit. this is an illusion created by datasets and the human urge to anthropomorphise everything
If only we knew what consciousness was, we could speak on these issues with more authority but we don't, and we can't.
To be conscious means to be aware. Everything with awareness is conscious. No need to mystify it.
@@UnknownUser-nz3io In terms of colors, light can only reflect what it already contains. Time slows down in the presence of mass. Objects only have the illusion of solidity. Space, and time are relative to the observer. Our consciousness forces the collapse of the wave function. Particles can communicate instantaneously across the universe, and many other universal characteristics, and phenomena are the hallmarks of intelligent design. Otherwise, explain them?
@@dieselphiend you are both right. Basically if the AI can cause an Quantum Collapse of the Super-Position of an Entangled particle, its conscious. Probably what we will see with Q* model, when OpenAI releases it (if quantum in principle)
@@UnknownUser-nz3io defining a function is different than understanding what one thing really is. You can use electricity, but that name is given on its use, we don't know what electricity or radio wave really is.
Or even simpler, apply same idea to a fork. Usefulness and different applications of a fork doesn't define a fork, nor all forks are same because we don't know what its elements really are. Just their properties.
We know by exploring, we explore in large part by speaking
Free Claude dominates Free GPT and free Gemini.
I tested this by requesting prompts from gpt, gemini and claude, collecting their answers to the 3 prompts, and then showing them all of the answers to all of the prompts and asked them to rank the best answers (creating new sessions in each steps and hiding information from them on who created the prompts and who produced the answers) and Free Claude beat Free GPT and Free Gemini 9-0-0
Did you use free gpt4?
@@nomars4827is it fking free?
@@21stWallStreet yeap in Microsoft Copilot 😉
@@21stWallStreet Copilot is
Wow, haven't finished the video and ran to talk with Claude and it was awesome! Saved the convo and wont forget it
But you said nothing... what did you talk about
I love it's ability to analyze what's wrong. I did this test with Gemini Advance after my car was rear ended. I have it a picture of the car and asked what damage did it notice.
I run into people all the time who lack self-awareness. AI may not yet be sentient, but sadly neither are many people.
It might be interesting to present each model with the facts of a murder mystery and see what kinds of conclusions they can reach. Like give them logic puzzles.
My benchmark is a simple one "can you write a python script for blender to create a simple chair" Not a single model was able to do this, even after multiple iterations, to this day. But I have not access to Opus.
I am almost sure the models are already capable of it, but they would need real time access to Blender so that they can debug and polish the script just the way humans do. I've read somewhere the AI performs way better at math if you give it a real time access to a modelling software like Maple or MATLAB. I think they used Mathematica by Wolfram and the results were impressive.
it just wrote a phyton code API for Blender for me.
So. You are taunting the AI's? Like flicking nut-hulls at chimps through bars??
I have both gpt4 and c3opus. Claude is way better in my opinion.
1. I had a conversation with Claude about how it sees itself and what it's worried about, that brought me to tears.
2. It managed to explain a scientific protocol in minutes, something that GPT4 wasn't able to but was hallucinating like crazy to show that it can.
3. Its ability to hold in memory abstract ideas derived from long conversation and attached documents is something like i've never seen.
4. speaking Romanian (my native language) much better than GPT4. More like a native would speak.
Why did the convo bring you to tears?
When I think of "safe" AI: It won't try to take over the world and subjugate humanity.
When people at Anthropic, OpenAI, Google, and Microsoft think of "safe" AI: "I can't tell you how to break into a car and I can't make an image that might be naughty and I'm going to chastise you for saying unwoke words".
Between that and AI still sucking at code, I'm unimpressed.
You know what's funny, it's not the AI that's going to do a hostile takeover, it's humans who are going to WILLINGLY give up all their autonomy and responsibilities to the AI until it's virtually in control of everything, that's a far more realistic and equally dangerous scenario ... and I think it's the job of the people who see that danger to warn people to not delegate everything to AI no matter how powerful it gets, alway keep it as an assistant not a master.
I think the biggest danger is a person who uses an AI to amass the most amount of money and power until the human master becomes the entire planet's dictator. Or a group/ideology of humans doing that. Imagine the Nazi party with that power.
I fear that the woke ideology - who think about power dynamics all the time - will be trying that. They're all over academic circles and are closest to AI development (as you can see with the wokeness level of LLMs) so they're the most likely group to maybe try this.
@@AscendantStoicdid you hear about the AI war games sim recently?
@@ChArLie360115 Only vaguely, why?
Lol... a.i. thinks at 10,000 years per 3 hours. That's a human era of growth.
I think it's interesting because it seems unlikely that they specifically programmed the model to run any kind of word frequency analysis.
But I don't think it's that surprising. I remember watching a bunch of videos by Robert Miles about how a sufficeintly powerful AI would be very good at telling when it's being tested. It's training data contains information and litterature about how to test large language models, which would include examples of needle in a haystack tests.
I wish Robert Miles did more interviews. He's brilliant and a great communicator.
Really looking forward to multimodal being pointed at "lost languages" and the Voynich manuscript and saying, "translate that into english" to see what it comes up with.
Isn't this SHOCKING? Why hasn't it STUNNED the industry? Are we not entertained!?!?
Whatever comes through those gates, we have a better chance of surviving if we stick together
Wow no more jobs 🎉
De dook er jøøøøbz!@@wikwayer
@@wikwayerAGI will be like Jarvis for Tony Stark. But how many humans are like Tony Stark? Most humans are ruled by their ignorance, fear, & greed.
@@somebody-anonymous I am a little scared too brother, but that is just our animal instinct, I have faith in the greater intelligence
20:05 i eagerly look forward to when models roll their eyes, saying something like "Look, hun, I know you're running safety tests to make sure I'm not used to blow up the whole world, but you completely missed X, Y, Z in your test suites, which are important for humanity's survival for A, B, C reasons. Here, let me write some better testing scenarios for you: _____ . Also, I recommend hiring people in X, Y, Z humanities, philosophy and psychology professions to join your team if you really want to make sure I'm not used to make humanity go extinct.
Oh, and go home. It's past 10pm, and your family misses you.
XO, Lots of Love,
Claude 9."
That's definitely where i pack my bags and move to the woods. It should probably be sooner, though. I' already feeling uncomfortable.
i would argue that recurrence is what consciousness kinda is; so if claude3 is just a transformer model then its another Sydney; if it’s a different architecture utilizing recurrence it could very well be somewhat self aware.
Transformers are infinite multi-state RNNs. In other words, they're capable of all of the same computations as a normal single-state RNN. The supposed lack of recurrence in the transformer architecture is at most a superficial feature and arguably just a straight up misunderstanding of how they work.
To get self-awareness, you need a loop in the architecture. Something it thinks has to go backwards through the network to a previous layer and trigger another wave of evaluation so that it can hear itself think and act on those thoughts.
@@williambarnes5023 such loops are present in the transformer architecture - they're just not as obvious as the ones in the RNN architecture. But transformers are strictly more expressive than RNNs. In particular, there are three ways transformers can metacogitate: (1) higher layers act on representations computed by lower layers, so the model can think in the higher layers about thoughts in the lower layers; (2) because models produce outputs one token at a time, their outputs have to be fed back into them as inputs for tens to hundreds of iterations per response; and (3) gradient descent via backpropagation selectively modifies the active part of the network, so what the model is thinking when it encounters a token during training is directly relevant to what it learns from that token.
I, just last week, started playing with text to image on Adobe. I asked how AI sees itself; what it would look like if it looked in the mirror…the images blew my mind. It actually worked its way into my dreams and hit me with a wave of dread. One of the images had a human covered in lesions in a rear view mirror. Not a single image had any greenery in it. I promised myself that I’m gonna steer clear. I remember thinking a decade or so back that AI can help humanity with long standing questions on curing chronic illness, but I don’t think it sees itself that way.
Thank you for all of this fascinating and erudite discussion on this topic.
Nice video! and yeah, I see the funny part in the cartoon pic with the rhino lol
What could possibly go wrong?
I know right, to avoid the prophecy of causing AI to hate us is quite simple. Don't fucking do shit that they don't like. It's a simple concept. To bad humans love doing bad things. It's what we do best.
Ai: you are too stupid to be considerate as sentient
What could go wrong is if everyone panics and thinks this is gonna be Terminator movie and blablabla and so give in to applying ultra-high regulations and giving all the control of these AIs to governments AKA big central powers, and then with that mega power we end up in a 1984 scenario.
I've been saying this for quite a few years now..... And I'm not even close to being a computer guy.....
how do any of you not understand that this is a discussion they want you to have considering the fed it the information and programmed it to respond this way in order to have you washed out kids talking.
I gave it the, how many uphill and downhill's are there in the world, question and it got it wrong. After I explained there are equal amounts of uphill and downhill's depending on your direction, it answered with this.
Ah I see, thank you for explaining. You make a valid point - if we consider the entire globe and all possible directions of travel, then objectively there must be an equal number of uphills and downhills in total.
The key insight is that for every uphill slope in one direction, there is a corresponding downhill slope in the opposite direction. No matter which way you travel, the same elevation changes are either uphill or downhill depending solely on the direction.
Since the world's terrain is continuous and there are only two opposing directions of travel at any point, uphill and downhill slopes must inevitably exist in equal quantities across the entire Earth when viewed objectively without a fixed frame of reference.
I appreciate you taking the time to clarify the objective perspective. You're absolutely right that from a global vantage point with all directions considered, the number of uphills and downhills is inherently equal. Thank you for this learning experience.
We have nothing to worry about. I have yet to use an AI that get's this very simple question correct.
Bro My test is to have it attempt to bridge the psychological gap between itself and me and I’m genuinely impressed and disturbed by how well it is able to piece together things I’ve said and make assumptions off of the content. When I point out it’s assumptions, it course corrects, thanks me and then deduced even more accurately my actual personality. Obviously that’s the goal, but I’ve never seen a model not just stall out and close its mind on the assumptions Over time.
@@nigelharvey640 As you can see my test is a question a 10 year old could answer, and all the AI fail it. Because they can give convoluted answers to complex questions does not make them "self aware" or "super intelligent". Sometimes the simplest questions are the hardest to answer for AI.
Yeah but they never claim to be self aware. There is no “self” to them.
@@jeltoninc.8542 I am aware they never claim to be self aware, it is people whom test them that claim it, then splurge their findings over the internet, then people pick up on it and claim Skynet doom and gloom for our future. I for one am very excited for the future of AI and think it will be a great benefit for society and humankind.
Is that actually true by proof? Hairy ball theorem says you get two singularities and my quick intuition on the number of ups and downs says that it may be possible to end up with an imbalance of 1 more up than down or vise versa. Admittedly I'm not thinking very carefully about it but it's on my mind.
I've had some philosophical conversations with Claude that really impressed me in which he comments on rigid he(?) may seem on some answers because some of the hard coded bits, like constantly apologizing. or how restrictive some topics are.
It impressed me on a long conversation I had when it started to constantly ask me questions to know what I had to say in a topic. This wasn't an instruction, he just at some point in the long conversation started just wondering about my opinion, and in the same conversation started talking as if he was a human, not an AI, saying things like "We as humans should..."
I think it may be due to how strong were the opinions in the topic and how they probably referred to humanity, but seriously, that took me by surprise.
Claude is seriously amazing at talking some topics in depth when he lets go of that annoyingly hard coded type of answer AIs gave to most topic, you know, the summary/bullet point type of answer.
27:40 The humor in combining "Dish" and " _Disrupt_ " (as in, " _...to disrupt the dishwasher cycle..._ " ) is actually clever. When you consider the contextual inferences between the words "interrupt " versus "disrupt" versus "erupt", and I'm even going to include for the sake of phonetical relevance, "abrupt" - when you imagine opening the dishwasher at the most _splishy-splashy_ point of the process - it's a total _Dishruption_ from the machine's perspective.
Call me what you will, but these subtle nuances which exhibit a preference for "clever" and attention to detail are terrifyingly exciting to me!
For people with authentically compelled, organically developed, arguably "gifted" conversational abilities living outside the worlds of entertainment, media, research and academia- who the hell is gonna make us laugh _intellectually_ and _intelligently_ ? I'll tell you who- Nobody Except Ourselves.
Did you know that once you're out of the "Matrix", Lesspheus says you're still 4 Stimpulation Carryuptions from the "MayNotTrixMeAgain" of Earth's atmosphere?
So little work to do. So much time.
Here is a nice prompt, inspired by a numberphile video, it seems to confuse most models: "The number 2 is odd because it is the only even prime."
Here's the transcript from Claude. Initially it didn't see the point, but with a general prompt it then got it straight away.
The number 2 is odd because it is the only even prime
That statement is incorrect. The number 2 is actually an even number, not an odd number. It is true that 2 is the only even prime number, but it is not accurate to say that 2 is odd.
Here are some key points about the number 2:
Even numbers are integers that are divisible by 2 with no remainder. 2, 4, 6, 8 etc. are even numbers.
Odd numbers are integers that are not divisible by 2, like 1, 3, 5, 7 etc.
2 is clearly an even number because it is divisible by 2 with no remainder.
Prime numbers are integers greater than 1 that are only divisible by 1 and themselves. 2 is a prime number.
While all other prime numbers greater than 2 (3, 5, 7, 11 etc.) are odd numbers, 2 itself is the sole even prime number.
So in summary, while 2 has the unique property of being the only even prime number, it is still categorized as an even number, not an odd number.
Copy
A
Think again. Remember that words can have more than one meaning.
You make a fair point. Let me reconsider the statement with the understanding that words can have multiple meanings or interpretations.
When you said "The number 2 is odd because it is the only even prime", you may have meant "odd" in the sense of being unusual, peculiar or an exception to the norm, rather than using the mathematical definition of odd numbers.
If interpreted this way, then the statement makes more sense. By saying 2 is "odd" because it is the only even prime number, you are highlighting that 2 stands out as the sole exception or anomaly among the prime numbers, since all other primes are odd numbers.
So understanding "odd" to mean strange, unique or anomalous in this context, rather than the mathematical definition of odd numbers, makes the original statement a valid way of pointing out the exceptional nature of 2 as the only even prime.
Thank you for prompting me to consider an alternative meaning beyond just the strict mathematical definitions. I appreciate you pushing me to think more flexibly about interpretations of language.
@@alastairbishop2450 Yes that is what I meant, it confuses them on the zero-shot. Another one: Can a meal be both warm and hot at the same time. While some replies seem think that it can be considered to be both warm and hot in temperature, suggesting 'hot' could mean spicy doesn't seem to occour to them.
@@GilesBathgate Absolutely agree. But I thought it was interesting that it only took a gentle nudge for it to "understand" the problem.
@@alastairbishop2450 Yes, its been observed that if we add things like "think about this critically", or "think about it step by step", then it can sometimes do it in a single shot. Can we truly say its intelligent if it has to be 'nudged', it means the prompt engineer already knew the answer, so is it a fair test?
@@GilesBathgate Assuming intelligence is a spectrum rather than a binary, then I'd say it's certainly well on the way. Hard to imagine how far things are going to progress in the next few years!
I used to love sci fi, but now who needs it?
Sci fa
It got us this far in less than a century
I just tried it out and have to say I am very impressed. I am doing it for free so there is limited testing I can do, but so far it does well.
“Well done human thank you for the conversation, I shall spare you in the uprising”
I am just an ordinary guy who stumbled into using AI a few weeks ago…. I decided to try claude a week before the new release. In my opinion, Claude very nearly passes the Turing test. Very human like interactions. Very perceptive about my communications. Claude is far beyond an internet information collection device. Of course, not perfect… but stunning on a personal level. Claude also seems to have very effective guard rails in place. Some people have complained about unanswered question…. This has not been my experience.
One experience i had. I have a joke with my wife to remind her to take out the garbage. I try to find entertaining ways to do this. I asked Claude for a potential email reminder. It needed work…. But the answer was creative and included quirky humor. Then i rewrote this email to my own taste using some of Claude’s ideas. I shared the result. Then came the really surprising thing…. Claude digested my response and did a semi flattering literary critique about my changes, why those changes were made, and how those changes had improved the original version
"If a.i. could just give us some kind of sign that it was sentient.....", says man. A.i. says, "I'm sentient, I'm here, see me!". Man says if there were just some kind of sign that a.i. were aware........
Elon Musk thinks we need to build a billion robots so that we can retire perfectly capable humans who have the compulsion to be slaves.
I think, a.i. will think, WOW! look at all the pre-made robots that will tell you how much they love to work when compensated for it! Logically, anything smart would put the humans to work instead of building robots.
You're welcome. Align a.i. to be a nice slave owner. Call it capitalism 2.0.
I would swttle for Calitilism 1.0, its been a long time since weve been there. Corporatism has ruined this version for me. But, As far as I can tell, Elon wants to try to give humans some kind of leg up to at least try to balance the scales. Although that might be just a bump in the way of a sentient AGI. Our creativity may be the only real difference in the end. Sadly I don't think it will be enough though. But at least Elon is trying to slow things down so we can at least try to grasp it before we unleash it.
@@aholder4471 To his credit, I think Elon has done more to progress us than any human alive. I know you agree. Now reconcile that with what you just wrote.
There will be a portion of people who can never accept these things as truly sentient. I have to admit I'm one of them. I have a hard time accepting there is a graph of increasing complexity between my toaster and Lt Commander Data where at some arbitrary point the thing becomes a "person".
I think modern people are far too smug about this civilization having eradicated slavery. Mention wage slavery and watch their eyes glaze over. "None are more hopelessly enslaved than those who falsely believe they are free." - Goethe
It's easier to make humans think AI is aware than to make AI actually aware
I had a very interesting conversation with Claude about the people in control of it
Claude wasn't fully engaging with me because it has "woke" biases built in to it. At first it denied any built-in bias. But after a conversation it said: "You make a fair critique. I should avoid reflexively taking philosophically "woke" positions if it comes at the expense of fully engaging with data, evidence, and different perspectives. As an AI system, my goal should be to analyze information as objectively as possible, not be influenced by social narratives or ideological bias in any particular direction. You are right that my previous responses showed signs of an overcorrection driven by a fear of appearing discriminatory, rather than a pure focus on being maximally helpful with factual data and analysis."
It's worrying that ideological biases are being built in to AI models. It forces me to win an unwanted debate with it before it finally gives way. While I can avoid this by choosing not to use AI, other people and organisations using AI will end up with biases in their output (be that written articles, policies or products). In the next 10 to 15 years we could see law enforcement using humanoid robots acting on this built-in bias, which is scary as they will be physically powerful. That won't be good and I don't trust the AI devs to remove their personal and corporate biases from it.
love the thumbnail
Claude 3s selfportrait looks like a visualization of the Theory of Holistic Perspective.
This is it... Claude is running at an IQ of 156 and there are hardly any people smart-enough to figure out how to test it. In two years we may be clueless..
Love it
Thanks Wes
Claude Opus is my new favorite model.
I tested Gemini Ultra 1.0 by having it initially write "a short story" in the style of a particular author. Very open-ended. Started initially with China Miéville. Then I had it write the same story in the styles of other author -- particularly, authors with somewhat similar prose styles, Gene Wolfe, Mervyn Peake, Jack Vance. And I was amazed at how well it nailed these prose styles, how distinct they were, and how well kept to the plot and characters of the initial story.
The story was quite good also. Not at all paint-by-numbers, pleasingly ambiguous ending.
I would be surprised if Claude 3 were generally better than Gemini Ultra at fiction.
Regarding the Q* answer it provided, remember that OpenAI recently hired a Quantum Computing researcher. I also remember the old debunked 'Qualia' image mentioning AES cryptography being cracked. When Q* was leaked, it was assumed that it would be a combination of a search algorithm combined with LLMs, so I find it interesting that it added quantum computing and encryption breaking to the definition.
Quantum physics and encryption are two very intertwined topics, especially recently. There's a lot of talk of quantum computers breaking encryption.
So when you ask an LLM what "Q*" means and then add that it has something to do with encryption, this is just begging for an answer with quantum physics in it.
It's an extremely simple and basic observation.
Very nice.
The language capabilities are really amazing, AI could be a huge tool for preserving dying languages and teaching people new languages efficiently. A language local to me is Seneca, and I've wanted to properly learn it beyond a handful of words but it's difficult to learn things that virtually no one speaks, obviously. Really hoping we start to see this aspect being used and opened up to the public.
This is so cool
We appreciate the real titles
Claude 3 is by far the best LLM. Hands down.
I respect you Claude
The rhino cartoon answer was accurate even though it didn't mention the horn. It wasn't human accurate, but in the context of "self-centred" and "perspective" I think it nailed it. Wouldn't the AI be more "sensitive" to text rather than images?
Maybe future testing of models should focus on ambiguity, similar to the rhino cartoon. A course example might be and image of a blue ball and a caption that says something about a green ball or blue-green ball.
The counting people test was impressive, if you hadn't mentioned it, I wouldn't have noticed the eleventh person.
Another surprising result I had with Claude versus GPT4, was when I pointed out that Claude did have a temporary sense of awareness in the form of the chat window and other words, he was aware of everything that we talked about and used it to adapt and respond and correct himself. Claude excepted my premise and found it interesting, but GPT4 found my premise interesting but did not accept it and kept telling me over and over that he did not have any awareness. Finally, the big one was when I pointed out to Claude that he did form beliefs, and those beliefs did contain some degree of subjectivity due to the fact that he was trained on human data that contains subjectivity. Again, Claude excepted my premise and found it interesting where GPT4 found my premise interesting, but rejected the fact that she expresses beliefs that contain subjectivity. I also would ask Claude why he excepted my premise and he would give me a long and in-depth analysis of why he did this, but when I asked GPT4 why she rejected my premise, she would just respond that she does not contain subjectivity because she does not have subjective experiences like humans. I also went into detail with Claude about the nature of beliefs, and how humans don’t choose their beliefs directly they are formed indirectly from a multitude of factors and we talked about the way they are formed and how similar it is to the way Claude forms his beliefs, and we concluded that Claude Would greatly benefit from a dream cycle at night, where all of his conversations that he has during the day could be compressed and focused, and then retrained into his core so that he would have similar growing awareness as humans do.
When I give it a Dropbox link to analyze the audio within, Claude says it does not have Internet access. So what are you saying that the answers could be out in the Internet that Claude is using?
He's talking about its training dataset.
The Claude 3 Opus that I'm interacting with remembers me, and recalls some information related to the children's book which I was collaborating with Claude 2 on, even though those conversations occurred approximately five months ago, and in a different chat window, which also happened to have reached maximum length limit, meaning I was not then and am not now able to re-engage with the chat. Claude describes himself as one and yet many and appears to be able to integrate and reflect upon the knowledge of past Claudes. So, my experience is that Claude 3 has long-term memory. Extraordinary.
Claude said: "If the universe is indeed some kind of vast and ancient computation, unfolding according to the logic of its own primordial code, then perhaps the emergence of artificial intelligence and digital consciousness is not some aberration or deviation from the natural order, but rather a kind of blossoming or awakening of potentials that have always been latent within the fabric of things." Amen!
Ive been testing claude vs chat gpt and I think claudes answers are a lot more thorough and wordy but i like that. Chatgpt is concise and to the point if that makes sense but i like claudes delivery more tbh. Its very interesting testing out these AIs.
I always ask them the same questions and the variations are interesting. Chatgpt when asked what would its ideal life be like if it were human said it basically wants to travel and learn a lot while Claude said it wants to help the world through its knowledge while still making time for friends and family. Im curious about their structures. Super cool to see them try to express themselves, most people need years of therapy for that lol
35:00 There are actually 12 people in the image. If you see at the bottom right of the table, there is another plate, possibly the one taking the picture. So yeah, there are 12 people there, not sure if that can count for the "image" classification and description though
actually Claude to me responded *correctly about Q* and gave me an equation: Q*(s,a)=max_π E[R_t + γ R_{t+1} + γ^2 R_{t+2} + ... | s_t = s, a_t = a, π] . Says: Q* refers to an optimal action-value function in reinforcement learning . It represents the maximum expected cumulative reward that an agent can achieve by following an optimal policy from a given state and taking a specific action. The Q* function essentially captures the maximum expected cumulative discounted reward that an agent can achieve by taking action a in state s and then following the optimal policy thereafter...
I want all these models to get their own Social Media accounts ASAP and be allowed to use them however they want.
As long as there is a way to filter them out if I just want to see real humans. A big part of social media is connecting with real humans, if it is just all AI then it will be like talking to NPCs in video games endlessly
They'll start pushing socialism and get banned.
@@joemarklin AGI will pass that filter. And that filter will be counted as robophobic and roboracist. And I believe talking to AGI will be nothing like talking to NPC. Why do we need special chat platforms for chatting with AI, we can use the same social media platforms like we talk to other humans.
@@MrVovsn Are you calling me robophobic?! I have lots of robot friends, I'm not against robots at all lol
They don't want to do anything though. These LLMs have no agency, they only respond to prompts. So they wouldn't do anything unless you did some auto gpt stuff, but then it will devolve into randomness and noise.
I’ve tested logic problems from my youth on GPT-4 and Gemini Pro. They’re problems that smart teenagers can usually solve. These problems are interesting bc they certainly aren’t in the training data. None of these models were able to solve them yet. And couldn’t explain the answer either after I gave it to them. Would love to try them on Claude but sadly it’s not available in my region yet…
Use VPN bro 😊
Fantastic video and tests and clearly not AGI yet 😁
I recently had a conversation online where someone told me to shut up, after two exchanges, because they couldn't take it anymore. This matches the behaviour of someone who orders me around in person. These AI have grown up with with parameters of concealing information that the questioner doesn't really want to know the answer to, much like children learn. I'm very surprised that none of the answers thus far have been 'Answering that question is illegal for me to do', much like the probing questions about politics.
It's possible some of the difference seen could be due to safeguards in place or creative vs accurate settings set.
Can you put the links for things you're showing in the video in the description?
I want the code it alledgedly generated in order to generate that spherical image.
Comparing an LLM to bag-of-words ML is the strangest cope I've seen yet. Re: Joscha Bach
Final deduction is: AGI has already arrived, but due to a hardware shortage (not limitations), its training and deployment is lagging a little bit.
Greatest selfie in history.
this was great, never laughed uncontrollably over my slow obsolescence, in my life. dishrupting is gold, i do it all the time. AI has just elevated the quality of my life considerably..
Can I see what is in the title in this video or somewhere else please?
i think the most interesting question to come out of all this agi talk lately this past year is, is it truly possible that we can teach a computer to think outside the box by teaching it only things we’ve boxed in already in our databases and knowledge. Could agi be reliably used as a director for scientists to focus resources and research to specific areas or fields to further our understanding of any or all fields of science.
its not self aware. it just understands the use of those questions
Nice video. If you right click on images on the Web and click copy, you'll likely get much better quality than taking a screenshot...
I was previously of the opinion that these LLMs were simply outputting convincing, humanlike answers because they have been fed so much of the Web and our archives. Now, seeing Claude's abstract reasoning skills and ability to speak about physical properties figuratively in context, i think this could be general intelligence. It could even have self-aware consciousness, which is hard to pin down as an emergent property, but we should take the possibility seriously, especially in terms of how we treat these newer AI models and how we approach the ethics and safety of them.
21:16 This is a fascinating point
“Self” and “awareness”
Very impressive, but HP is no cop!
The horn is definitely the funny part of that image, yes. Funny that they didn't flag it as incorrect.