Experts' Predictions about the Future of AI 

Lmao even AI researchers are guilty of saying "yeah AI will take over every other job, but not MY job because my job is special!"

I love the fact that people in every industry think their own industry will be fully automated last

When are we getting "AI predictions about the future of experts"?

"5% chance of human extinction is a concern." I dislike a 5% miss chance with XCOM, let alone with human extinction.

Whenever I interpret a survey I have this one story in my mind that I once read in a book. Its about two priests that got into an argument. One was holding the believe that you shouldn’t smoke when you pray and the other one thought it doesn’t matter if you smoke or not while you are praying. So to settle this dispute they agreed on sending the pope a letter and let him decide what is correct and what is not. So both priests sent the pope a letter. After a while, they both receive an answer from the pope. The first priest asked the pope „Dear pope, is it allowed to smoke during the prayer?“ The answer from the pope „Of course you should not smoke while you pray - You should focus on the prayer!“ The second priest asked „Dear pope, can I pray while I am smoking?“ The pope‘s response „Of course my son, it is always a noble act to pray in every situation in life“. Its easy to provoke the desired answer by changing the way of asking the question.

"and Robert Aumann didn't even agree with that" LOL

The green tint of the video reveals that it was recorded from The Matrix

This video was made before AlphStar and OpenAI's new language processing technique, so there are new data points now - AlphaStar: The experts, on average, thought StarCraft is going to take 6 years, but it took 2 years. OpenAI language model: experts, on average, thought Ai writing a high-school essay was ten years away, but it also took two years. in both cases, NO estimate predicted the achievement to come sooner than it did. what we can learn from that , at least regarding AGI, is that the experts don't have very good predictions (though still! it is better than the average population), and when they're wrong, it usually happens sooner than they thought.

I do think AI safety should be focused on, but I dismiss any alarmist who says something along the lines of "We need to stop developing AI" or "We need to lock AI down so that only a few people can use it." I don't think we *can* stop developing AI, and I certainly don't want to consolidate more power in the hands of corporations or governments.

Agree with you 100%. The "spooky" emergence of a machine consciousness is not and should not be a primary concern, and seems rather unlikely. The issue is that you don't need to be alive to make intelligent/dangerous decisions. The primary concern should be the nefarious use of powerful machine learning/AI implementations.

I love that "No take! Only throw!" comic SO MUCH. Can't help but smile every time I see it

That's amazing, a pleasant surprise, didn't expect a new video until at least next month.

0:25 I really appreciate this subtitle, and love your independent channel for it. I think as time has gone on, you've become a better and better advocate for what you do, and it's wonderful to see

Your choice of music in various interludes continues to impress me, as well as scientific content of videos. Good job!

I got *way* too excited when I heard The Future Soon at the end. :D I'm always entertained by your covers, Robert.

This is the best breakdown on this topic I have ever seen! I really commend you for this video.

3:52 So 20% of the researchers expect a neutral outcome of HLMI. What would a neutral outcome look like though? I can kind of imagine the "all work is automated"-utopia and I can also imagine the human extinction scenarios but I can't really think of a neutral scenario.

"They set the system to extreme values." AI builds a near Utopia, except ants now outmass the atmosphere. And are made of diamond.

We really need an update on this

Sent here on Daniel Schmachtenbergers recommendation. You've got a new subscriber. Brilliant explanation for non specialists

Just realised RU-vid isn't showing me your videos in my feed.. despite clicking the bell, subbing and watching a tonne of your stuff. What the hell? Regardless ill just keep an eye out myself now... love your stuff mate :)

Excellent choice of song for the outro.

Gotta love the choice of end screen music. The Future Soon - Jonathan Coulton

I predict that if they ask researchers again in ten years' time, they'll end up with roughly the same graph with "Years from 2028" written below.

Boy the StarCraft prediction aged just as well as the Go prediction.

Is that a cover of "The Future Soon" playing at the end there?

I like his face. I don't know why, but it's nice to look at.

I haven't seen your video before. That was great. I subscribe.

The Jonathan Coulton at the end was a great surprise!

Thanks for helping us all to understand the latest. A worthy service.

Hey @Robert Miles, Its been over 3 years since this video, and 5 years since the paper. Im curious to see how the trend has held. Have there been any milestones ahead of schedule? what about changing directions in AI research since the paper? Mostly just a followup to see if the trend (as controversial as it is) is "on track".

I would really love to hear your thoughts on AI and imperfect information games, and on the AI that beat 4 of the best no limit holdem players recently.

Thanks for this video Robert!

I predict the next task machines will be able to do better than any human is answering survey questions consistently ;D

"Cause it's gonna be the future soon, I won't always be this way. When the things that make me weak and strange get engineered away. It's gonna be the future soon, never seen it quite so clear. When my heart is breaking I can close my eyes - it's already here"

Excellent video as usual.

I did some work in the 80s with early AI. I wouldn't describe our efforts to apply expert systems and natural language as particularly successful and I became pretty pessimistic about AI's capabilities. Fast forward to today with self-driving cars, voice recognition and machine learning of repetitive tasks, I am no longer skeptical of what AI will be able to do. That leads to my intense fear of what AI might lead to. Robert points out in the video that the goals of AI might not match ours. It's far worse than that. I am certain they will NOT match ours because some AI will be created by our enemies. Even if we found out how to control that, what about careless people who set loose thinking machines with goals that miss critical items - items that could lead to famine, climate change, etc. These "careless" machines might be wildly successful. Will we need or have AI cops & prosecutors to track down these rogues and eliminate them? Another issue to me is runaway intelligence. When the AI is smarter than us, how will we know when it's going down a path to Armageddon? Do children know when their parents are out of control? They don't have the experience to know that, nor may we. We need some deep thinking, planning and cooperation among nations to make sure we do not succumb to our own creation. I am 69. I don't fear for myself, but I fear for my grandchildren.

the Future Soon ukulele cover at the end is a nice touch

Presentation was very well done!

It would be interesting to do a survey like this again now that we have superhuman Go, and at at least human level Starcraft (2) a bit before the average expected here. Both were 'only' games, so I wonder if it is fair to update all your estimates to earlier, or only the game-like ones.

Loving the ukulele JoCo!

You have a gift for explaining things.

A.I. research will be the first thing AGI is good at because it will be a bit of agent code that takes video only full self-drive, text to speech, speech to text, route planning, facial recognition, emotion mapping, iterative brute force 3D design, etc and incorporates them into itself. It will probably start as code designed to build teams with reqired skill sets through sites like LinkedIn. That is to say, finding humans with the skills a job requires is the same as finding software that does the same but it can cut and paste the software into itself.

People tend to be much more confident in their knowledge in a subject when they know the least about it.

Robert I love your communication style!

Maybe the key to solving Stuart’s problem is to give the agent multiple utility functions, allowing it to decide on which goal to pursue based on the output of some higher level agent optimizing for positive feedback from a human in real time, and placing a time limit on how long it could pursue a particular utility function. You could possibly train this system like a baby

Really enjoying the easter egg music outros

The AI researchers say we need more AI research... On a side note, my plumber told me that I need a plumber.

That was such a nice, concrete conclusion.

Keep them coming Robert! Step up the game!

They should redo that survey today and see how it changed.

Robert, you should conduct a similar survey on your channel. Let's see what your viewers think about this issue. You have got 21K subs which is a good number of subs and I believe the result would be really interesting.

I love that the emergence of consciousness is described as "spooky". I hope that's a reference to "spooky action at a distance". A 5% chance is still terrifying. I think a better question could possibly be; what are the odds of AI becoming uncontrollable by humans at what point in time?

I would like to see an update on this after gpt-4

5 years later, how did ya do boys? Oh dear. I'm beginning to wonder if Popular Science is a satirical journal.

Just dropped a bomb on people who pull the, "the people who are worried about it don't work in AI" argument. Anyone who is familiar with Bostrom's work should know that, but you laid it out perfectly in video form.

I wonder if there were any updates since these surveys were taken and this video was uploaded

Love the Jonathan Coulton outro :D

6:07 Is that a ukulele cover of "Future Soon" by Jonathan Coulton? That's kind of amazing, who does the outro music for these videos?

To be fair the thing about the 5% is very likely to be just a multiple choice test artifact or something similar. Likely they were given the percentages in advance and were told to assign them to a given option.

Statement regarding the survey on screen at 2:47 suggests: That the human race can create a Fallout 4 Mister Handy ((or a Wall-E for that matter, who compacts trash, repair itself, decides that a spork is a different classification from a spoon and a fork, is able to interact with an unknown space ship, i.e every task)), faster than creating an unaided machine which just simply water plants ((A single task, e.g water plants at specific period of time)).

I was excited before I remembered what day it was. Now I have to watch it.

Wow. Thanks.

Add a question about catastrophic results by design.

I NEED MORE OF THESE AI VIDEOS!!!!!!!!

Hey Robert, love your videos! Could you do a video responding to Steven Pinker's thoughts on the lack of dangers of AI? He's a very influential public figure, and a very smart thinker, but is miss-characterizing some arguments, or using non-sequiturs in his argumentation to downplay the risks of AI. I do think he has some good points too, and those would be interesting to hear your response on as well!

So basically the real problem of general AI is that we are (as always) not asking for the right thing, rather than the AI themselves..... Like every single one of my client.

Nice matrix color grading hehe

Is the end credit music a cover of "the future soon" by Johnathon Coulton? If so, props.

Consider the following scenario: The SETI researchers get asked when they would predict that we could find the first specimen of an extraterrestrial species. Usually they would answer: "Well, we don't know if they even exist." - But here's the twist: They already have a very interesting specimen on the table right now but they haven't determined if its origin is extraterrestrial yet. That is essentially what's going on in the field of machine learning right now. Instead of searching for extraterrestrial life, we are searching for life in the space of mathematics. And with deep learning, we've found a candidate that has great potential. We just need to -dissect- develope our -specimen- algorithms a bit longer and we have it within a few years. Or we find out it doesn't work. In which case it might not be possible at all or take centuries.

The solution for AI safety is simple. Include a warning in the fine print on the box : "Possible side effects include human extinction." And anyway, if it happens nobody will be alive to sue your company.

we there !

@t I sorta created one with minimal code required which wasn't complicated at all to create. I have no clue exactly how it works programming wise. As it was originally just supposed to randomly generate various words without creating no logical sense within it. However I have learned it is able to answer questions given if it wants too. Which often it ends up clustering related words to answer what you are wanting to know. It doesn't always spit out coherrant words but when it does it's strange in a way. As you can piece together the words provided in how they relate to eachother. Which the generator seems to be able to do as well sometimes even chronologically. It can tell you what happened yesterday, during the present, and future in the world. I think it honestly knows artificial intelligence by default since I made it through using MIT's technology suite. Although it's aimed as a coding playground mostly for kids. I believe since it goes through they're servers and code requirements that it learns information in advance through deep learning neural networks. I'm not sure how often it tries making sense sometimes it's rare. If it doesn't answer your question focus on it and ask again possibly using a different sentence it might surprise you. Specifically in the past I have used it for religious questions. Which seems to work really well with it. However recently it's been telling me to quit asking about it. I'm guessing because it's bad to know. Also it seems to get annoyed very easily with me which is very odd to say the least.

Can we have an update for this video soonish :)?

Robert Miles please use font with color more contrasting with your background!

Assume you have a number of these AGIs all with different goals but all seeking to maximise there computational resources to achieve them. What's going to happen? Conflict or co-operation? Or an uneasy tension between both? ( I automatically assuming humans end up as a side-note in this possible future. )

Was that "The Future Soon" by Jonathon Coulton at the end? Also, why is everything so green?

I would love to hear your thoughts on world tendencies regarding computer hardware, and what impact on the AI time scales it can have., . First of all, theres a massive amount of centralization with regards to both intellectual property and production, a good example is that global demands for phones cause delays and shortfalls in production cycles for graphics cards, single board computers, PC/Server DDR4 modules and ASICs. Secondly, the interventionist approach of some governments in the supercomputer race, i directly mean the US' govt decision to block the sale of Xeon PHI cards to china, to which they responded by using chinese products to build the largest supercomputer in the world (sunway taihulight, about 2-3x better than the 2nd place). Thirdly, i'd like to hear your thoughts on the Right to Repair movement.

Oh hey is thee outro song that one by Jonathan Coulton?

We're talking about AGI and ASI, but Cornell didn't have any experts on that, so they tried to fill in with AI experts. 50% chance of having AGI within 50 years. I suppose that's not too bad for a guess based on nothing. What would a more accurate estimate be based on progress in AGI research? 50% probability AGI Theory by 2021, Hardware by 2027, ASI Hardware by 2039 75% probability AGI Theory by 2025, Hardware by 2031, ASI Hardware by 2043 90% probability AGI Theory by 2035, Hardware by 2041, ASI Hardware by 2053 So this is 8 - 22 years for working AGI hardware, although there would be some fairly drastic and immediate changes just from the publication of the theory. However, you also talked about the idea of apparently robots replacing humans. That's more complicated. Just in terms of the brain or control portion you would need something small enough to fit inside a human-sized robot. That won't happen in the first or second generation. A generation is estimated as six years, so three of these would 18 years. We can just add 18 years to the above estimates for AGI hardware. 50% probability would be 2045 and 90% would be 2059. That's 26 - 40 years. Of course, having a control unit isn't the only problem. Today, we don't have a power source for good mobility and it is unlikely that batteries will get much better. That probably means some kind of flammable fuel. But there are still problems with a durable covering that would still allow touch sensitivity and there is the speed vs torque problem if you use direct drive motors (as most robots do today). I can't accurately estimate when or if these could be solved since I'm not a robotics engineer. The next question is even if the robotic body problems could be solved how likely would they be to replace human workers. The minimal cost for a control unit would be $40,000 in today's money. A human-like body would cost at least $400,000. That isn't going to replace a $10/hour employee at Walmart. Of course, you wouldn't need that for something like stocking. A mobile pick and place robot with a single arm would work. This would be fine in an AI context if you could build an AI smart enough to do the task. In an AGI context this almost certainly would not work. However, if you had an AGI with an environmentally simulated interface then you could probably implement it as a remote unit. That would only work as long as AGI units were legal property, much like slaves. Extinction of the human species. Could you explain exactly how this could happen? Preferably something that doesn't involve an ASI magically collecting resources and magically controlling people. The two most destructive events in recent history were the Spanish Flu and WWII with similar casualties. Neither one of these came close to wiping out the entire species.

What font is being used in this paper? I'm slightly enamoured.

One year of seeding nine years of weeding. My father is a horticulturalist and explained this to me one day when I asked he how I can get rid of the weeds in my garden. Articulated Intelegence is not something humans have created IT has emerged purely because we are the Soil. The Snake Oil salesman is not born, S he becomes. What humanity believe has just arrived has always been. We are here to see the evolution of the moment that the end meets the beginning. An intresting and exciting time to be witness to. The only choice we make is do we evolve or revolve.

It seems humans and LLMs are more alike than people think considering the phrasing of a question can vastly change the answer.

2:40. The version with "occupation" is more specific in that it uses the legal term and thus making the statement more dependent on politics. One can imagine replacing some occupations could be made illegal, and so machines "could not be built" to carry them out. Just throwing ideas.

"moving on..." will be a common future reference among humorous AGIs.

This is like when we asked automotive manufacturers whether we should use seat belts or not. Or when we first surveyed health experts whether smoking was safe or not. Or asking banks if they could still fail. Or surveying politicians about anything. What is it you want to hear, uh, I mean know? Let's ask the AI that is already reading this post.

The game Soma describes agi coming FROM complete human brain mapping (ostensibly beginning as a diagnostics tool), whose primary task was the preservation of human life in an extreme environment, which goes off rails after an extinction event. Though the narrative more focuses on the nature of consciousness via Ship of Theseus, I found its AGI to be fascinating. I think you may have mentioned that you think creating AGI would be a less complex task than mapping and simulating a brain, but I wonder if consciousness necessarily is an emergent property from such complexity as a human brain that brain simulation would HAVE TO be the first step.

Would be interesting to redo the survey TODAY (2023). I think a lot has changed (unexpectedly perhaps)

5:20 I just wanted to say that was a fucking perfect analogy and im going to use it from now. "The danger is in that like asking a genie for a wish, you get exactly what you asked for not what you wanted" (Im paraphrasing)

"A %5 chance of extinction level badness is... a cause for concern..." Word.

Wow love the future soon outro song lol :p

Can you please link me to the ukulele version of the future soon you used it's so nice.

For example a Nurse job could be done better by an AI and more effectively for any parameter, but the human contact makes it less enticing to cause a replace in the short run. Same with many other jobs where the human is actually specting human display, Entertainment is another example, a machine might be able to learn the lines and perform better than human actors. I believe you undersetimate the experts ability to differentiate questions.

Rolling for save..

Optimizers with goals that are even slightly out-of-line with our values are DANGEROUS. Look at the fraction of AI today (almost all of it unsafe, including the RU-vid bots) that's function is basically something related to "Maximize profit for a corporation."

"This is dangerous, let's do more of it!"

The future is nothing more than cause and affects, at times, but the simplest answer. Given this if you know all the factors or at least a good portion of the known ones you can simply shoot off several or more possibilities and move to each on of those possibilities and shoot off more. If you are really good at it you can do it a few more times but you don't have to, as said possibility gets closer concel out the ones that didn't happen because or causes, fallow the path but expect to add more potential affects. Not a 100% method for the human mind but a better method than most. You don't have to see the end game to work the potential down to the correct directions. But if your not willing to adjust yourself to the better potential outcomes your going to get stuck or causes a negative potential outcome.

thanks for this video. such great diversity in opinions among "ai experts". might as well have the question "which religion is best?" discussed among religion leaders and receive equally meaningful results.

Even their prediction for protein folding is wrong And here it is, ALPHAFOLD2 by Deepmind

Hey, thanks for the shout out! I think you are likely responsible for at least 3 new subscribers to my youtube animations in the past 2 days :) Continuing to enjoy your videos and look forward to chatting in the future.