> After decades of investment, oversight, and standards development, we are not closer to total situational awareness through a computerized brain than we were in the 1970s.
Hard to see how that could be true. In just about any field, computers today provide much better situational awareness than was possible in 1970.
The article makes the usual complaints about self driving cars:
> Despite $16 billion in investment from the heavy hitters of Silicon Valley, we are decades away from self-driving cars.
Yet cars are much more intelligent today than they were in the 1970s. And we are not decades away from self-driving cars - Waymo runs self driving cars today in very specific locations.
FYI, it was written by a woman. I looked up her book, Kill It with Fire, and as a mainframer I have to say it seems pretty interesting.
I think what she alludes to in this essay, though, is more like that AI cannot solve socioeconomic problems of humans. And even humans seem to struggle with it.
Whenever I read stories where the metrics became the targets, and the like, I am reminded of Varoufakis' book Economic Indeterminacy. He doesn't give any answers there, but there is this "strange loop" in rationalism that nobody really understands.
I also think that AI might be a wrong target, because you need to understand the problem before you can solve it, and once humans understand the problem, they don't need AI anymore, they just code the solution as an algorithm. On the other hand, if humans don't fully understand the problem, it's extremely difficult (except artificial circumstances like games) to explain to AI what the problem is, so it would arrive at a "reasonable" solution (and avoided, at the very least, killing all humans).
If we can't understand the problem, will we be able to understand the solution presented by that AI? Or we'd just apply it, trusting blindly the unfathomable reasons the AI used? Do we have an AI where the decision tree can be grasped by humans?
It seems inevitable to me the moment AI capacity seriously surpasses human (as a whole) capacity in any specific topic, it becomes an oracle.
I hear of efforts to translate machine decision making to human understandable terms, but if it is a question of raw intelligence, it will quickly become impossible to understand.
It's interesting to see how super advanced ML-based engines have changed the field of chess in recent years - top engines are way better than top humans, so, in a sense, act as oracles.
Top players routinely use them in preparation, to study new lines, get hints about what moves make sense in a position and also to generate new ideas or tweak the principles they apply in the game. The engines don't explain their reasoning, but provide something closer to the "correct" move in any given position. It's up to the humans to do the legwork and understand _why_ the recommended move is strong.
Clearly, chess is not real life, but the impact of these oracle engines has been broadly positive (with the exception of using engines to cheat in online play).
This is a good example, because Chess rules do not surpass human intellectual capacity. With enough time, you can understand the reasoning of an AI.
But when it comes to raw intellectual superiority, say, explain the logic behind new mathematical or physical concepts and discoveries, very high level predictions, we come to our human limit and cannot surpass it.
The AI can give us implants to improve our intelligence, but only to a certain limited degree, our biological brain will slow us down significantly.
That is the barrier we cannot possibly cross in a timespan of human life.
And to be honest, I wondered about that with GPT-3. Maybe it could give a more profound answer to a given prompt, but it chose not to, since it found the prompt to be too silly, and it responded in kind. Just like adults are able to entertain an imaginary universe of children.
So even to explain that we want a "serious" answer might be difficult.
How much would you bet that human interviewers' opinions of a candidate after a video interview wouldn't be affected if they were visibly in a room full of books? Or if they wore glasses, or had a painting hanging on the wall, or the various other things the researchers found made a difference to the AI's assessment?
To be clear, I am super-skeptical about the ability of AI systems to do a good job of judging an interviewee's personality from a short video clip. But (1) this seems obviously to be a really hard problem, and one that couldn't even have been attempted in 1970, and (2) I am also pretty skeptical about the ability of human interviewers to do it.
> "But (1) this seems obviously to be a really hard problem, and one that couldn't even have been attempted in 1970"
Apollo 11 landed on the moon July 20th 1969. You think that didn't take a degree of "AI" and "ML"? Or maybe we just had a different name for these things back then...
I think that's just a matter of 1. resources and 2. risk tolerance.
1960s USA had a less superficial political culture, so was more willing to tolerate PR-sensitive risks in pursuing major technological achievements, and it could allocate a much larger proportion of GDP to the Apollo program, as government directed social welfare spending consumed a far smaller share of GDP: https://fivethirtyeight.com/features/what-is-driving-growth-...
Its easy to point at the bookshelf example and say "Haha AI is stupid", but its actually quite impressive. One could easily argue that most human interviewers have similar bias, and that it can detect such complex signals (books, glasses etc) IS impressive.
The problem is this case is the data and/or wrong objectives, but the "AI" here has a lot of awareness, just on the "wrong" signals.
I don't see how this is relevant. Computers in the 1970 had no situational awareness about people interviewing for jobs. So yes, that software might be crap, it still has infinitely more awareness.
what even is "infinitely more awareness?" is that an actual metric? Computers today have exactly as much awareness as they had in the 1970s, situational or otherwise, which is none. The algorithm in question does not know what a bookshelf is, does not know what a job interview is and it does not know how the two relate. It correlates a bunch of pixels and creates the illusion of having awareness, but this is an anthropomorphization and nothing more.
> Yet cars are much more intelligent today than they were in the 1970s
Therein lies the problem. Your definition of intelligence presumes that it is a simple quantitative scale, measuring I guess, something like "system complexity".
The relevant sense here, in which no progress has been made, is qualitative -- ie., it is a distinct property. And this property has not been acquired.
What is the property? It is dynamical, not formal. It is more like gravity (, pregnancy) than it is like addition.
It is the ability many animals have of adaption in the shifting and challenging environments in which they are embedded.
That type of adaption is not formal: it is not adaption in the sense of "updating a weight parameter". Rather, of the cells of their bodies coordinating themselves differently, and thus of their tissues, and thus of their organs, and thus of their whole brain-body system. Both from a top-down command ("I want to run now, and so my cells...") and from a bottom-up ("my cells... so I ...").
What enables animals to be fully embedded in their physical environment, to cope and adapt to its radical shifts, is this capacity. The type of "crossword puzzle" "intelligence" we obsess with is entirely derivative of this more basic --and vastly more powerful -- intelligence.
Cognition is just a semi-formal process, parasitical on the body's intelligence; whose role is simply to notice when it fails and problem-solve it.
We have, at best, merely the architecture of this formal reasoning. But there is still nothing for it to reason about. And in this sense, computer science has made no progress -- and indeed, cannot. It is not a formal problem.
And yet computers continue to perform tasks that were talked about for years as something uniquely human / intelligence driven. This is a nice philosophical debate, but in practice I think it falls flat.
I dont see any single case of that. Rather in every case the goal posts were moved.
Can a computer play chess? No.
They search through many permutation of board states and in a very dumb way merely select the decision path that leads to a winning one.
That was never the challenge. The challenge was having them play chess; ie., no tricks, no shortcuts. Really evaluate the present board state, and actually choose a move.
And likewise everything else. A rock beats a child at finding the path to the bottom of a hill.
A rock "outperforms" the child. The challenge was never, literally, getting to the bottom of the hill: that's dumb. The challenge was matching the child's ability to do that anywhere via exploration, curiosity, planning, coordination, and everything else.
If you reduce intelligence to merely completing a highly specific task then there is always a shortcut, which uses no intelligence, to solving that task. The ability to build tools which use these shortcuts was never in doubt: we have done that for millenia.
> They search through many permutation of board states and in a very dumb way merely select the decision path that leads to a winning one.
> That was never the challenge. The challenge was having them play chess; ie., no tricks, no shortcuts. Really evaluate the present board state, and actually choose a move.
Uh-huh. And how exactly do you play chess? Do you not, perhaps, think about future states resultant from your next move?
Also, Alpha Zero, with its ability to do a tree search entirely removed, achieves an ELO score of greater than 3,000 in chess, which isn't even the intended design of the algorithm.
A rock will frequently fail to get the to bottom of a hill due to local minimums vs. global minimums. A child will too sometimes.
> Uh-huh. And how exactly do you play chess? Do you not, perhaps, think about future states resultant from your next move?
Not quite. You'd need to look into how people play chess. It has vastly more to do with present positioning and making high-quality evaluations of present board configuration.
> rock will frequently fail to get the to bottom of a hill due to local minimums
Indeed. And what is a system which merely falls into a dataset?
A NN is just a system for remembering a dataset and interpolating a line between its points.
If you replace a tree search with a database of billions of examples, are you actually solving the problem you were asked to solve?
Only if you thought the goal was literally to win the game; or to find the route to the bottom of the hill. That was never the challenge -- we all know there are shotcuts to merely winning.
Intelligence is in how you win, not that you have.
> Not quite. You'd need to look into how people play chess. It has vastly more to do with present positioning and making high-quality evaluations of present board configuration.
That is what Alpha Zero does when you remove tree search
> A NN is just a system for remembering a dataset and interpolating a line between its points.
Interpolating a line between points == making inferences on new situations based on past experience.
> If you replace a tree search with a database of billions of examples, are you actually solving the problem you were asked to solve?
The NN still performs well on positions it hasn't see before. It's not a database. The fact that the NN learned from billions of examples is irrelevant. Age limits aside, a human could have billions of examples of experience as well.
> A NN is just a system for remembering a dataset and interpolating a line between its points.
So are human brains. That is the very nature of how decisions are made.
> Only if you thought the goal was literally to win the game; or to find the route to the bottom of the hill. That was never the challenge
So then why did you bring it up as an example other than to move goal posts yet again? I can build a bot to explore new areas too. Probably better than humans can. Any novel perspective that a human brings, is, by definition, learned elsewhere, just like a bot.
> Intelligence is in how you win, not that you have.
Sure, and being a dumbass is in how you convince yourself you're superior when you lose every game. There are many open challenges in AI. Making systems better at learning quickly and generalizing context is a very hard problem. But at the same time, intellectual tasks are being not only automated, but vastly improved by AI in many areas. Moving goalposts on what was clearly thought labor in the past is just handwaving philosophy to blind yourself from something real and actively happening. The DOTA bots don't adapt to unfamiliar strategies by their opponents, and yet, they're still good at DOTA.
Let’s say that you have the ability to know the state of every neuron, and the interconnect map between them, at all times. You watch a chess player make a move, determine what is going on, and define the process the brain follows as an algorithm. Now that you have an algorithm, you have a very powerful piece of silicon execute the algorithm. Does that piece of silicon have intelligence? You would probably say no, since simply executing a pre-defined algorithm is a shortcut. Intelligence means the ability to develop the algorithm intrinsically in your head.
So fine, we take a step back. Instead of tracing all the neurons as they determine a chess move, we trace all the neurons as they start, from a baby, and learn to see and to understand spatial temporal behavior and as they understand other independent entities that can think like they do and as they learn chess and how to make a move. Then we encode all of that into algorithms and run it on silicon. Is that intelligence? To me, it sounds like it is just a shortcut - we figured out what a brain does, reduced it to algorithms, and ran those algorithms on a computer.
What if we go back further and replay evolution. Is that a shortcut?
To be fair, you did claim that the ability to adapt and make tools is what distinguishes real intelligence. But I wonder if ten years from now, you will saying that a tool making computer is just a shortcut.
I think intelligence is more generally how an agent optimizes to be successful, objectively and subjectively, across a wide variety of different situations.
> Can a computer play chess? No.
> They search through many permutation of board states and in a very dumb way merely select the decision path that leads to a winning one.
This is a perfect example of moving the goal posts. The objective was never to simulate a human playing chess.
If we understand something, we can describe it with an algorithm.
If algorithms for intelligence are by definition impossible, then understanding intelligence is by definition impossible.
So if true intelligence is something beyond our current understanding of the world (beyond algorithmic description). To me, this feels like god in the gaps applied to intelligence.
If a lookup table can predict human decisions with high accuracy given access to its senses and feelings, then either a human is just another can opener or intelligence isn't real.
The objective was to make a machine that could beat anybody at chess. Nobody on the Alpha Zero team believes Alpha Zero is an example of general AI. Teaching a system to understand a complex system is a necessary subcomponent of general intelligence.
And I think it looks more or less like, "organic-in-relevant-way or not".
"Relevant" here means a type of organic-physiological adaptability which is able to operate on incredibly short time scales: ie., you're able to physically adapt your body as your environment changes at the second, minute, hour, day, ... decade, timescales.
Typing is a type of organic adaption which takes years to complete. As are essentially all of our skills.
With digital machines and robots we're maybe able to emulate our cognitive processes as they consider our body-environment relationship. But we have not emulated, at all, our ability to have this type of body-environment relationship.
A dog's environment isn't roads. And no car has been design for a dog to drive.
Were such a car to exist, it is clear the dog would win in very very many environments (almost all). As would a mouse, let alone a dog.
That it may be possible to rig a human environment to be replete with so many symbols (road signs, etc.) that an incredibly dumb automated system can follow them is hardly here-nor-there.
Personally, I dont even think that will be possible. Self-driving cars may work on highways and motorways; I don't see there being any in cities. Not for centuries.
(Absent pretty big engineering projects to make cities so overly sign'd that a non-intelligent automated system could navigate them. Consider, eg., existing automated trains & train networks.)
> Were such a car to exist, it is clear the dog would win in very very many environments (almost all). As would a mouse, let alone a dog.
This seems incredibly unlikely. AI vastly outperforms 99.99% of humans on various video games, and 100% on many others. I'll bet on a well trained ml model over a dog every time.
> That it may be possible to rig a human environment to be replete with so many symbols (road signs, etc.) that an incredibly dumb automated system can follow them is hardly here-nor-there.
We already have above average human performance with just normal road signs, and could also simply use digital information.
> Self-driving cars may work on highways and motorways; I don't see there being any in cities. Not for centuries.
IMHO the biggest problem is the moral problem; even once the tech achieves better reliability in general (as compared to human drivers, who are quite crappy but we're all used to them), the cases when it will fail will be so spectacular and cause so much outrage because we're ill equipped to deal with situations where there is nobody to blame: we always try to find somebody to hold responsible. When there is none, we make them up (deities and whatnot).
When natural disaster strike, people feel plenty of emotions, including anger. Often that anger though cannot be directed to anybody in particular. ("God isn't easily sued
When machines misbehave, being by definition human made, it's harder to accept it "as just the way the world works".
Centuries is a long, long time in science and technology. It is 2021. If we take "centuries" to mean two centuries, railways with steam engines were not yet a thing in 1821. Cars without horses much less so.
None of us can predict the state of computer science in 2221.
A waymo self-driving car is not made of flesh and bones but I have yet to see a waymo self-driving car that performs better than a dog at: rescuing humans, detecting cancers and disease, shepherding and playing with other animals, protecting family from threats...
my point was just a crude attempt to dispel the myth that flesh has anything to do with efficiency at some task. What matters is design, and the immense research powers that nature has through eons of natural selection does an impressive job and doing things for which it has been "trained for" (hence your example).
You have a valid point here. But it might be that for practical reasons the progress in that direction won't be needed. Like, brute forcing it might be enough to reach a level higher than we can grasp. And if we can't grasp it, it's all Greek to us anyway...
What self-driving cars do is closer to something an animal like a deer does, not something a bacterium does. And you'd generally say a deer uses some intelligence while moving.
To answer your question. No. Single cells organisms can't drive cars. A multicellular organism can.
We are talking about self driving AI as a system - inputs and outputs and it's relative complexity. I mapped that system complexity to single cell motility.
Let’s say you and I were going to race each other by walking. You start on the east side of Los Angeles and I’m in Santa Monica.
Does your lead mean you’re in a much better position than me? What if the finish line is in Amsterdam?
That’s how I see AI (particularly self driving tech) today. Yes, technically there’s been advancements but we don’t even know whether it’s possible to get to the finish line today.
> Let’s say you and I were going to race each other by walking. You start on the east side of Los Angeles and I’m in Santa Monica. Does your lead mean you’re in a much better position than me? What if the finish line is in Amsterdam?
I can't figure out how to parse this. You refer to my starting location as a "lead", but they ask if it means I am in a better position - that is the definition of "lead". I think your point is that we are so far from what is needed that it is hard to know if we are even moving in the right direction.
Which is a weird argument. My brother drove me in his Nissan Rogue today, which does automatic lane following. You don't have to steer your car, or use the gas or brake for many driving conditions. That is unambiguously an improvement over full manual control.
It's an improvement in what it does now, but I think the point is it is not necessarily bringing the end goal closer. Like a side track that runs dead at some point.
No matter how much faster horses may have had become by selective breeding, that did not bring 100km/h travel closer.
With regards to autonomous driving, improvements like this absolutely bring us closer to the end goal.
Each improvement in adaptive cruise controls, lane following/holding assistants and any other partly-autonomous assistance system does its part in acquiring experience and technology for "the end goal".
You missed the word “much”. Is a lead of one step a much better position? If it’s a 2 step race, then for sure. If it’s a 2000 step race, almost for sure not. If we’re not sure how long of a race it is, then who can say.
If by “finish line” you mean super-human cognition in every single sense, then sure (although it is possible). That might be a century away while AI has nonetheless been stupendously successful in several areas.
This author makes broad sweeping claims, supprting them with numerous references that (in all instances that I checked) actually counter their argument. I'm not even sure the author knows _what subject they want to talk about_, never mind what argument to present.
I waded through half of it hoping a coherent point would emerge before heading over to hn comments to confirm my suspicions.It’s a mash up of a couple of different pop opinions on the state of ML without any real insight.
at least one of them is true. The effort spending "cleaning data" which I think by her description she means "connecting pipes" is underestimated. I work for a company that deploys an ML model that works on a "lowest common denominator" between multiple different downstream SAASes, and this week I have been struggling with a field that should be an integer, but is a string in one SAAS provider, (and there are entries that are not parsable as integer). I can't simply convert it to a string because the ML model is not expecting it.
Yes. "AI" (read Deep Learning/LogReg/SVM models) do indeed perform better given more data. I can vouch for this myself. And there was also a paper regarding this.
Amazon made $645 million net profit in 2008, $476 net profit in 2007, and $190 million in 2006.
Where did this myth of "Amazon doesn't make profits" come from? Why are people seemingly unable to check publicly shared historical 10k and fact-check themselves before making statements like this?
Yeah, the 2008 number is wrong, but the meme comes from earlier. Its first profitable quarter was Q4 2001, three years after IPO, and it's first profitable year was 2003, six years after IPO.[0] This seems like a long time, especially for the late nineties/early 2000s. (Though tbh, IPO three years after founding feels early to me too.)
Additionally, I seem to recall that they talked this up. Not "we're working on becoming profitable" but instead "We plan to continue losing money for several years. Deal with it."
I believe that prior to 2016, any profitable years were pretty much entirely thanks to Q4, and they were pretty small for its size[1]. A profitable year is good, but three quarters of losses each year will stand out. Sure, they're retail, but they're also tech. Sky high margins are expected year-round.
Amazon was unprofitable because they poured all their opearating profits into growth projects, not because they were subsidizing operations with investment.
Like many retailers, the business is seasonal and Q4 has more shopping. This is modeled as part of the business.
We don't say lawn care businesses are unstable because they do most work in the summer.
From what I remember, Amazon's strategy early on was to take as much revenue as it could and invest it back into itself. It intentionally ran in the red to try to grow faster.
Title aside (which is silly since AI is a toolset for solving a variety of problems), it is just so poorly written that it's not until more than halfway through it that I think I see it's main points (that present day A.I. systems are too dependenct on 'clean' data, and some nebulous discussion of how AI contributes to decision making in organizations). And the main point wrt data quality is rather silly in itself, because plenty of research is done on learning techniques that take into account adversaries or bad data. And all the discussion wrt how AI should be used to improve decision making is just super vague and makes it seem like the author has little understanding of what AI is and how it is actually used.
It is interesting that the author assumes that the intent of industrial applied AI is to make better decisions - from my experience, in the vast majority of cases companies are applying various techniques (both AI/ML and hard-coded heuristics) with the explicit intent to get cheaper decisions, knowing very well that they aren't going to be as good as a dedicated, caring human could make them.
The goal is either business process automation (do the same thing with less people) or to enable processing at a scale where doing it manually is impractical. For example, nobody would assert that an automated email spam filtering system is going to better than a human filtering my email, but an automated filter is quite useful since most of us can't afford a personal secretary. The bar for "good enough to be useful" often is lower than "human equivalent".
This point is totally valid but in the case at my work it is actually both. The old saying in marketing is "I know i'm wasting 50% of my marketing budget I just don't know which 50%." It still holds true for companies with large budgets. We have applied XGBoost to produce many and better models for how to best allocate these budgets. The results are both better and cheaper outcomes.
>People don’t make better decisions when given more data, so why do we assume A.I. will?
Because humans aren't computers. Computers are much better at being able to handle processing large amounts of data than humans can.
>we are decades away from self-driving cars
Self driving cars already exist. In college I had a lab where everyone had to program essentially a miniature car with sensors on it to drive around by itself. Making a car drive by itself is not a hard thing to accomplish.
>the largest social media companies still rely heavily on armies of human beings to scrub the most horrific content off their platforms.
This content is often subjective. It's impossible for a computer to always make the correct subjective choice, no humans will always be necessary
I read the whole article and thought it was worth my time. I liked to broad strokes of goals of anti fragile AI.
I have been thinking of hybrid AI systems since I retired from managing a deep learning team a few years ago. My intuition is that hybrid AI systems will be much more expensive to build but should in general be more resilient, kind of like old fashioned multi agent systems with a control mechanism to decide which agent to use.
People don’t make better decisions when given more data, so why do we assume A.I. will?
It's recognized that many machine learning systems today need very large amounts of training data, far more than humans facing the same task.
That's a property of the current brute-force approaches, where you often go from no prior knowledge to some specific classification in one step.
This often works better than previous approaches involving feature extraction as an intermediate step, so it gets used.
This is probably an intermediate phase until someone has the next big idea in AI.
Yep, 20+ year learning based on 10.000+ years of wisdom taught from generation to generation. People take both things for granted when comparing humans with "AIs".
This style is very funny indeed. Sometimes it's used in my language (pt-br) aswell on some articles (probably because they bought the piece from reuters or something and translated). It's very strong in writers like Gay Talese, who take it so serious they even DRESS the style LMAO
I can't imagine anyone actually reading these articles in 2021. I'm pretty sure most people buy the New Yorker to put it on the coffee table, because of the decorative covers.
* analysis of trained neural network so they're not just black boxes.
* arrangement of real neurons in actual brains of ants, mice, flies and other small animals.
* some philosophical questioning of how conscience, intelligence, awareness emerge, including a good definition and differentiation on how the brain is able to recognize causality from correlation.
* some actual collaboration between psychology AND neurology to connect the dots between cognition and how an actual brain achieve it.
Unless there are more efforts towards those things, machine learning will just be "advanced statistical methods", and programming experts will keep over-selling their tools. Mimicking neural networks is just fancy advertising about a simple graph algorithm.
1,2,4 are already occurring so I’m not sure what you’re on about. The third is completely irrelevant and seems fairly pseudoscientific, leave that to philosophers, we’re not trying to create souls.
> advanced statistical methods
Furthermore plenty of methods in machine learning, including some methods of training neural nets, are completely astatistical in nature. Unless you want to grow the definition of statistics to be so large as to consider all of maths and every science as ‘statistics’ these will rightly remain distinct fields of study (though they do overlap just like stats is used and overlaps with most sciences).
Hard to see how that could be true. In just about any field, computers today provide much better situational awareness than was possible in 1970.
The article makes the usual complaints about self driving cars:
> Despite $16 billion in investment from the heavy hitters of Silicon Valley, we are decades away from self-driving cars.
Yet cars are much more intelligent today than they were in the 1970s. And we are not decades away from self-driving cars - Waymo runs self driving cars today in very specific locations.
Wondering if this article is written by GTP-3.