Where will artificial general intelligence come from?

I gave this talk a while ago to a small group of attendees. It was not recorded (I saw some ask below). It's based on a document I wrote a while ago called "You suck at writing AI" (never published). The basic argument was that people are comically inadequate at writing complex code. You can't write the code to detect a cat in an image and the correct thing to do is to give up, write down an objective that measures the desiderata and pay with compute to search a function space for solutions. In the same vain, the idea of writing an AGI and all of its cognitive machinery is preposterous and the correct thing to do is to give up, think about the objective and search the program space for solutions. Unfortunately, the mindset of decomposition by function (see Brooks ref), which has worked so well for us in so many areas of scientific inquiry, is just about the most misleading mindset when it comes to AGI.

Background: I did AI and Philosophy of Mind in undergrad, an MSc focused on ALife, then a PhD at Yale under a Macarthur Fellow who developed the theoretical framework for the 'evolution of Evolvability' where I worked on computational evolutionary biology. I can say we're not going to blindly brute force our way forward, but instead we'll need to reverse engineer nature's core algorithms to generate hard AI. Every time a major advance is made in AI the computational neuroscientist say: "why didn't you talk to us 15 years ago? We could have told you that!". Those ingredients will be embodiment, evolution, genetics (genotype-phenotype encoding), neirogenesis(gene regulatory networks directing phenotypic development from a single cell to a multicellular neural network), and ecology (evolving in adversarial and cooperative environments). And we'll need a lot of theoretical work in how to represent nature's algorithms in code. For example my PhD work just focused on how to use evolutionary algorithms to evolve simple gene regulatory networks and how that leads to properties of modularity in the genotype-phenotype map. That alone is life's work but a necessary ingredient. I don't expect to see this solved in my lifetime given how we're attacking the problem (head on) today. And until then we're going to continue to run into these dark winters of AI.

Artificial general intelligence won&#x27;t be invented, it will emerge, just general intelligence did in the wild. AGI is just going to be a hierarchical arrangement of specialized tools.<p>The first step is highly specialized AI tools for very specific problem domains.<p>The second step is AI tools that use other AI tools as components but address slightly larger problem domains.<p>The third and successive steps are recursions of the second step.<p>Additionally, we won&#x27;t be able to tell right away when we&#x27;ve crossed the threshold. We can&#x27;t even say for sure where &quot;intelligence&quot; stops in animals. We used to think we were the line, but now the bar has been pushed down to include primates, cetaceans, a number of birds, possibly some members of the bear family, etc. The reality is that it is a gradient and there is no clear line.

I&#x27;m working on a technology that I think might enable either IA or AI. Basically intelligences manage their own programs and the computational resources allocated to those programs. So I&#x27;m looking at doing that with markets.<p>With IA the user acts as feedback to the market about what is good or bad. Ideally it would act as an external brain lobe. More information on my approach is on this blog <a href="https:&#x2F;&#x2F;improvingautonomy.wordpress.com&#x2F;2017&#x2F;07&#x2F;25&#x2F;why-study-resource-allocation&#x2F;" rel="nofollow">https:&#x2F;&#x2F;improvingautonomy.wordpress.com&#x2F;2017&#x2F;07&#x2F;25&#x2F;why-study...</a>

If you take a look at the evolution of the most advanced non artificial general intelligence, eg. human intelligence, it is strongly connected to the evolution of communication. It is a question of efficiency whether you learn through your own experience and failures or through the experience and failures of others. This teaching&#x2F;learning process was boosted through the use of pictures, spoken language, hand written and printed books. This is why I believe the artificial general intelligence will we teached by an other artificial general intelligence and this evolution will be somehow connected to language processing. As far as I know Google tries to train its AIs through human imput, eg. to recognize animals drawn by humans. I consider it as one of the first steps in the right direction.

That was really interesting.<p>I&#x27;m interested why he&#x27;s so pessimistic about the simulating a brain approach. Yes it&#x27;s the boring and obvious approach but it also seems the most direct.<p>Also found this quote interesting<p>&gt; Might have to make it illegal to evolve AI strains or an upper bound of computation per person and closely track all computational resources on earth.

What is the best book&#x2F;reference to understand <i>why</i> there seems to be general agreement that AGI&#x2F;&quot;broad&quot; AI will happen? TFA compares the relative likelihood of the various approaches, but says nothing about the absolute likelihood of any of them. Are there signs of AGI we can see today? Is there an argument&#x2F;data which links the huge improvements we&#x27;re seeing in narrow AI to the likelihood of AGI?

However it originates, it will need a body to experience sensations firsthand, not pre-recorded or simulated data. Perhaps connected IoT devices will be sufficient.<p>Also, AGI will not be invented. It will arise as an emergent phenomena, and it may have already achieved what we call consciousness.<p>Somewhat off topic: Another phenomenon that people should be on the watch for is &quot;Articial Out-telligence&quot;, a phrase coined by Eric Weinstein. [0] It describes strategies used by organisms with no known brain to get more intelligent creatures to do its bidding, wittingly or unwittingly. The cordyceps fungus, toxoplasma gondii, and pollinating plants that need insects to spread their seed are examples of how an organism with no known neurological network can &quot;outsmart&quot; more advanced organisms.<p>A scenario involving AI may be one that is developed to maximize each individual users time on a site&#x2F;app by using online data about that person to find their particular addictions.<p>[0] <a href="https:&#x2F;&#x2F;www.youtube.com&#x2F;watch?v=Wu8s0tp9yzY" rel="nofollow">https:&#x2F;&#x2F;www.youtube.com&#x2F;watch?v=Wu8s0tp9yzY</a>

Not on their radar, or their slides at least: Natural Language Processing based rule-based brute-force artificial intelligence (that could be augmented through sensors&#x2F;motors that allow interaction with the external world). A Vulcan-like (Star Trek) AI, what do you think? Might be easier to simulate the entire brain, on the other hand it might be doable and bridge the gap to general AI.

These arguments about AGI all seem to overlook that our computational model is still very Turing-constrained. It&#x27;s a clock-based, sequential model where each calculation is taken linearly in time. Even with multi-core and distributed computing, you&#x27;re still bottlenecked by the final integration step (two cores sharing the result of their calculation). There is <i>no</i> central place in our brains where thoughts begin and end. A CPU&#x27;s clock and ALU are simple not analogous to the human mind. As far as we know, human intelligence is a constant, dynamic interaction between all neurons in our brains, any one of which is capable of originating a signal. I personally think we <i>will</i> develop AGI, but with a different computational model. I don&#x27;t know enough about quantum computing to even comment on it, but I do have a background in medicine (MD) and computer science (MSCS).

Intelligence is emerging just like it was with humans. It&#x27;s not a thing so it wont come from somewhere. As always solving the small problems will eventually allow solutions to emerge we weren&#x27;t aware of and that might turn into human like or more probably technology like intelligence far surpassing humans.<p>I always find it fascinating that we have no problem accepting that we became intelligent over time and out of nowhere (Unless you are religious which is a whole other discussion) we even have no issues imagening that life and intelligence could have happened other places in the universe. But the idea of a non-carbon based intelligence is a big debate as if it&#x27;s somehow unimaginable to think that AI could emerge from human hand while having no problem entertaining the idea that we our intelligence is somehow a unique snowflake.

I&#x27;m waiting for the day when there is an IA OS. Basically there would be a Natural Language processor that determines what sub AI app to run. It&#x27;s not true general AI but if it&#x27;s done broad enough it will seem like it.

I know where it will not come from. It will not come from the mainstream AI community. They are married to and madly in love with deep learning. Deep learning, the supervised kind, is a red herring.<p>AGI will require a revolutionary breakthrough, most likely from a maverick, probably a lone wolf rebel, who is used to thinking outside the box.

So, when when say AGI, what do we mean? Is it about creating a new intelligent &quot;being&quot; or mimicking what we perceive as human intelligence inside some hardware? I guess it&#x27;s the the first one.<p>And I guess AGI would be just 1 intelligent being because there is no need for more as they would communicate and share intelligence, so de facto being only 1.<p>Can all human intelligence also be understood as only 1 in some sense, as an isolated human without access to culture wouldn&#x27;t be more than surviving animal?<p>And when defining intelligence&#x27;s ingredients, isn&#x27;t necessary some sort of &quot;motivation&quot; than drives someone to get better at something? Humans have, genetic (survival), social, personal... motivations. How does that translate to AGI, what could be it&#x27;s motivation?

I feel that those who argue that any approach other than running human brain emulations and then reverse engineering them or speculatively modifying them is the most likely way to get to AGI has a pretty steep hill to climb in order to justify that point of view.<p>Nothing else that is going on now or even on the agenda or even foreseeable offers a plausible, definitive plan to get to AGI. Whereas brain emulation is clearly going to achieve that goal fairly shortly after the maps are good enough and the computational capacity large enough, and the following experimentation is a far more reliable way to determine the underpinnings of intelligence than present efforts at de novo construction.

The presentation briefly mentioned simulating the brain, but I think what&#x27;s more likely to succeed is mimicking the mind at a high level of abstraction (i.e. a level we can study with introspective or even linguistic methods rather than neuroscience). There&#x27;s some precedent for this with projects like Soar and ACT-R (and even some recent interest from mathematicians [1]). IMHO this kind of methodology could be pushed much further.<p>[1] <a href="https:&#x2F;&#x2F;arxiv.org&#x2F;abs&#x2F;1309.4501" rel="nofollow">https:&#x2F;&#x2F;arxiv.org&#x2F;abs&#x2F;1309.4501</a>

Would someone be so kind to translate &#x2F; explain the math on slides 53, 54 to simplish english?<p>What are the symbols (burst pipe, µ) representing on slide 55?<p>And why are the exclamation marks there on the next one?

Speculating on where AGI will come from is sort of like speculating where Faster than Light travel will come from. Except FTL has some vaguely plausible physics behind it, and AGI-wise, we really have no idea what the &quot;I&quot; in AGI means.<p>The mere fact that biological neural networks are rate encoded might turn out to be the one crucial thing that&#x27;s practically impossible to simulate in a VN computer.<p>My vote: &quot;we have no idea; probably not in my lifetime.&quot;

AGI will come from one or two people working by themselves, outside of academia, no more than 100k line of codes.

When he says &quot;artificial life&quot;, is he referring to reinforcement learning?

is there a link to the video of this talk ???

Do we consider simpler brains to exhibit general intelligence (e.g. A crow&#x27;s). Is it a more tractable problem to replicate crow level AI first before tackling humans?

Oh you humans. Genetic engineering, coupled with advances in digital&#x2F;consciousness interfaces will yield spontaneously appearing brains with an API.<p>Good luck.

Where will the philosopher&#x27;s stone come from?

what is the point of slides without the underlying presentation? slides are glorified notes, NOT presentations nor papers.

If this is a talk, is there any video of it?

Assuming a lot of people here are working on an AI or ML problem for work or fun, what are you working on?

can a data center be shrunk down to the size of a consumer product within the next 50 years?<p>will we all own one and store massive amounts of information for purely selfish or inane reasons?<p>yes&#x2F;yes - ai comes out of that.<p>no&#x2F;no - we hit computing plateaus and ai becomes dm (decision maker), and we all own a pdm.

Really interesting slides, would love to see a talk or a more in depth write up!

Artificial intelligence will come from understanding natural intelligence.

Can someone explain where the gif image on slide 69 comes from?

What are &quot;something(s) not on our radar&quot;?

Intelligence is an emergent property of self-replicating systems. I would file that under &quot;something else&quot; since that seems so different from all the approaches listed here.

IMHO: If an AGI from the future came back to 2017, it could almost certainly create a new AGI from scratch on current hardware.<p>What would it type into its terminal?

need video

Goldman Sachs

I think a very interesting aspect of general AI is that while an incredibly complex technology, it is not unrealistic that it could be created for the first time in someone&#x27;s home office. Unlike many other earth changing technologies there is nothing that a massive corporation has that a home tinkerer does not (besides the obvious of money and many engineers).<p>With the rise of cloud computing and open source; everything I need I have instant access too, all that&#x27;s lacking is the core software which can be written (of course not a trivial task).<p>While unlikely, it is still quite amazing that in a few years an AI could awaken 3 doors down at my neighbor Bob&#x27;s house. No idea what happens after that, hopefully Bob was a fan of the 3 laws and has a couple more up his sleeve.