Rise of AI Conference 2018 · Berlin

The Good, the Bad and the Ugly

Daniel Jeffries · SciFi author and futurist

At Rise of AI 2018 in Berlin, science-fiction author and futurist Daniel Jeffries argues that despite loving the concept in his own fiction, he does not believe the technological singularity is near, because no one currently knows how to build artificial general intelligence or abstract reasoning. He traces AI's history as decades of failure and "winters," credits Geoffrey Hinton with keeping neural net research alive, and explains that people predict the future badly because they project a technology's current, flawed iteration forward instead of picturing a category maturing over time. He calls sentencing algorithms in US courts an example of the "ugly" side of AI, since black-box systems trained on biased historical data reproduce the same injustices against poor and disadvantaged people. He argues the more realistic future is not artificial general intelligence but narrow, "semi-alien" intelligence augmenting humans, illustrated by the MIT-covered "Rise of the Centaurs" chess result, and applies this to medical monitoring, smart homes and self-driving cars, cautioning that these benefits will roll out more slowly than people expect while automation, as in past eras, will create new jobs rather than eliminate work overall.

Key takeaways

  • Jeffries argues no one currently knows how to build artificial general intelligence or abstract reasoning, making the singularity unlikely in the near term.
  • He says people predict technology poorly because they extrapolate a flawed current iteration instead of thinking in categories that mature over successive versions.
  • He calls US sentencing algorithms an example of AI's ugly side, since systems trained on biased legal history reproduce injustice against poor and disadvantaged people.
  • He points to the MIT-covered "Rise of the Centaurs" chess result, where amateur players plus AI beat AI-only teams, as a model for human-AI collaboration.
  • He argues automation has repeatedly eliminated most existing jobs throughout history while creating new ones, and expects the same pattern with AI.

Detailed notes

Why the singularity is fantasy

Jeffries says that despite writing the singularity into his own fiction, he thinks Ray Kurzweil and others will be disappointed at the end of their lives, because no one currently knows how to build artificial general intelligence or how to do abstract reasoning. He describes AI's history as roughly 50 years of failure and repeated "AI winters," during which most researchers gave up and Geoffrey Hinton was kept going largely by a single university grant, until GPUs applied to existing techniques drove the deep learning breakthroughs of 2012, such as surpassing humans on ImageNet-style perceptual tasks.

How to actually predict the future

He argues most people are bad at forecasting because they extrapolate a technology's current, flawed iteration (citing early home video and CD-ROM books) instead of thinking in categories that iterate over time, such as VHS to laserdisc to DVD to Blu-ray. He attributes this partly to humans having lived as hunter-gatherers for 1.8 million years versus only 12,000 years of agriculture, leaving minds focused on immediate survival rather than long horizons. He says one can responsibly predict the general characteristics a mature technology needs, such as good screen, portability and battery life for what became the Kindle or iPad, but not unpredictable "black swan" inventions.

Abstraction, automation and the limits of narrow AI

Jeffries frames societal progress as a history of abstraction and automation, from the hammer and nail enabling multi-story buildings onward, and says today's AI systems are precursors to future intelligent machines but still cannot do abstract reasoning or generalize the way humans do, citing failed attempts at hand-coded symbolic rule systems that collapse into a "towering inferno of brittle rules."

Sentencing algorithms and the ugly side

He identifies sentencing algorithms already used in many US court systems, black-box statistical tools that recommend bail or jail decisions, as a clear example of AI's "ugly" side, arguing that a system trained on the history of the legal system will reproduce its history of injustice against poor and disadvantaged people. He blames the humans behind the data and design, not the machines, and says the future is neither purely good nor bad but exists in a gray area.

Rise of the centaurs

Jeffries argues the near-term future is not artificial general intelligence but "semi-alien" narrow intelligence assisting humans at tasks they do poorly, such as politicians setting budgets inefficiently versus an AI analyzing sensor data on failing roads. He cites the MIT-covered "Rise of the Centaurs" story, in which, after Deep Blue beat Garry Kasparov and IBM declined a rematch, a subsequent centaur chess tournament was won not by a grandmaster but by amateur players teamed with AI, beating AI-only entrants. He argues automation has already eliminated roughly 95% of past jobs multiple times in history while creating new ones, pointing to how few people today work in agriculture or as hunter-gatherers, and how women's job options expanded far beyond the three available in the 1950s.

Where AI will help: medicine, smart homes, self-driving cars

He describes plausible near-future benefits: a smartphone that flags a suspicious skin spot to help doctors triage urgency, a wearable tracking blood markers and nudging behavior before a heart attack, and a smart home that identifies a real fire versus a dead alarm battery and can call the fire department with context. He criticizes media coverage of self-driving car crashes as misleading given that 1.25 million people die in car wrecks yearly, but acknowledges real limits, noting Yann LeCun's self-driving car effort could not yet interpret road workers' hand signals. He predicts all of this will unfold more slowly than people expect, that neural nets will hit diminishing returns and possibly another, smaller AI winter, but that society will keep creating new kinds of jobs as it has throughout history.

Quotes

The history is a long litany of failure, actually, for about 50 years.

Daniel Jeffries

I actually worry more about narrow intelligence controlled by dumb, ignorant, violent human beings than I worry about a machine rising up and deciding to kill us all.

Daniel Jeffries

The future is not set in stone. It is also not binary.

Daniel Jeffries

Questions answered in this talk

Why does Daniel Jeffries think the technological singularity is not near?

According to Daniel Jeffries at Rise of AI 2018, no one currently knows how to build artificial general intelligence or how to do abstract reasoning, so despite using the singularity in his own fiction, he expects it to remain unrealized for the foreseeable future.

What does Daniel Jeffries say is dangerous about sentencing algorithms?

According to Daniel Jeffries at Rise of AI 2018, black-box sentencing algorithms already used in US courts are trained on historical legal data, so they reproduce the same injustices against poor and disadvantaged people that the legal system has always produced.

What does the "Rise of the Centaurs" chess story illustrate about AI and human work?

According to Daniel Jeffries at Rise of AI 2018, in a chess tournament the year after Deep Blue beat Garry Kasparov, amateur human players paired with AI beat AI-only entrants, showing that human-AI collaboration can outperform either humans or AI alone.

Mentioned

a"": ""According to Daniel Jeffries at Rise of AI 2018

Keep exploring

More from Daniel Jeffries

More from Rise of AI Conference 2018

More on AGI & Future of Intelligence

Transcript

Auto-generated transcript, not edited. · 4,257 words

Show the full transcript

so I guess the real question is where is all this stuff going we'll talk a bit about it over the next 10 to 20 years I'm a little bit down on the singularity as a sci-fi writer I love the concept I put it into a lot of my fiction but the truth is I think Ray Kurzweil and a lot of people are gonna have an unfortunate surprise towards the end of their lives when they're not immortal I don't think we're actually going to get there in the near future and that's because we'll do some predictions here predictions are basically a tricky business so I'm going to be careful but what you know what does all this look like in the next 10 to 20 years the fact is we actually have no idea how to build an artificial general intelligence sorry but it's just the case we we actually don't know how we do what we do we don't know how to abstract things what we have now are a lot of mimics of intelligence

we have a lot of narrow intelligence that gives the illusion that it is intelligent but it's intelligent really in no way in fact it's kind of stupid if we think about the history of artificial intelligence the history is a long litany of failure actually for about 50 years sorry guys but it's the it's the case and for many years we weren't actually able to do almost anything with artificial intelligence it was generally a toy if you know the history about it we went through a series of AI winters where people just basically gave up on the technology Geoffrey Hinton who's now you know one of the godfathers of neural nets one university kept neural net research alive that with a grant and it was basically everybody working in this space was kind of a pariah they told you don't study this stuff it's never gonna work okay and and that's because it's actually in Applied Science this was something where you you can't just do the work with a pen and paper like you can

with physics you actually need processing power and data and things like that if you really think about where a lot of the breakthroughs came with deep learning in 2012 was basically applying GPUs to something we kind of already understood and we started to get better at some of these perceptual tasks like image net to the point where we have gotten better at humans so how do we see the future in sort of an effective way and I think actually most people are really terrible at it and then the reason they're terrible at it is because what they do is they look at a current iteration of the technology and they project that indefinitely into the future instead of looking at a category of technology so a good example that I always use is home video recording is a category of Technology Betamax VHS or iterations of Technology and so people will look at the current iteration they go look at all the flaws this is terrible to look at doesn't record very much this is

how it's always going to work and the reason is because they're not able to then iterate through VHS laserdisc DVD blu-ray DVR in their mind they can't see those things happening the reason is because I think mostly our minds are still living in the forest we were hunter-gatherers for 1.8 million years agricultural revolution has actually only been around for 12,000 years scientific revolution 500 Industrial Revolution 200 right and when you're a hunter-gatherer you're you are focused on the now at all times is a tiger going to eat me where is my next meal where am i sleeping in order to argue that the vast majority of people still live within that framework it's very difficult to see the future right I would say that entrepreneurs futurists are kind of the shaman and the medicine man right love of the global tribe right looking into the future and what does it hold I think that we can actually predict some things about the future by generalizing by doing the one thing that AI can't do and

that is abstracting and projecting out those things so we can say take something like the cd-rom book which came out in the 80s had a number of advantages over a normal book you could copy it you had color you know it was it was a very interesting technology it didn't take off why didn't it take off well you could scratch the disk the screen was horrible to look at it was a big computer you're not gonna lug that around I'm gonna read my book on the subway with my mind right so eventually that actually is in the Kurzweil book that's sort of a false pretender stage where that initial technology fails what we can do though is look at all the flaws of a current technology and say here is what the characteristics of a future technology that works well look like so we can say okay we need to have an excellent screen it would need to be portable it would need to hide away the medium of storage for us we need to

have a long battery life we'd need to be hard to destroy now I've basically just talked about what is a Kindle or an iPad so you can predict the general characteristics but you cannot predict are sort of Black Swan events you can't predict somebody who invents something that just doesn't exist and one of my earlier sessions I essentially said you cannot explain to an 18th century farmer what a graphic designer is and that's because a graphic designer is the end result of a series of steps all of the technology that exists today is a series of steps built on other pieces of technology okay so let's talk about some of the things that are going to happen Society moves forward by automating things and abstracting things so today we sort of fear some of these things but it's not to be feared when I when we created a hammer and a nail that is a form of abstraction and automation it allows us instead of building mud huts we can now hammer things together hold

them in a consistent fashion now I can build more complex structures without that I cannot build multi-story structures until that original precursor technology exists so what we have today are a bunch of free cursors - future intelligence machines but we do not understand how to do abstract reasoning in fact if we when we've tried some of these symbolic reasoning ideas when we try to come up with all the rules we go well we'll just put all the rules into a system a bird is an animal that flies and has wings except when it doesn't have any of those things that it's still a bird and so what you end up with this is towering inferno of brittle rules that doesn't work we also don't know how to ask the correct questions about certain things right we don't know how to actually get to a point where we say where we can generalize how do I look at something that cuts me generalize that to something sharp jagged so the next time I see a rock

that's jagged I understand that that cuts me just like a knife would right so we're able to abstract and reason in a way that computers are not able to so what I see is a lot of these narrow intelligences augmenting what we do we're going to talk about some of the fantasy of of the future we're going to talk about some of the things actually that I think will actually happen and I need some water the singularity is fantasy number one it requires a huge leap in processing technology and an understanding that we just don't have it I don't know where it comes from maybe it emerges out of nowhere but I I can't foresee it that would be an absolute Black Swan event the other thing is things like you know terminators are this sort of mass consciousness this falls into the ugly side for me I actually worry more about narrow intelligence controlled by dumb ignorant violent human beings then I worry about a machine rising up and deciding to kill us all

we don't need any help being jerk humans did not get to be the top species on the planet by being nice in what we were really worried about when we worry about artificial intelligence as we worry about AI making ethical decisions if you really look at where we fear it the most we fear it in areas where it makes an ethical decision so sentencing algorithms are another example of the ugly for me these actually already are at work in many court systems in the United States where and people don't know this it is a black box algorithm that spits out a number of whether somebody should get bail or go to jail or not those people don't know this this is a terrifying horrifying prospect in many respects if you take a narrow intelligence today and you put it on which is nothing really but a data gathering system right a box of Statistics DARPA calls neural nets spreadsheets on steroids that's all that they are if you have it go back and study the

history of most legal systems what you are going to find is the history of injustice in the very same people that we have always put in jail poor and the disadvantaged are the people who are going to continue go to jail they are going to ingest the horrific ethics that humans have now I don't blame them the machines for this I blame the humans we have to ask better questions the future is not set in stone it is also not binary this is the other thing that we start to see in the popular press it's either good or bad there's nothing in between and the fact is life is in between the beauty of life is in between the gray area is where we all exist in life so I actually see us not needing to create artificial general intelligence which is fine but actually creating semi alien intelligence now what do I mean by that I mean that we are creating we are creating narrow a eyes that it can assist us at the

things that we do terribly one example I gave earlier is politicians setting budgets the politicians set budgets they do a terrible job at it they do a terrible job in it because they are trading favors with each other and so what you have is a tremendous amount of waste and inefficiency of what should be a simple calculation where are my roads failing if I have a sensor network of various systems I can understand which roads are breaking down if I'm using satellites to look down upon it with artificial intelligence I understand where the money could be allocated I can probably do a better job with budgeting than politicians ever could in addition I see the AI is working with us there was a great article out of MIT called rise of the centaurs and in that article title comes from a very interesting thing that most people don't know about after deep blue.the chess playing at AI beat Garry Kasparov Garry wanted to play the system again in IBM said sorry we're packing up our

stuff and going home we already beat you so he sponsored a contest called centaur chess the next year where anybody can enter a human and artificial intelligence or an artificial intelligence and human team the team at one was a human and AI working together which beat the pants off of an AI and interestingly enough it was not a grandmaster human it was a couple of regular amateur chess players working with an artificial intelligence so when you think about artificial intelligence as an augment to our to our lives you can begin to see how we're not necessarily going to destroy all the jobs I think this is another fallacy we've already destroyed 95 percent of the jobs multiple times in history nobody here is a hunter-gatherer I don't think just mentally nobody here is is still in it most people there's three percent of the people who are in agriculture it's ok we every time we destroy all the jobs it's not that there's no downside to some of these things short-term pain or disruption every

time destroy all the jobs we create new and varied jobs if you're a woman in the 1950s you had a choice of three jobs secretary homemaker teacher now women can do all kinds of things we have a brand we have more jobs than we have ever had in the future artificial intelligence it's going to accelerate and create new types of jobs that we cannot imagine imagine for instance I have a call center and I if you've ever called a call center it's a painful experience right you're calling up someone who's not generally paid to think critically who generally lists all the reasons they cannot help you in a very Road fashion every once in a while you get someone on the phone who can think clearly and it's so exciting to you I think my gosh this person can actually do my return and they understand what the problem is thank you so much can I please talk to your manager to say that you deserve you know and raise what if we could model

some of the better decision-making processes of the best people on the team and create a system that allowed people to learn from those other people right so that they I could not just along the way level up the entire quality of the team so that more of the people effectively are thinking critically because the machine is helping them to do it when I think about things like medical technology I gave an example earlier that's more advanced but a simpler one is this imagine I wake up and I have a spot somewhere and I'm worried about this spot what do I do now well what I do now is I call the doctor and I say I've got this spot I'm worried about the doctor says that's great I can see you in a month and a half sorry you go no no but I'm really worried about this right now and the doctor has no way to differentiate whether I actually have a spot that needs to be looked at right now right or something

that can wait however in the future I just point my smartphone at that spot and it gives me an indicator that says you better call the doctor it could be could be malignant it's good there's a good chance I in the future bill to send that report to the doctor's office the doctor will now be able to make better decisions about triage and scheduling they'll say ok this is actually a priority this person should I'm in the other person who called me with hypochondria can wait at home right so we'll be able to make better decisions in the future I might have a watch that has a little hairlike pinprick that's constantly taking my my blood it's giving me a sense of triglycerides cholesterol maybe I'm not working out as much maybe I'm deluding myself into thinking I'm working out as much for them getting more of a ponch but the AI could then begin to nudge you the more vegetables do things now we're humans so we might avoid that and so the AI

might actually be able to predict that we're going to have a heart attack but instead of having the heart attack and collapsing into my oatmeal then hoping that the paramedics get there in time the AI might say by the way the next day you're probably gonna have a heart attack you should probably call the automated ambulance the self-driving ambulance to come pick your ass up right now okay so these types of things are going to be tremendously beneficial in terms of medical technology in our ability to get the types of care that we need right in the field and it's only going to continue to progress also think about something like a smart home when I have a fire alarm go off in the middle of night it's just this annoying screech I have no idea whether it's true or mostly it's just the battery's dead if I have a smart home it might say hey by the way there is actually a fire there's a voice telling me that there is a fire in

this room and I should probably avoid going out that way and I should get the kids and the things that I care about and go out the back door as quickly as possible now think about something like the Google assistant that they just demonstrated instead of making hair appointments now the house is actually calling the fire department and it has contextual awareness of this situation where the house is burning what type of fire it is is it a grease fire was there an explosion of some kind did somebody fall asleep with a cigarette it will have contextual information which theoretically could then be given to the first responders or the fire department and now they arrive on the scene with additional understanding and information that they didn't have before they're not walking into a fire blind right so these are the types of benefits that we're going to see with artificial intelligence some of this is going to take time for the infrastructure to roll out a lot of the stuff that we're seeing here

is still very much in the research stage I would say in some respects don't necessarily believe the short-term hype do believe the long-term vision we don't really understand all the things we can do with artificial intelligence but we're going up a sigmoid curve and at some point we we hit a point of diminishing returns where it doesn't live up to our expectations right and potentially we have another winter I think it would be smaller at this point in time but eventually we're gonna run up against what neural nets in our in our current understanding can do I've already run up against it I tried to create something to name one of my novels and I fed it all the coolest novel titles in history and it gave me a bunch of gibberish right and you know it some people get very excited about if I fed it a bunch of novels it could generate a novel it's not generating a novel anytime soon I'm not gonna my job's not in any trouble anytime soon what

it is generating is the format of a screenplay or the format of a novel but it's not saying anything it has no contextual understanding and other types of things that we that are going to be exciting when you think about sort of self-driving cars right now the popular press is all about every time there's a crash right there it's like they're excited about the crash it's disgusting right that's what they're reporting on and they're reporting on it in a misleading way every article title now is Tesla with autopilot crashes that's like saying man with legs runs they're not accurately reporting that the Tesla had its autopilot on what they are doing is deliberately misleading you to create fear all right here's here's some real fear you have a better chance of dying in a car wreck tomorrow than just about anything else in your life 1.25 million people die of car wrecks every year the AIS are going to be much better at that much better at survivability over the long term when in the

future then they ask people what it was like in the past when you tell them that you drove a car they'll look at you like what are you talking about it'll only be a specialized thing for like a sports car or or something that you own because you really wanted to drive and they'll be shocked that that many people to hide in that that was acceptable to society at the same time there are many things that self-driving cars cannot do at this point Yan Laocoon just came out with a self-driving car company said it can't understand hand signals of road workers and there's no way to teach it to do that yet this is one of the best AI Minds working today and he has no solution to this problem so they're using a limited pilot set in a small town where the workers are aware that it is happening because they are going to have to communicate with the cars in a different way and as a society think we will find mitigations for

some of these things over time in addition we're going to have to find how to give a higher level of meaning to the artificial intelligence when an AI crashes today any obstruction is exactly the same to the AI a piece of plastic debris a person or a divider but a human knows that that plastic thing if I hit it it's probably not good but if I hit a human there are serious consequences if I hit that wall there I am going to die right so we are gonna have to get better at teaching computers to understand meaning have abstract reasoning all this will happen much slower than people imagine right when you think about how long thinks to develop people always overestimate the speed of these things society is speeding up things are changing they are not changing so fast that we are going to be living in a in a god-like AI Network where we're all virtualized in the next 10 years it's not happening now if it if it does happen I'll happily

be there with all of you guys living in this virtual paradise but it ain't going to happen I'll bet my career on this one so we don't have to worry about that we also don't have to worry about it destroying every single job that exists we've already again destroyed all the jobs and many instances in the past and we progress a society by abstracting automating certain types of jobs people become very attached to doing certain types of jobs almost in a perverse Sisyphean way that are not necessary when I was a Systems Administrator I took a perverse pleasure in how long I could stand there and take the pain of standing at a keyboard in a cold data center for 14 hours drinking Red Bull it's insane there are all because I took perverse pride in doing these repetitive tasks when my energy and my intelligence could have been spent better elsewhere if much of that has been automated and we're already seeing that in the tech industry right where you're seeing DevOps and these

sort of immutable systems being developed and that allows people to move higher up the value chain that is the history of mankind when everybody doesn't have to be a farmer somebody can be a lawyer somebody can be engineered somebody can be a venture capitalist right and and that is what specialization brings we're gonna see an increasing it's so acceleration of that type of specialization and it is not a bad thing it is not a thing to figure it as a thing to embrace and it is going to make our lives better there are definitely things that are not good we cannot control all the things that an AI will do we will not stop governments from creating killing machines it will not happen it doesn't matter if we create a treaty where every single country on earth signs up to it I guarantee you I will bet anyone here a million dollars that they will build it anyway with black budgets and they will do it secretly anyway it doesn't matter so there are certain

things that we can't stop that are bad but there are so many things that are positive that to me the future looks very very bright I don't worry about these kinds of doomsday scenarios humans love to come up with horrifying doomsday scenarios they almost never work out that way we are wired to see big flashy threats not important threats that play out over a short period of time or a longer period of time we will be fine humans are very creative and intuitive we will invent new jobs and the economy will move faster and faster and life is gonna be very exciting over the next 10 to 20 years thank you everybody

Rise of AI 2027 · Berlin · May 2027

Be in the room for the next one.

Ten years of talks started as conversations in one curated room in Berlin. The next edition brings 300 selected leaders together in May 2027.