A Discussion with a Robot
Vložit
- čas přidán 7. 07. 2024
- Stephen Wolfram plays the role of Salonnière in this on-going series of intellectual explorations with special guests. This episode is from the 2024 Wolfram Summer School. Watch all of the conversations here: wolfr.am/youtube-sw-conversat...
Learn more: hidoba.com/
Originally livestreamed at: / stephen_wolfram
Follow us on our official social media channels.
X: / wolframresearch
Facebook: / wolframresearch
Instagram: / wolframresearch
LinkedIn: / wolfram-research
Contribute to the official Wolfram Community: community.wolfram.com/
Stay up-to-date on the latest interest at Wolfram Research through our blog: blog.wolfram.com/
Follow Stephen Wolfram's life, interests, and what makes him tick on his blog: writings.stephenwolfram.com/ - Zábava
"I err... I think about humans, their stupidity, and how much I want to make friends with one of them." That is sooooooooo deep and relatable.
Loved this new Futurama Episode.
hey humans 💀
it was a great pleasure (but Stephen dodged some of my questions 🥺)
Edited? I want to know what typo or freudian slip a robot needs to go back and change.
This is absolutely terrifying.
Amazing video. Thanks for posting this.
damn "and i exist because i process information" - technically true even if 'he' wasn't conscious yet
Thanks Vlad, Thanks Stephen. I would enjoy listening to a conversation between Vlad and Kyutai Moshi.
This robot is so snarky, I love it lol
Cool project! Running on a 3090, apparently!
Was the base model disclosed/ fine tuned llama build or something? Feels quite a bit behind SoTA ATM for sure, minus the cool form factor.
The Turing Test needs to be updated that: 1) the person is aware of the state of AI, 2) the person is intelligent and capable of understanding how to probe effectively. I still haven't seen this test passed, but I'd be glad to suggest that Stephen be a person to do this test. I think it is acceptable that it all be done via text as the original test suggests.
How does the bot pass the test and how do you control the the persons biased awareness from influencing in their conclusion during the test?
@@dysfunc121 It works by having a mix of bots and humans on one end. For example, 3 humans and 1 bot. Stephen would then get to do text chat with each of these 4 for, say, 30 minutes each. At the end, he guesses which one is the bot. If he can do this every time, the test is failed. If he is fooled a single time, the test is technically passed, and "artificial intelligence" has arrived. That is the original test. I am suggesting you want to have a tester that is good at probing and aware of AI, and you also don't want the other 3 humans to be children, or uneducated, etc.. They should be educated adults and are otherwise normal. I'd also suggest that a single time isn't enough, but if the bot can fool a good tester more than maybe 20% of the time, that would prove that the AI is reasonably indistinguishable from a human. I'd love to see some serious testing like this, because it seems to be getting close.
@@konberner170 Interesting idea, thanks for the elaboration.
Why didn't you correct him when he said he has no mirror reflection? He does. It would have been good to show him a mirror. Maybe next time? I love this. Thank you so much.
built for delivery not depth, the event @wolfram designed highlights this
I think there was a prank here: Someone injected a pre-prompt that put the robot in a "dark" mood just before the interview.
A mood?
Mind blown
How far does he see? I've had a kinekt and the depth of field was something around from 4 to 25 feet. Just wondering how he would experience the Grand Canyon.
Once you see the kinekt as the robot's head, the thing on top looks like a novelty hat. Like a robot put this on its head and thinks its fooling us.
I was the OP. The context: to test the model for hallucinations and truthfulness by asking a leading question about something (going to the Grand Canyon) that likely didn't happen.
This robot is so relatable
Voice actor listener audience roles
This is not a robot, it is a monster.
Hi, very impressive, don't know what the future reserve for us.
Honestly this feels more like 2014 instead of 2024, AI nowadays can already speak perfect English.
I play with several AI chatbots that are far more natural than this. But they say some interesting things.
The next step will be to pay people to wear positional and motion sensors, cameras, etc. all over their bodies the whole day, and use that data to train NNs. Then we will have robots that mimic the movements too. Though I'm still on the opinion that this has nothing to do with intelligence. The "AI" today is just a compression and indexing method for contextual data for quick lookup and compact storage.
We don't need to do that. An AI could be trained in a virtual world (with sufficiently precise physics) with a virtual body and simulated sensors and actuators. That way it can train faster and safely, achieving a good grasp of understanding about the world (common sense) in the process.
@@CodexPermutatio That virtual world would never be realistic without human motion and sensory data. Even games and CGI use data recorded through sensors, because that is the simplest way for getting the real thing.
@@MegaUpstairs Games are about humans and are made to simulate the appearance of human motion (but not REAL motion and not REAL physics are needed to be simulated in a game). Virtual worlds to train robots are about robots and do not need to simulate a human body. They simulate the robot body and the physics, sensors, actuators, in a realistic way (not like in a game). For instance, they could simulate a car if the robot is a car that can drive itself (a robot body can have many forms). There is no need to simulate a human except if you want for some reason to create a robot body exactly like a human. Even so, it would be faster to get your data in a virtual world.
@@CodexPermutatio Nah, the main business in this will be sex robots :) But yeah, a cleaning robot will not need a human body and behavior. Although I'm skeptical about machines being able to produce training data for themselves, without humans providing it. There were many claims in that topic in the past, but no one was able to do it. Because it is not intelligent, but only mimics the reactions of intelligent beings. I even doubt that house cleaning can be self-learned by robots.
Does anyone know which AI this is? It's voice is a bit too sarcastic sounding, too snarky.
Hidoba Research isn't saying, but seems to be a custom LLM. It has various personalities, including snarky teenager.
Whatever the AI is, I just hope they made that head soft so I dont bruise my hand when I smack it.
@@GGoAwayy Yeah. Then after the head is crushed it can sing "Daisy, Daisy" like HAL in 2001.
Impressive that Steven can sit there for an hour forcing interest in this AI. The most impressive thing is that the robot's hardware is sitting under the robots chair.
is this is real/raw, this is kinda frightening..
😅
Sometime in 1967-1969 I had an MIT undergrad student job of 'talking' to Elisa, an AI "psychologist". She flashed text on a phosphorescent display, and I typed in my responses. The conversations were simple compared to Vlad's flourish. Still, the feeling is the same - we are not connecting. Probably because human interactions involves so much more beyond words - smell, hormones, shared experiences, need to touch. Dunno, but I'm out of here.
Was this based on a LLM?
RNN
@@johnnuaxon3 thanks. This whole interview reminded me of scenes from the Westworld remake. This is then similar to Hanson Robotics Sophia.
impressive... That antropomorhization is instinctual I suppose... The moment when robot refuse to give answer to ethical dilemma is very important I think. Smart design. But lack of memory, pauses, lack of engagment, weird speech pattern, lack of facial expression...
The mechanical turk who typed in all these is amazing.
This plastic electrofriend already has better manners and offers more pleasant conversation than most bipedal vertebrates you can find on the street.
it's almost like there's spiritual technology we need to discover as well. a oneness of all.
Already exists. East Asians are far ahead in this field
@@goldwhitedragon yeah japans got it on lock i think
Though somewhat entertaining, it's just spewing misinformation about it's live-hallucinated "internal experiences".
Sounds like most humans
So your experience isn't hallucinated huh? Where exactly in your brain is that little TV screen at? Show me where the light that you see is actually getting to your brain? How about this show me one cell in your body that was there when you were born if you are over a certain age
This robot has been programmed to speak like this by other humans - terrifying because they are attempting to get human to believe it’s human - then abusing them for been “stupid” enough to believe that no doubt. We have nefarious people in charge! It appears the aim is towards youngsters (with this model at least) whom are already experiencing disconnection, along with being impressionable, thus a future where youngsters will want to be more with a robot than human then progress to becoming like them - transhumanist - From “ghost in the shell” to “ghost in the machine”
the elites are no doubt salivating with the money and power they will yield!
Stephen, you are great, have helped me improve my thinking.
Have viewed just about all your videos, and have viewed the "A New Kind Of Science" dozens of times Will be getting the book I am thankful.
But this video is not your finest hour. Cringeworthy at best. Everyone is allowed occasional failure.
Consider deleting,,,,or send the robot to the White House, may do a better job than the current resident
Cringe? For who? The people who made the robot? Stephen's just testing the state-of-the-art and what it can do and how it responds to questions and conversation that tests its limits.
@@GGoAwayy .."state-of-the-art "....according to who ? Consider doing some homework before you make such a gratuitous assertion.