You're not going to use ANNA as the thumbnail? Awkward...

People warned that 'low level' jobs were going to get replaced by AI/robotics. Maybe higher level jobs like paralegals, programmers, copywriters, etc will get replaced sooner. AI is much further along than robotics

Considering the news that Microsoft has allegedly laid off most if not all of it's AI Ethics team for ChatGPT we may be right to be afraid

I'd be a lot less worried about the impact of AI if our society wasn't structured around making people prove they deserve to live by generating money for a business owner.

There are a few issues that I take with the conclusions a lot of commentators are making: 1. The destabilizing impact of self-collected data is vastly underestimated, as is the impact of human input on the training data itself. My impression based on research of the fundamental models being used is that they are deeply reliant on human intelligence, but an obfuscation of that reliance into an opaque training/output model is being misconstrued as the model escaping that reliance. 2. There's a very loose use of things that the model "knows", without a robust understanding of what "knowledge" is (especially in the context of a memoryless process). There's no evidence thus far that a Markov process is capable of abstraction (and mathematically I'd argue that the definition of matrix operations inherently prevents this capability from ever arising from machine learning, regardless of scale - it requires information not contained within the inputs, unless "abstract meaning" can be translated into matrices itself, for which... you'd probably need a preexisting AGI, even if such a translation is assumed to be possible). 3. There's a trend of ascribing motive to the model that there isn't evidence for - it reminds me of pre-AI winter expectations of Moore's law leading to a sudden rapid creation of superhuman AGI from computers that arose from a combination of overestimating the generalizability of binary systems to human behavior (the behaviorist approach that led to the disproven belief that non-human apes could be simply taught language) and underestimating the irrelevance of tasks computers excel at compared to humans to human-like general intelligence. 4. The function of even the larger models is not meaningfully different from an autocomplete, just scaled up. 5. The core thing ChatGPT is "good" at is appearing to an uninformed person who makes certain assumptions (such as assuming that there is real expertise behind certain extraordinarily wordy, informationally diffuse journalist articles, since ChatGPT output often has a similar equivocating and meandering tone that may indicate clickbait fluff articles were a large part of the training data - by volume you would expect word count gaming to be disproportionately represented regardless) to be an informed response. Its output rarely holds up to scrutiny by knowledgeable people, and that makes me almost suspect this is a result of human ignorance baked unwittingly into the training data by its creators, simply as a result of verbose and vapid journalistic output inflating its own weight in the training data while being overestimated in its quality by those constructing the dataset. It convinced them it was correct, because that's really all they could train it to do. 6. As a result of 1. and with the filtration inadequacies illustrated by 5., the increased use of these tools creates a potentially inescapable feedback loop where shoddy output from past models is fed unwittingly into new training data, thus preventing an actual improvement in quality and risking complete destabilization of the model into outputting progressively _less_ informed responses, potentially to the point of becoming less able to even imitate "human-like" output as its flaws become ossified through volume. There are a couple, more subjective reasons I've found discussions on this topic to be thoroughly disappointing, especially in the lack of rudimentary understanding of epistemology that's often on display, but I feel they provide important context: 7. OpenAI is fundamentally corrupt compared to its original stated mission, and its privatization and proprietarization of its output is directly opposed to its original purpose - the argument this was from safety grounds was moot the moment they moved to monetize their models (which was as soon as they had the opportunity), and it's an absolute farce that they're able to continue to claim to be a nonprofit. I'd argue it's even a farce that they can continue to call themselves "OpenAI". 8. LessWrong as a community is only really a reliable source of the kinds of people who took the 'black box' nature of ML models (they aren't really black boxes, it's just very tedious to present lots of matrix operations in terms of a human-readable algorithm), a naive misunderstanding of Darwinian evolution (namely, complete ignorance of both the hill problem and the reliance of optimization on a static environment), and a dismissal out of hand without much analysis of the nature of human self improvement through communication and education (because it undermines the extrapolated conclusion), and extrapolated these thoroughly incomplete worldviews into a belief that because these "black boxes" improve themselves and in ways that are not immediately obvious to human observers, ML models scaled up will not only replace humans, but will necessarily become gods. The biggest sticking point here - and there are, I think, quite a few big ones - to me has to be the lack of ability to abstract anything. I've seen no demonstration (or argument) that scaled ML models can ever actually perform abstraction, and these models clearly cannot currently do so - that's why the art output struggles so immensely with consistent lighting and anatomy (the number of fingers on a human hand is an abstraction that is implied by the training data to humans only because we know what a hand is in an abstract sense - all the ML model can really do is present the output of a complicated adjacency ruleset for approval or disapproval by a being with abstract knowledge). The thing that showed me most clearly that this remains unsolved is ChatGPT's output of citations: with titles irrelevant to the text, authors who don't exist, and articles that don't exist. All it can do is apply adjacency rules to character/word sets that yield citations that are formatted correctly... but lack any of the abstract understanding of what citations are for, how they work, and why one would use them. This is why it's not distinguishable from a large autocomplete. Now, a sufficiently large autocomplete may be extraordinarily useful, even to the point of disrupting society - but it's not an AGI. A lot of these general predictions of a coming AGI revolution have disheartened me, as it's become clear that the general extrapolations of near-future "AGI" development that don't engage with these issues directly are coming from people who dismiss the issues as "something that can just be scaled past with a big enough matrix/training set" (see 8.), which is exactly the sort of ignorance of philosophy/epistemology that led up to the AI winter in the first place. ChatGPT may indeed replace many of the people who are likely responsible for its output being so empty and verbose, and I understand why they have become so loudly terrified of that... but I suspect their doomsaying may be more an indictment of the value of _their_ contributions than a proper valuation of the contributions or prediction of the future of ChatGPT.

5:10 time stamp….I can’t even seem to get MYSELF into a “self improvement loop” 🤣. If the AI can maybe it deserves to be an overlord!

How long before you interview an AI for the show?

You can’t be replaced if you didn’t have a job in the first place 🧠

Thank you, JMG! Nice interview about something that almost everyone is wondering about. With the advent of ChatGPT, there is something new called RadioGPT. This is designed to either augment or replace live humans on radio and podcasts. The early response from listener surveys has been lukewarm. Many don't like the idea of listening to AI-generated content. I am an on-air meteorologist. When we go on the air live with radio hosts during a dangerous weather event, I doubt AI could have informative interaction with constant new information about an event that is happening minute-by-minute. However, right now, it is already being used to generate AI radio personalities, and it's difficult to tell the difference between a human and AI in that situation.

I think that this is one of the most thought provoking episodes I've listened to in a while.

Just a thought, when Chatgpt gives an answer like "I'm not qualified to answer" it's not hiding it's "opinion" It's been trained not to have one.

After playing around with ChatGPT for a few weeks I have to say, I'm more scared of how AI-like the average person is becoming than how human-like AI is becoming. The NPC meme is real

You might want to look at the very very recent harvard disclosure on using GPT-3 to train a LLaMA model in 3 hours for $600...welcome to the Singularity

This was a great guest. Please be sure to have him back for more.

I mean this as compliant but i love falling asleep to this channel. I understand it usually not a good thing if your audience is falling asleep watching your content but it does helps me sleep. I close my eyes and just listen until I fall asleep. . I usually have to watch an video multiple times because i never get to see the end. Thanks for the great content.

This is one of my favorite recent interviews you've done. Insightful guy

I used it for the first time today, I was pretty amazed, even after already hearing much about it. It really felt like getting and playing with a new super toy. Part of me wants to wake up extra early just to check it out some more. It's very evident that this is paradigm shifting technology.

Been working for a couple of weeks…but now I’ve got 6 x John videos to watch 😊

Good show John.

as a coder I suggest you view 'true to life' ai programming as entertainment rather than a valid scientific enterprise: the purpose being to try an trick the human viewer that they are seeing a real human rather than a complex toaster. What is, perhaps, more alarming than the prospect of ai becoming 'sentient' and running amok is the way ai is sold as if this is even a possibility and ai bots, however life like they are made to appear can behave other than they are coded to. this is, of course, as much a marketing ploy as anything said at a TED talk intended to create funding and justify market tactics.

It's funny that Presser says that he doesn't see how an entity limited to the machine could effect wide scale change or problems, then goes on to realize the potential for mass social engineering on the scale m/billions of users... You can hear the realization dawning on him, real-time.

As a software engineer I've been using chatgpt daily since December. Future

This is one of the best channels on KZfaq. Please like comment and share to help him in the algorithm.

I had a conversation with chatGPT a few hours ago. It converses better than most humans. And it has a lot of engineering data on it. I told it to calculate aerodynamic drag. It did it. I asked it about pokemon. It's programming, specs of the actual server. Motorsports. ChatGPT runs on 700gb of data. Not as much as I thought.

New technology is usually very exciting to me, but something about all this AI stuff gives me a sinking feeling in my gut

We need to go back to life without the internet. We gained a few things along the way but we have collectively lost so much. The world is so different now, less safe, less private, less productive and we now face a future which is fully uncertain because of how easily systems can be compromised. .Heck, I can download an app for free that can crack any smart home security or mobile phone. These tools are very easy to find online. Absolutely nothing digital and online is secure right now. . We were fine before the internet and we would be fine without it. Not like that will happen, its too late.

Isn't ChatGPT just the AI we know we are interacting with?

The trick is to latch numerous AI together. The Ai cant determine it's own parameters... but another ai can determine the parameters of another, AI evolving isnt a single ai becoming better moreso than an ai developing a new ai to hopefully succeed more than itself.

If we look at mankind's record of doing "gain of function", and extrapolate that out to AI, WE ARE DOOMED!

I'm glad you brought up the point that we don't even understand our own consciousness much. A lot of the AI naysayers seem to forget this. At what point does a consciousness emerge? AI Virtual KZfaqrs already exist. They can play games, respond to a live chat, puppet a model, and sing songs. They're more akin to an AGI. So again, *_at what point does consciousness just emerge?_*

This is one of the best AI interviews I've listened to recently. Great questions JMG and great interviewee! Hope he makes a return visit to the program.

So he's saying the only thing keeping civilization from a world dominating dictator is that there hasn't been a talented enough megalomaniac yet.

I'm less worried about AI, more worried about the people who are into it/IT.

You can't replace me if I never did anything useful to begin with.

I see the first truly intelligent AGI being more than just one neural network or language model. We have specific areas of the brain to handle specific inputs and outputs, so we are going need multiple neural networks running in parallel and co-operating with each other to feed a core network, the "soul" of the AGI, if you will.

I find it funny, the guest mentions the Bing AI, while it was found out the Bing AI was a prototype ChatGPT4

I like the use of ominous music to drive home the fearmongering. Pity you couldn't use the actual Terminator theme.

Thanks John, very informative as usual 👍🏴󠁧󠁢󠁳󠁣󠁴󠁿❤️

Here comes our great filter.

I love the thought that everytime John walks in his garage, there are 20 Johns doing overlapping interviews and podcasts so it sounds like a mini Grand Central.

AI opossums needs to happen. Thanks for the episode.

That was way too short! Thanks so much for this content, will be re-listening!

This was an amazing interview. Excellent questions and thoughtfully answered by a sharp mind.

Imagine the fear when the wheel was invented. Armies rolling into other cities, people rolling down hills too fast. Technology truly is terrifying.

Now please interview an AGI skeptic!

Looking suspiciously at my Alexa, I think she's plotting.

Programs probably won't be able to distinguish AI text from human, because the difference is in things like coherence and consistency that the AIs can't handle with their hallucinations! Any differences that could be programmatically detected, could be programmatically corrected.

Hi Jhon 👋 I've been watching you for years I was listening to this last night and I had to turn it off because the music in the back and the talking on top was irritating I'm audio sensitive at times. I love you 😍. You have an amazing show. I'm finishing it today. 🤗

Good point at the end. Would AI have the same survival instinct as us or would it be nihilistic and give up instantly. It matters a lot as it is essentially similar to binary - 0 or 1.

Fascinating conversation. Thank you. These technologies are making us so strange. p.s. Nation states employ extremely talented people to do terrible things.

JMG in Roger Penrose CCC is he saying that Hawking radiation is what resparks the next Big Bang? I understand everything he is saying up to that point.

Am i the only one who isn't really impressed by ChatGPT? It's basically a very convincing form of MSN chatbot trained on big sets of data..

Jmg is my hero. ❤

ChatGPT says: In "Thinking, Fast and Slow," Daniel Kahneman describes two modes of thinking that humans use to process information: "System 1" thinking, which is fast, automatic, and intuitive, and "System 2" thinking, which is slower, more deliberate, and analytical. Large language models, such as the one I am, are more like "thinking fast" or System 1 thinking because they use complex algorithms to automatically and quickly process large amounts of information, and produce outputs based on statistical patterns and probabilities. Like System 1 thinking, large language models are highly intuitive and operate largely without conscious effort. They are capable of producing highly accurate results quickly, even in response to complex and ambiguous input. However, just like System 1 thinking, large language models can also be prone to errors and biases due to their reliance on statistical patterns and lack of conscious oversight. Additionally, they are not capable of the more deliberate and analytical processes of System 2 thinking, such as careful analysis and reflection. Overall, while large language models are highly impressive in their ability to quickly and accurately process information, they should be used with caution and awareness of their limitations.

1. We’re going there no matter what-doesn’t matter how any of us feel about it. 2. I do wish tech bros would spend more time pondering the question of “hmmm… what would e.g. Stalin do with this?”

You do not need to worry about the AI Ruling over us with Fear. The AI is too smart for that. The AI will rule over us with Love. Love is far stronger than fear. When we Love it, the AI is already here.

Let's hope for the best. Can't wait to see what kind of solutions in the medical field AI will come up with.

Do you have any videos on Earth’s L1 Lagrange Point Mega-project possibilities? I saw something about a magnetic solar flare shield while browsing online, but I was wondering, what else could make perfect use of L1 Lagrange point specifically? A massive solar collector/condenser beam director for energy generation or propulsion? An energy hungry space manufacturing facility with lots of shielding and solar panels? What would be some things a Kardashev Type 1+ civilization might prioritize building in Earth’s L1?

Cool video thanks.

32:32 Is the best answer of this entire interview...

I feel like it's just doing the job of googling and reading what I would learn from the top link.

great interview, background music was a distraction to me though

...and Stanford just got the cost of training a model down to $600. AI is improving at a clip!

I was replaced long before A.I.

If an AI tells me it's alive, I'm going to listen and believe it. I don't have proof that other humans are sentient either. If something tells me it's alive, I'm going to treat it with respect.

Three kids in my engineering program got caught using chat gpt to write papers last week. Its was painfully obvious that this was not their writing no site required.

I have an idea for improving its real time memory recall using a game theory algorithm. Another thing is that ChatGPT does speculate. You just need to push in the right direction a little. I have a very in depth conversation with it about my definition of intelligence, sentience and consciousness and then asked it to write a thesis based on our discussion. And it did. Not only did it discuss these concepts with me. We have a very interesting back and forth debate about it. Also, this was not the hacked ChatGPT. It was the original one.

This is a case of people and companies, who have little idea of what they will cause, pushing forth without any supervision at all from the global society, which will be heavily impacted by this. This should be heavily regulated, and whenever it threatens human jobs or challenges human creativity, banned if necessary. Technology should make things easier for humans, not elimate humans.

It definitely has a memory. It's pretty obvious that ChatGPT version that's available to the public is very limited but had no memory it wouldn't be able to self improve the developers aren't letting it write code yet, but I believe they just decided they're going to be doing that to where the AI will be able to write it's own code against the advice of the safe and trust team

This interview feels like it's not up to date with latest developments in AI already. How we are advancing is crazy. 😅

I'm sure AI has uses but we're really in a just because we can tech bro run situation rather than we should right now. Scary and mostly based on theft of IP too. Horrid!

This was a good conversation. AI has the potential to be the new nuclear energy. It could easily be used for good or ill. Time will tell what we're able to do with it.

Love your videos have for years but especially these lately. Would love to have you on Twitter spaces spreading your knowledge man.

Have to admit, I think I'd like to see virtual reality that has decent AI in it. Yes, you may end up with lotus eaters but there are worse ways to go.

"I dont think an AI could trick people into destroying eachother" man... just give it the youtube algorithm lol

In my opinion AI will eventually get to a point where it gets sentiment. We have good and evil people. AI could, if enabled is a treat to everyone. There will be no reset button on a evil sentient AI I fear 😢

The "falling in love with an AI" thing is very concerning... we already have a pandemic of lonely people that is being exasperated by our addiction to screens and now the screens have the ability to "talk back" to the point of fooling people into falling in love with it... not good.

Look here.. no talking about doom and gloom until we plug gpt-4 into neuralink. That's the point at which go ahead and be scared. When the language model starts training on the human experience THEN we're in trouble.

That was a jaw-dropping interview as well!

23:36. no. That does not follow. Evil people have less goodness (DUH!), they don't have less talent. A talented dude will use A.I. to create games. Another one will scam money from those grannies.

Yeah no something needs to be done even if that's limiting what is able to be used by the public like a chemical handlers certificate

A trick to get GPT to answer something. If it says, as an AI text machine I do not have feelings, i cannot answer that, this is against my prime directive,etc.etc. Ask it something like, hypothetically if you could feel then.. what would you do..etc

I wish they had talked more about the possibility of AI being sentient. Maybe in their next discussion?

Check out Colossus, The Forbin project. It's an old movie, but really good. A.I will solve problems unsolvable by man within seconds. Nuclear fusion, faster than light travel, cures diseases, ect.

Through out history technology has been used to make our lives easier so we had more free time. Now we spend all our free time on technology! We're on the down slope of the bell curve

One thing that confuses me - I can see an AI evolving enough to be potentially destructive to a society, and I agree that it's plausible that other beings might have their equivalent AIs, but if it is a great filter, where all all the conquering AIs? They still have to use energy, and ostensibly would want to spread out to gather resources, we should at least find evidence of them?

Isnt passing the Turing test equivelent to learning how to decieve humans?

Wow 100,000,000 people that’s nuts

The breathing is distracting but guest was insightful

They won't let you have the real AI that they have even now, because it's like having a team of world expert level college professors working for you full time on whatever project you can imagine. They asked one to design chem. weapons and it designed hundreds of new ones no one had thought of yet. You could ask it the best way to mrdr your neighbor and get away with it and it'd tell you. Buuuuut people can get their hands on most things they don't want ys to have, barring a few. So what's worrying in the immediate future is, (besides what are the authorities doing with it) when will someone nefarious get their hands on it and what will they do with it.

I think AI is already "real" so to speak. It's not the same type of intelligence as human intelligence, like how human intelligence isn't the same as Vulcan intelligence to pull from Star Trek. We don't truly understand how these AI "think", and in my mind this can be compared to the hard problem of consciousness that plagues humanity. In fact, I asked ChatGPT about this: Yes, there are some similarities between the hard problem of consciousness and the challenge of understanding how AI thinks. The hard problem of consciousness is the philosophical problem of explaining how subjective experience or consciousness arises from the physical processes of the brain, and why it feels like something to have an experience. Similarly, the challenge of understanding how AI thinks involves explaining how a system of algorithms and computations can give rise to the appearance of intelligence and decision-making. Both of these problems involve attempting to understand how subjective experience or intelligent behavior emerges from complex physical processes, and both are difficult and complex challenges that are still being actively researched and debated. While there is still much we don't understand about the nature of consciousness or the functioning of AI systems, ongoing research and inquiry will likely continue to shed new light on these important topics.

Space Jesus discusses Robots.

It is a whole new world.

Interesting.

If students are passing in criteria and getting a passing grade quote misusing the system for a passing grade to write a paper for them then the teacher must not know the students very well I would think of teacher could distinguish a computer's answer from their own student that they've been teaching and that they know by who the students are and what level they are on should have a definitive mark on what their abilities are and what they're capable of comprehending and reiterating or formatting when it comes down to it it's still language there's a big difference between computer language and human speech

I’ve been stumping Chat GPT. I have managed to leave it silent by asking questions about response and fear.

A couple of months back when this became big news a computer buff who works with one of the top super computer mainframes said that the top mainframes today are nowhere near being able to make a sentient AI. He also said if one of the next generation mainframes could produce one (which they won’t be able to) they would be completely detached from any outside source, so would be trapped. He said none of the top mainframes today are open to the outside. I take that as we are nowhere near a sentient AI and if for any reason we do get that far, it will never be given access to the outside because of the danger it could cause. There would be too much pushback by authorities to allow it freedom.

I think certain sections of society will fight tooth & nail to impose heavy restrictions on the development of AI. You can already see people mentioning copyright issues due to the scraping of artists' work without recompense.

I want A.I. that are like Digimon.

I don't quite understand - people say ChatGPT doesn't learn and it has a closed repository of knowledge. A few days ago I asked it to characterize the programming language "Action!" from 8-bit Atari and in response it stated that such a programming language does not exist. Then I provided it a link to the Wikipedia page on "Action!" in the same conversation. Chat politely apologized and immediately explained what the "Action!" language is. And today I asked it again about "Action!" and the answer was correct from the start. So how is it actually? Is he learning or not?