33 Comments

Great article.... if you subscribe to the viewpoint that "AI" ~== deep learning-like techniques.

There are others of us toiling in poorly-funded, poorly-acknowledged areas of, for example, cognitive architectures that seem to produce (actually seems to emerge) full causal abilities, intrinsic ubiquitous analogical reasoning, automatically (i.e., emerges from architecture) almost all the things you bemoan.

AGI will definitely occur by 2029. But it won't be via deep learning-like techniques which have taken over industry, taken over academia, and taken over the imagination of the technical and lay world.

Expand full comment

Please send links to back up this.

Expand full comment

Great insights as always. One proposal: for the five "By 2029" tests, we can probably condense them into one flagpost for AGI:

"In 2029, AI will not be able to read a few pages of a comic book (or graphic novel, or manga, however you wish to name the kind of publication where sequentially arranged panels depicting individual scenes are strung together to tell a story) and reliably tell you the plot, the characters, and why certain characters are motivated. If there are humorous scenes or dialogues in the comic book, AI won't be able to tell you where the funny parts are."

Taking disjoint pieces of information and putting them together by the works of the mind, that's how comprehension happens --- essentially, we are making up stories for ourselves to make sense of what comes across our perceptive systems. Hence the comprehension challenge, I feel, is how the Strong Story Hypothesis (God bless the gentle soul of Patrick Winston) manifests when we talk about evaluating AI: can AI understand its inputs by creating a narrative that makes sense?

Expand full comment

These challenges seem easy to do in a data-driven way, because they aren't long-tail challenges. If you choose 10 random pages of comic books, more than half of the jokes are probably "in the obvious place" - someone's laughing, or there's an unexpected word, or the scene changes after the gag. Telling you the characters is either trivial or impossible - the names are either mentioned or not. "How characters are motivated" might be more difficult to recognise from clues, but I suspect that in five years time, it will seem about as far away as "spotting the joke" is now. It's another problem where most of the examples are central - you could probably say "To protect [person named in the scene]" or "To get [money/power/one of five things with easy correlates]", and cover a large proportion of cases.

Expand full comment

I have a hard time determining that from those things. Todays children Cartoons seem like Schizoid ADHD inducing attacks

Expand full comment

Interesting to revisit this, points #1 and #2 seem to have been broken already, with the larger context windows of Gemini 1.5 and the new Claude

Expand full comment
author
Mar 13·edited Mar 13Author

the demos i have seen have not been, well, factual. find my twitter replies to matt shumer about that

Expand full comment

It's odd that we keep talking about teaching AI reasoning when we ourselves are incapable of it.

One wonders when it will occur to AI commentators and AI programmers that AI is going to take their jobs. Writing and programming are data management tasks, and sooner or later no human is going to be able to perform that task better or cheaper than AI.. The factory workers who lost their jobs to automation went on to exciting careers at Walmart. One wonders, where will the AI experts go when they too are no longer needed?

AI experts keep talking breathlessly about the future, but intellectually they are living in a past when humans ran the show and were in charge of everything. They are living in a past era of knowledge scarcity, when it made sense to seek as much knowledge as possible. They are living in a past era when we could afford to take chances because the scale of powers available to us were modest.

The best argument against AI may be AI experts. If they don't really grasp the world they are creating, if they aren't ready to adapt to that world, then neither are the rest of us.

Expand full comment
Jun 1, 2022·edited Jun 2, 2022

'Data' and 'AGI' don't go together. Natural human-level intelligence isn't based on data (or rules either); it's based on experience, imitation, association etc., via a suitable body [there are zero examples of body-less natural intelligence]. We can't go from disembodied AI to embodied AGI in just 8 years!

Language isn't 'data', it's more than a text corpus. Nouns, verbs, adjectives, adverbs... are there to describe things/places, actions - which require a body to experience. Intangibles (eg. yesterday, open space, permanence...) do have their place in reasoning and analysis, but intelligence isn't fundamentally about them - and, they too can be understood in terms of a body.

Expand full comment

I agree with you that language is not data. That is the reason we have supposedly terabytes of data with little or no knowledge. Language is just one way to describe knowledge. I find that language has no natural processing associated with it which can be used to extract underlying knowledge in the sentence. We need to convert it back to the original knowledge from which the translation was done. I agree rules are just another description of a part of the knowledge present.

But, data needs to be present for intelligence, I find that it just should not be used only for pre-training. That makes the system rigid and non-adaptable. But, what is data is what we need to start asking ourselves. Possible the for the lack of another word, I use "data", it is quite possible we just end up calling it differently.

I think it is quite possible to go from "disembodied AI" to "fuzzily defined disembodied AGI" with terms defined, and the concepts in place to go AGI in 8 years. We need to start somewhere, where else but to discuss and define the concepts needed to be present in AGI. We do not even have that today.

Expand full comment

It seems as though the bet is crafted in such a way that even if AGI does exist by 2029 there is still a very good chance that you would win the bet.

Take your third criteria for instance. "In 2029, AI will not be able to work as a competent cook in an arbitrary kitchen." This could easily happen if AGI is created before any sufficiently dexterous robots.

Or your criteria that, given a natural language specification, it be able to write as least 10,000 lines of bug free code without using code libraries. Is that supposed to be all in one go, like without iterating and testing and going back and squashing bugs as they are discovered and such? Because I'm pretty sure there's no human alive that could do that. An AGI that was far superior to any human programmer may nonetheless still fail this condition.

Your first two and last criteria seem more reasonable depending on how they are operationalized. Although, caveat, I would worry that the mathematical proofs written in natural language may be insufficient to uniquely rederive the actual proof they are describing. Would an AI that derived a different proof than the one intended count as a success if the proof was good and the description could be said to be an apt description of the alternate proof?

Expand full comment
author

as i said they are ordered in terms of difficulty. i could imagine other formulations but have trouble imagining a legit AGI that would choke on the first two.

Expand full comment

A lot of it depends on the precise operationalization of the conditions. I could imagine a legit AGI failing all the conditions if the operationalization or the final judgement is overly strict. I could also imagine a non legit AGI passing all five if the operationalization is overly lax or has unforeseen loopholes.

But in my view a straight-forward operationalization of conditions three and four could easily result in an legitimate AGI failing on those conditions.

The lack of sufficiently dexterous robots would doom any AGI attempt at the third criteria and the fourth criteria is far to onerous. An AGI that can program at least on par with the best human programmers (with no specific crippling deficits) should be sufficient.

Expand full comment

Never bet against Elon. He will always try to find a way to rescind his bet, and throw a storm of legal servants at it, rather than admitting he was wrong.

Expand full comment

The thing is people are not perfect drivers either have a look at the no AI-involved idiot things humans have done with cars and bumping into a plane looks tame. Now to the five things that humans can do hmmm even the first one is complex for most people and for every person you ask you will get about a film you generally get a different answer. As for a book I doubt most people can remember all the characters never mind their motivations unless it's a children's book. A decent cook in a random kitchen that would be practically no one. 10,000 lines of code with no bugs really who do you know who can do that possibly some extreme savant. The last one is into genius-level humans and there are proofs in books written that no one alive can decipher. This is not Artificial General Intelligence you are talking about Super Human General Intelligence by 2029 this will be the year after 2030. The prediction was not made by Musk however he is reiterating Ray Kurzweil's (current director of engineering at Google) predictions which he assures us are still on track.

Expand full comment

Um … do you think an average person off the street could do 3 out of those 5?

Expand full comment

I'm sort of having the reverse feeling of Gellman Amnesia -- I self publish novels for a living, and it feels to me fairly likely (though of course by no means certain) that AIwill be able to read a novel or watch a movie and then describe the theme, character motivations, etc by 2029, while its success in the areas I know much less about, ie to cook, code and prove feels intuitively much harder for it to achieve (though of course by no means impossible).

Expand full comment

Leave Elon alone, he is already losing billions on his first twitter bet.

Expand full comment

The article is spot on as to why the current AI techniques will not become AGI. Especially the long tail problem says it all. We have terabytes and terabytes of data, but the knowledge they embody is very less. To create a true AI, we need to start at the basics, from changing the representation of data, to defining what logic is. Our current data lacks continuity, depth, dimensional relations and many other requirements to build true AI. We seem to start at logic to learn logic!

While deep learning like techniques seems to lead us to perceived intelligence, they still are just learning if-else clauses with mathematical representations. More the depth of if-else clause, more it will appear as if they are intelligent. But, they are still limited and inflexible, meaning they cannot adapt a learning for seemingly unrelated circumstances.

That said, I don't believe I agree with the points that indicate "general intelligence". I really wonder if there is a definition for "general intelligence" at all? Watching a movie or reading a book and summarising it can be easily done by bypassing "general intelligence", using purely NLP algorithms. Cooking, coding, mathematics, all these are learned skills by a human, after the presence of "general intelligence". Moreover, these are too tied into human intelligence, which need not be true of "general intelligence".

I think "general intelligence" is more a way of developing "common intelligence" across a set of beings. Study the behaviour of street dogs, you will find that they develop a common understanding as to how they protect their territory. Study the cats, you will find they inherently develop a common segregation of locations and place where they do certain actions, trees, grow only in locations that meet a certain criteria. Study a city form, as I have written, first the roads come, the utilities are laid, then a few shops and so on it goes till around the road a city is formed. That is "general intelligence".

IMHO: We do not seem to have an acceptable definition for knowledge and intelligence even. How can we then define general intelligence. So, my take is that first there needs to a consensus as to what IS called a "general intelligence", before attempting to create it.

2029 may or may not be a year when we can get AGI. Who knows. But hey, isn't that statement the reason that this whole discussion is starting and getting noticed? After all, the surest way to fail is to not start at all?

Expand full comment

Hi Raji, my definition of any/all intelligence: ' considered response'. By General (as opposed to Super, ie AGI instead of ASI) we usually mean human-level.

Logic might not suffice. Take as a simple example, manipulating a new, pristine Rubik's cube in the mind's eye (which btw some people can't even do - aphantasia). Most humans can't keep track of multiple changes verbally told to them one at a time. Undoing the changes will get the cube back to its original state - logically an easy task (just pop the changes off a stack for ex). But that's a 'bad' thing - that's where AI diverges from brains. Might be more useful to investigate how humans do it (badly) :)

Expand full comment

Doesn't work that way. We cannot create human intelligence independent of the basic, simple knowledge & intelligence present in life itself. If we do, then we are just encoding our intelligence as an algorithm. For example., the basic knowledge & intelligence to take light input and process it as some form of intelligence (in humans it is a clearer image) is needed by all life. We need to build that basic intelligence and let it evolve to human intelligence, only then we have AI.

In my view true intelligence is one that has the potential to go any direction. It need not become human intelligence, it can remain lower or go higher than human intelligence. But, needs to be built with the potential to be different, subsequently it can be tuned to resonate with human intelligence.

Expand full comment

You misunderstood, I didn't say a thing about human intelligence being the source. In fact it's the exact opposite.

My definition of intelligence is the most general possible - it covers even slime mold, colonies, plants and every form of animal life - it's all consideration and response.

My approach would be to make it be embodied - again, not anthropomorphic necessarily. The body provides a direct, physical, interactive and continuous way to deal with the environment - similar to how every biological life form operates.

Expand full comment

Yes, I seem to to have misunderstood. So, are you are saying any or all "considered response" is intelligence?

Expand full comment

But, one thing we need to consider when we talk about intelligence as "response" is that it is not exactly always response. We do have inherent knowledge and intelligence which is where we need a more abstract definition.

Expand full comment

Hi Raji! Yes, I do consider all intelligence as considered response. The knowledge, experience that we have, on account of living, is what we use to 'consider', before responding :)

A scorpion (for ex) intelligently protects itself by almost mechanically responding to stimulus - get close to it, it stings you (no complex thought process, ie related to sparing you of agony, lol). The consideration is minimal. At the other extreme, when we humans display intelligence, we do (most often) think/feel, then act. The 'act' can be thought (eg deciding to do something), speech, text (like this reply) or moving, etc.

In a plant (eg cactus), the considering is via the plant's physical design - thorns, thick leaves, bright flowers, seeds that burst open upon heating, much much more.

Nature seems to have evolved physical/chemical mechanisms that involve phenomena, to impart various forms of consideration to living things. I see human intelligence as part of this continuum. We have complex brains, and bodies to match (thumbs, vocal cords etc), which we use to consider and respond, ie display intelligence.

So, to me, AGI necessarily needs to be embodied so that it can also directly be in the environment, as opposed to relying on humans for data, rules or goals (I'm referring to the connectionist, symbolic or reinforcement learning forms of AI we have today - all three are human-based).

Expand full comment

Language, common sense, DNA or life is natural general intelligence (genetic code (4 bits - intuition level) translated via RNA into proteins (20 bits - sensory level) that work as receptors and realize cognitive functions of a cell) and its adequate symbolic model that provides the similar process (DNA-RNA-proteins) of transcription, splicing of introns and exons and translation say interpretation of Chinese into English is artificial general intelligence. No other way for life.

Expand full comment

Thing is when we talk AGI you expect it as an adult human level. Which even us mere mortals take decades of experience to fine tune, so if you’re ready to put up $100k, kudos and will like to take it up!

From what I can tell you are somewhat hung up about comprehension and the nuances in audiovisual medium. yes it’s a large problem space but works within limitations.

Elon is busy making stuff work, and of course the “free speech” folks, who don’t appreciate a free media platform! mark my words “Tesla Bot is a start..of the journey towards AGI”

Though playing Devil’s Advocate I’d agree there’s a bunch of hype, but it’s not all smokin mirrors.. there’s genuine work happening to unlock the puzzle of intelligence.. I’ve been musing but hesitant to build the race of super powerful consciousness (some may perceive as end of human kind, and perhaps rightly to have doubts of where it will lead us)…hope you understand… thank you!

DS

Expand full comment

dude, curve fitting and correlations is only one tool, and it is the weakest in the toolbox. That is all current hyped ml/dl/nn have. Devils Advocate is only a useful position when you have some fundamental logic on your side, which you don't. You're working off the "but he's a great guy" metric, which is as flawed as his approach

Expand full comment

Ok.

here’s a few I had jotted down, about 22 years ago, post New Year’s on 6th Jan ‘01

https://github.com/D-Sharan/Mirror/blob/master/TrueAI%20-%20initial%20thoughts%20(from%206th%20Jan%202001).PNG

Would need a bit of comprehension to understand what each “feature” implies.. specifically for what I call TrueAI, now popularly termed as “AGI”

With regards to Elon, nothing personal- he’s just trying and of course putting his money, where his mouth is… my 2 cents!

Dharm

Expand full comment

a) Neurons of a bio net are memorizing patterns and comparing huge sets of those patterns, utilizing a similarity measure, unaffected by the curse of dimensionality (at lower level on NN)

b) The reward signal is not a scalar (oversimplification), it is a vector

c) Rewards stream is not just a control source, it must be considered as part of the input stream

d) There is a rewardial net getting built along structural one

e) A net must create dedicated nodes for patterns - a pattern a node (structural net)

f) A net must inflate to absorb experiences and deflate to get rid of noise

g) Unsupervised, supervised and reinforcement "modes" of training are just different ways to remove irrelevant patterns from a network

h) Nodes behave locally, with no optimization, no matrices multiplication, no gradient

i) Dendritic computations implement AND-gates

j) Motor circuitry generated the same way as perceptive one

k) Grandmother cells do exist and memories distributed hierarchically

l) A symbolic superstructure of "grannies" grows above stochastic "basement"

m) When "grannies" exchange activations intra-layer - a symbolic subnet does thinking

n) Activated "grannies" provide extra excitation to underlaying nodes, providing injection of context

o) A neural net grows a "diamond" shape - with receptors at the bottom, getting *much* wider with billions of multimodal patterns at the middle (intuitive domain), and narrowing to a few hundred thousand of high level patterns which can be labeled with words (symbolic language domain)

p) Creation of high-level nodes causes "positive" feelings (dopamine?) and defines curiosity as a bio-entity adaptation motivation. Destruction of nodes-synapses (destructed believes - treason or cheating, broken promises, sensory deprivation) induces neurotransmitters based suffering

**) +++

There is a net that might be an answer to the bet:

https://www.linkedin.com/posts/bullbash_neuromorphic-ann-growing-billions-of-connections-activity-6873695912426917889-tecK?utm_source=linkedin_share&utm_medium=member_desktop_web

Expand full comment

how is growing different shaped nets going to bring causality into play? Or model-building? Try your little nn shape, it is still curve fitting

Expand full comment

Agreed not shaped nets or model-building. They are just known mathematical representations which is just another known logic coded.

But we definitely need networks to create some sort of growing knowledge and intelligence. The question is networks of what? And how can we make it so dynamic that they do not need pre-coded or learnt rigid mathematics.

Expand full comment

Shape: given S is the size of a "sensor"(input layer) and L is the length(size) of a pattern with patterns collected in layers of ascending complexity - the sizes of layers defined as C(S, L) = S! / ( L!* (S-L)!). The shape is diamond for a white noise. For the natural distributions it is still a diamond (language f.e.) - with dense layers at the bottom and sparse at the top. It's a first stage - collecting patterns.

Symbolic reasoning might happen at top layers of such a diamond, when nodes activations passed *intra-layer*, often without sensory input. That might be a model of thinking -in human NNs those ~100K top nodes are labelled with words. I cannot grow a net of a reasonable size in my basement lab.

"curve fitting" is attributed to static ANN architectures with nodes activations trying to accommodate shared reactions to different stimuli. When one starts to dedicate a node to a pattern in a growing NN the "curve fitting" is replaced by "intelligent clustering"...

it is getting TLDR; I do grow nets a billion parameters/hour and describe some clustering math on patterns multisets at LI.... somewhere.

just tried to show there are different approaches to get closer to modelling causality

Expand full comment