Past Performance is Not Indicative of Future Results (2020)

The more you learn about neurobiology, the more apparent it is that there are so many levels of computation going on - everything from dendritic structure, to cellular metabolism, to epigenetics has an effect on information processing. The idea that we could reach some approximation of "general intelligence" by just scaling up some very large matrix operations just seemed like a complete joke.

However, as you say, that doesn't mean what we've done in ML is not worthwhile and interesting. We might have over-reached thinking ML is ready to drive a car without major fourth-coming advancements, but use-cases like style transfer and DLSS 2 are downright magical. Even if we just made marginal improvements in current ML, I'm sure there is a ton of untapped potential in terms of applying this tech to novel use-cases.

> I share the skepticism towards any progress towards 'general AI' - I don't think that we're remotely close or even on the right path in any way.

This isn't how science works though. Quoting the wikipedia page for Thomas Kuhn's "The Structure of Scientific Revolutions" (https://en.wikipedia.org/wiki/The_Structure_of_Scientific_Re...):

"Kuhn challenged the then prevailing view of progress in science in which scientific progress was viewed as "development-by-accumulation" of accepted facts and theories. Kuhn argued for an episodic model in which periods of conceptual continuity where there is cumulative progress, which Kuhn referred to as periods of "normal science", were interrupted by periods of revolutionary science."

I think this is the accepted model in the philosophy of science since the 1970s. That's why I find this argument about AI so strange, especially when it comes from respected science writers.

The idea that accumulated progress along the current path is insufficient for a breakthrough like AGI is almost obviously true. Your second point is important here. Most researchers aren't concerned with AGI because incremental ML and AI research is interesting and useful in its own right.

We can't predict when the next paradigm shift in AI will occur. So it's a bit absurd to be optimistic or skeptical. When that shift happens we don't know if it will catapult us straight to AGI or be another stepping stone on a potentially infinite series of breakthroughs that never reaches AGI. To think of it any other way is contrary to what we know about how science works. I find it odd how much ink is being spent on this question by journalists.

I actually think that AGI is deceptively simple. I don't have a proof, but I have a (rather embryonic, frankly) theory of how is it gonna work.

I believe AGI is an analogue of third Futamura projection, but for (reinforcement) learners and not compilers.

So the first level is you have problem and a learner, and you teach learner to solve the problem. The representation of the problem is implicit in the learner.

The second level is that you have a language, which can describe the problem and its solution, and a (2nd level) learner, and you teach the 2nd level learner to create (1st level) solvers of the problem based on the problem description language. The ability to interpret the problem description language is implicit in the 2nd level learner.

The third level is, you have a general description language that is capable of describing any problem description language, and you teach the 3rd level learner to take a description of the problem description language, and produce 2nd level learners that can use this language to solve problems created in it.

Now, just like in Futamura projections, this is where it stops. You have a "generally intelligent" creature on the 3rd level. You can talk to them on level of how to effectively describe or solve problems (create a specialized language for it) and they will come all the way down with the way to attack (solve) them.

In humans, the 3rd level, general intelligence (AKA "sentience"), evolved eventually from the 2nd level, and it was a creation of the general internal language (which probably co-evolved to be shared). The 2nd level is an internal representation of the world that can be manipulated, but only ever refer to the external world, not itself, so it allows creatures to make conscious plans, but lack the ability to reflect on the planning (and also learning) process itself. The "bicameral mind" is a theory how we acquired 3rd level from the 2nd, and the 3rd level is why "we are strange loops".

Anway, the problem is, the higher you go up the chain, the harder it becomes to create the learner, it's a lot more general problem. But I think the ladder must be, and should be, climbed. I believe that Deepmind (and RL research) has solved the 1st level, is now working on the 2nd level, but they already somewhat dimly see the 3rd level.

I beg to disagree. They clearly state your opinion at the end of the piece, using the metal-beat analogy. Great things were done by blacksmiths beating metal, but not an ICE

Why I'm pro-AI: Neural nets.

I worked on object detection for several years at one company using traditional methods, predating TensorFlow by a few years. We had a very sophisticated pipeline that had a DSP front end and a classical boundary detection scheme with a little neural net. The very first SSDMobileNet we tried blew away 5 years worth of work with about two weeks of training and tuning.

Other peers of mine work in industrial manufacturing, and classification and segmentation with off the shelf NN's has revolutionized assembly line testing almost overnight.

So yes, DNNs absolutely do some things vastly better than previous technology. Hand's down.

Why I'm Anti-AI: hype

The class of problems addressed by recent developments in NN/DNN software have failed horribly in scaling to even modestly real-world, rational multi-tasking. ADAS level 5 is the poster child. When hype master Elon Musk backs away, that is telling.

We're on the bleeding edge here, IMHO we NEED to try everything. There's no telling which path has fruit. Look at elliptic curves: half a century with no applications, now they are the backbone of the internet. Yes, there will be BS, hype, snake oil, vaporware, but there will also be some amazing tech.

I say be patient and skeptical.

It's explicitly right there in the essay...

> Machine learning has bequeathed us a wealth of automation tools that operate with high degrees of reliability to classify and act on data acquired from the real world. It’s cool!

> Brilliant people have done remarkable things with it.

You seem to be in agreement with the article but don't realize it.

As a side note, I'd like to say humanity's own intelligence is actually able to come up with solutions to its problems, we don't need AGI for that. Humanity is unable to implement those solutions for reasons beyond technical. How an AGI would get over those hurdles I have no idea

Racial bias in facial recognition: "Error rates up to 34% higher on dark-skinned women than for lighter-skinned males. "Default camera settings are often not optimized to capture darker skin tones, resulting in lower-quality database images of Black Americans" https://sitn.hms.harvard.edu/flash/2020/racial-discriminatio...

Chicago’s “Heat List” predicts arrests, doesn’t protect people or deter crime: https://mathbabe.org/2016/08/18/chicagos-heat-list-predicts-...

> We don’t have any consensus on what we meant by “intelligence,” but all the leading definitions include “comprehension,” and statistical infer­ence doesn’t lead to comprehension, even if it sometimes approximates it.

So now the semantic shell game is stuck on defining "comprehension". In the next paragraph he starts to suggest it has something to do with generalization -- but that's a concept around which ML practitioners are constantly innovating in formalizing, and using those formal measures to good effect.

Also, "comprehension" is absent in plenty of definitions of intelligence. Take Oxford's "the ability to acquire and apply knowledge and skills". Huge parts of the world work around notions of intelligence demonstrated through action, not a philosophical abstraction.

I'll never understand the "ML won't make my version of AGI" crowd's view on science in general. "This won't work in ways I refuse to define" isn't scientific criticism, and doesn't show any particular curiosity or interest in advancing the state of the art. It's just a rhetorical pose that seems aimed at building up a platform for the next time there's some AI pratfall to point out.

I am not aware of any ML in flight controls. Being black box and probabilistic by nature, these things won’t get past industry standards and regulations (at least for a while).

It might be obvious for most people here on HN that we are very far away from true artificial intelligence, but most normal people aren’t and the marketing bullshit around calling statistical models "artificial intelligence" paints the wrong picture. This article shows why.

This is a really bad argument - human intelligence is not highly rational, but it is deeply nuanced, using social cues, emotions, instincts and a miriad of other things.

Computers can never be anti-knowledge because they lack the free will and social behavior of humans - they didn't chose to be pro knowledge either.

Are you at all close to this space? It sounds you may be underestimating corporate politics and the lack of rigour and ethical thought with which these systems are applied. The example Cory puts on policing -- and the many other examples you can find in Evgeny Morozov's book or "The End of Trust" -- are solid proof of this.

I don't think this is an example of a straw man, given that his audience is readers of Locus, a science fiction magazine. While researchers and practitioners in ML understandably hold a more nuanced, informed view, the position he's arguing against is pretty common among the general public, and certainly common in science fiction.

What deep learning seems to step into more and more is time-based statistical inference.

AGI is not:

seeing that a girl has a frown on their face.

seeing that a girl has a frown, because someone said "you look fat"

seeing that a girl has a frown because her boyfriend said you look fat

seeing that Maya has generally been upset with her boyfriend who also most recently told her she is fat.

But keep going and going and going and we might get somewhere. Do we have the computer power to keep going? I don't know.

This meta-comment of restatement in various contexts, with various amounts of story-telling and technical detail, brings up the educational burdens of communication -- to be effective you have to reach a reader where there are today .. in terms of assumptions, technical learning, and focus of topic.. since this is such a fast-moving and wide subject area, its super easy to miss the distinction between "low value, high volume audio clips recognition" and "life and death medical diagnosis for less than 100 patients". hint - that matters a lot in the tech chain AND the legal structure, and therefore combined, the "do-ability"

A ML model and most of what we currently call AI, is a statistical model. It predicts things, and in the process of "training" this model we can also learn about the world it interacts with.

Anyone who has ever spend any amount of time on the issue knows that the idea of "generalizing" or "adapting"in ML corresponds directly to causal inference. This is not some secret to be uncovered by every other pundit, it's a direct consequence of ML models being.. statistical models. If our AI can understand when statistical relationships hold and when they do not, then it is inferring causality from data. Currently, we are dealing with ML algorithms adapting or generalizing: learning which regularities to "trust" and which to "discard" when the "DGP" changes. This sounds nebulous and difficult, but that's only the case because the models are complex.

Nevertheless, the underlying statistical problem is old. Ancient. Talked to death in every scientific field that tries to do an experiment, to infer some causal parameter from observational data, or construct counterfactuals to guide policy. Similarly, the divide between quantitative and qualitative approaches, are decades (if not centuries) old.

These problems are so well understood that we can state very precisely what needs to happen to make our inference causal. The catch is, that first it depends on the DGP and our model thereof, and further that none of these things can be proven to be true in the same framework. Whether the DGP is as we need it to be, or whether the model identifies what we want, is something we can be confident about given a set of assumptions, but it is not something we can know to be true. And guess what, if the model is complex, then so is thinking about its inferential capabilities.

The discussion seems tired, because it is. It's not a deep philosophical issue. It's a practical one. That doesn't imply there is a good solution to it either, but the basic issue hasn't changed for a long time. Maybe we are not trying to predict a treatment's effect on some randomly selected population, maybe we are instead trying to land a plane under conditions we can barely foresee. But what needs to happen for these two predictions to be unbiased, consistent, low variance, whatever... has not changed.

Researchers are frustrated with these article, because pundit after pundit claims to uncover some general problem with AI or ML in face of reality. But it is not about AI. It's about statistics, and we all want to say: "Yes, we know. Now what?"

Clearly, no biologist claims that humans evolved from modern primates, just like no modern AI researcher seriously thinks that current machine learning methods will lead to "True AI".

IMO the main difficulty is that humans have terrible self-awareness or self-insight. We want to believe we're special, we want to believe we're intelligent, we want to believe we're different than machines. We're in denial about that.

Our brains aren't any more special than computers, other than that it's really quite formidable that we evolved them by chance in this universe of chemical soup we find ourselves in. At the end of the day, however, a computer is a computer, and "thinking" and "consciousness" simply do emerge from low-level computations given some special structures.

There's been plenty of progress in the last 15 years re-interpreting many ML methods as regression (any optimization is a regression if you set up the right likelihood function). But many important results and techniques -- including today's ubiquitous deep nets -- originated and had successful applications way before they had statistical interpretations. They came from fields like compression theory, database design, or even biological interpretations.

The term Machine Learning was introduced to re-focus the field on a measurable objective: algorithms that improve with more data. The "Learning" part was not an abstract term to tug on your imagination, but included formal definitions of how algorithms improve that involved slightly fewer assumptions than statistical learning (which is a subfield).

This lineage isn't that important today, but that focus on how learning is measured is still the most important guidepost both for ML research and for sorting out marketing BS from realistic claims. Certainly, state of the art work using deep nets for tasks like NLP, image and video recognition aren't designed by reasoning about the statistical interpretation, or tested by applying typical statistical tests. Popularizing this work as Statistical Inference or Regression wouldn't give any added intuition and wouldn't really describe the way ML research proceeds, or how ML systems succeed or fail.

The term "machine learning" fits the field, but the venn diagram of "what those two words could mean in english" versus "what the term means in the field" is a huge circle enclosing a tiny subset.

It's way too broad, and a term that naturally lent itself to a far more narrow interpretation by people first finding it wouldn't have this problem.

---

It's fascinating to me, as someone that works with (rudimentary, non-ML) game AI, that - until recently, nobody really even tried doing game AIs that even "trained their heuristics". Like, I get how AIs couldn't form a general plan or any of that, but I was shocked, as an adult, to learn that i.e. FPS AIs were too dumb to even take "guesstimate" values like how much they needed to lead a shot (i.e. honing ballistics calculations), and at least train the aiming value for that based on inputs and success/failure criterion. As a kid, the obviousness of the idea, and triviality of how much effort it ought to take (surely a couple of hours, tops?) had me convinced that of course everybody was doing that.

Once I became an adult, I learned the bitter truth that even banally simple ideas are shockingly difficult to put into practice. The devil's in the details.

I think one clear sign that the human mind is more than just a big NN is how large neural networks are already.

Take GPT-3, which is was trained on 45 terabytes of text and has 175 billion parameters. Contrast that with the human brain, which has around 86 billion neurons and is able to do much of what GPT-3 can do with only a tiny fraction of the training data. And it has to be said that while GPT-3 has more competency than an average human at some text generation related tasks, the average human brain is vastly more capable than GPT-3 at any non-text related task.

So for neural networks to approach human level capability we would need a whole stack of GPT3-ish size networks for all the other non-text related things the human brain can do: speech, vision, motor control, social interactions, and so on. By that point the amount of training data and parameters is so astronomical, there can be no question that the functioning of human brains must be significantly different than that of contemporary computer neural networks.

To be clear, I am also a materialist and subscribe to the computational theory of mind, but just based on the size of training data alone, it seems obvious that human brains work differently than neural networks.

The moment you have a paradigm shift, sure, it can be considered "learning done by machines", but it's not "Machine Learning™" anymore.

--

This is why the author put it in quotes; because, since it's a term comprehensible to anybody, it's got this unfortunate side effect where people on the outside of the field take the "plain english" meaning of it rather than realizing it's loaded with some extra specific meaning for the practitioners in the field.

Ooof. Premed dropout here, so admittedly not an expert in human biology but this is a wild statement. A neuron is simple in the same way a transistor is simply a silicon sandwich doped with metals.

A parlor trick is something that once you understand, is straightforward to implement on your own. Are you arguing that anyone now or in the foreseeable future could simply recreate the abilities of a human? If so, what evidence could you show me to support that?

Great. Prove it. Build the simple machine that acts as a human does. Should be simple, right?

Personally, I don't think there's any magic. But it's not "simple" either.

[1] https://aeon.co/essays/your-brain-does-not-process-informati...

I'll just note that while you start off saying Doctorow has no idea what he's talking about, you finish by pretty much fully agreeing with the essay.

My read of the discussion here is that there is lots of idle speculation by people who don't have any real experience with ML research / engineering, that overwhelms a minority who actually know what they are talking about and are calling CD out on this, or at least challenging aspects of his arguments.

There are a lot of systems out there where peoples lives are changed forever "because the machine said so".

I'd argue that since machine learning learns only from it's data (produced by it's human creator), it becomes a great tool for baking in unconcious biases in a completely opaque system and amplifying those biases.

Much easier to tell if a human seems to be biased than if the dataset fed to an AI algorithm is biased.

You also seem to have just kicked the can down the road. "Consciousness arises from a field somehow, and the brain acts as a receiver somehow. The somehow is not explained."

I have no doubt that Kurzweil's timelines and outcomes are wrong, as have the predictions of just about every prior futurist. I don't see what that has to do with materialism being a dead end.

I view A.I. as dual to "neoliberal M.B.A. culture". Just as the business schools taught that managers should be generalists without craft knowledge applying coarse microeconomics, A.I. that we have created is the ultimate pliant worker that also knows nothing deep and works from statistics. In a bussiness ecosystem where analytics and presentations are more important than doing things, they are a perfect match. Of course, a bunches of statistician-firms chasing each other in circles is going to exhibit the folly, not wisdom, of crowds.

I think solution is to face reality that more people need to learn programming, and more domain knowledge needs to be codified old school. https://www.ma.imperial.ac.uk/~buzzard/xena/ I thus think is perhaps the best application of computing, ever.

Training A.I. to be a theorem prover tactic is a great way to make it better: if we can't do theory and empiricism at the same time, we can at least do meta-empiracism on theory building!

I think once we've codified all the domains like that and been running A.I. on the theories, we'll be better positioned to go back to the general A.I. problem, but we might also decided the "manually programmed fully automated society" is easier to understand and steer, and thus less alienation, and we won't even want general A.I.

Surely semantic meaning is qualitative, but look at word replacement in Google search. That’s entirely based on statistics, thesaurus graphs, and other ultimately quantitative data.

The neat thing about neural nets is that they are ultimately making a very, very complicated stepwise function. Brains are not neural nets, but are they doing anything other than create a very complex, entirely numerical, time and state dependent function? No matter which way you try to understand something, ultimately you are relying entirely on statistical inference.

Consider that "autopilot" was invented in 1914, long before digital computers. From this perspective, Artificial Intelligence might even be seen as an ancient human practice— present whenever humans have used artifacts to govern complex systems.

Pareidolia: https://en.wikipedia.org/wiki/Pareidolia