In the popular consciousness, there's a lot of understandable confusion about the role of machine learning (ML) in artificial intelligence (AI) and in the pursuit of artificial general intelligence (AGI). There's substantial conflict in the opinions on this matter between ML optimists and pessimists, or perhaps "believers" and "nonbelievers". But a common trap in non-technical discussions surrounding these topics involves the attempted comparison of fundamentally incomparable ontological categories. I like to understand the relation of these terms and categories, and, more generally, the role of ML in AI, through an analogy about aircraft, propulsion systems, and advanced manufacturing processes.

AI is the domain and study of systems and algorithms that perform "intelligently". I'd liken this domain to the category of aircraft. There are very few qualifications regarding what the category of aircraft includes. Any man-made thing that flies is an aircraft, whether it be a paper airplane or a stealth fighter. Similarly, regarding AI, there are no real rules as to what defines intelligent behavior. An intelligent behavior of interest can be as narrowly focused as the ability to play chess or as broadly performant as a general social capacity. An AI system may or may not have memory, may or may not have a learning capability, and may or may not demonstrate human-like intelligence, let alone consciousness. That all depends on the type of AI (aircraft) we're talking about.

AGI is a term for AI that matches human performance across a broad spectrum of general and abstract reasoning tasks. I like to think of AGI as a specific category of aircraft with extreme capabilities. I further like to imagine this category as performance-defined. Let's call it aircraft that can break the sound barrier. That certainly requires some hell of an engine. But this category is agnostic as to whether that engine is a jet, a rocket, or some other technology entirely. Similarly, I think AGI is best defined in terms of loose performance standards. For instance, can this AI system independently perform high-quality scientific research? Can this AI system write a decent book?

Humans write good books because the human brain is a rocket that propels us past the sound barrier. The first AGI systems might just use jet engines. They might not function like a human mind at all. Notably, a jet engine doesn't function in space, while a rocket does. Whatever the hazily understood quality of consciousness is, I think of it like the quality of spaceworthiness. That's different than breaking the sound barrier. An aircraft might break the sound barrier without being spaceworthy. It might be spaceworthy without being able to go supersonic in Earth's atmosphere. Like breaking the sound barrier, there might be multiple propulsion types that enable spaceflight.

As humans, we associate general intelligence with consciousness because our human brain is a sample size of one. We only have one supersonic aircraft—our brains. It's a rocket, and it's also spaceworthy. But I think that conceptualizing AGI in parallel to the categories of rocket-propelled or spaceworthy aircraft would be a mistake. A performance-focused definition of AGI that is agnostic to system design and secondary characteristics like consciousness is the right one because the safety considerations and economic impacts of AGI are indifferent to how it works. An unconscious system that looks nothing like a human brain but that can still design bioweapons, execute cyberattacks, generate disinformation, and replace technical jobs requires special societal consideration.

I see the point of defining the ontological category of AGI to be that of aiding society to direct special treatment towards advanced AI systems that pose unique hazards to public safety and economic stability, while simultaneously promising unique upsides. Those unique aspects require a unique social and regulatory treatment, just like supersonic aircraft require a unique regulatory framework to prevent our living spaces from being regularly disturbed by sonic booms. That framework requires a lexicon tailored to identifying risk. Of course, distinct categories like "artificial consciousness" and "artificial emotion" may eventually also become relevant in guiding the legal discussions regarding the potential rights and sovereignty of artificially intelligent organisms.

A lot goes into making a supersonic aircraft. There's the engine itself, whether it's a rocket or a jet. And each of these categories also includes subcategories. There are solid-fuel and liquid-fuel rockets, and there are turbine jets, ramjets, and scramjets. Whatever engine type we use, the engine doesn't fly on its own. The engine must be integrated into an airframe and with a control system. The engineering of the aircraft as a whole is related to but categorically distinct from the engineering of the propulsion system. And a great jet engine may not power just one aircraft design. It may be incorporated into a wide variety of aircraft with different purposes, some of which may be capable of supersonic flight, and some of which may not.

In the current AI ecosystem, I think of the engineering of advanced aircraft in parallel to the engineering of AI systems and agents. Each of these AI systems requires a core intelligent model of some kind, which is analogous to the propulsion system. There are multiple types of models that may be able to power AGI systems, including both language models and world models. Within the category of language models, in particular, there are a variety of architectural variations. The transformer architecture has proved highly capable as a basic design for large language models (LLMs), but has now speciated into many variants, and is not even the only way to build a language model in the first place. A good model, moreover, can be used as the engine of a variety of different AI systems and agents, some of which may qualify as AGI, and some of which may not.

So, what is the role of ML in this modern AI ecosystem, and where does an ML scientist fit into AI design efforts? I like to think of ML as a revolutionary manufacturing technology, like 3D printing. While traditionally, we build large, complex things out of small, reusable components like sheet metal and rivets, 3D printing allows us to manufacture complex, custom-purposed things, all at once, with microscopic precision. Similarly, rather than designing complex algorithms based on modular components through the explicit specification of behaviors, ML allows us to learn extraordinarily complex algorithms, all at once, from data. The implications are vast, and the science of improving ML techniques is a domain in and of itself, just as research efforts in improving 3D-printing technologies is a domain in and of itself.

And indeed, the implications of 3D printing extend far beyond aircraft alone. Biotechnology researchers are experimenting with 3D printing skin directly onto the human body in order to aid in the recovery of burns and wounds. Similarly, ML has a wide variety of impactful usages beyond traditional AI systems. Speech-to-text technology, for instance, is not really an intelligent process. Or at least, it is a simple conversion that does not require a reasoning-based approach. But ML enables the learning of speech-to-text algorithms, from data, that would have been prohibitively complex to design by hand using traditional methods. The applicability of ML science extends to every case in which we require complex algorithms that challenge direct design but that are in some way correlated to patterns expressed in data.

To be sure, 3D printing can make a hell of a jet-engine part. The types of jet engines we can manufacture with the benefit of 3D printing are far more diverse and sophisticated than those we were limited to manufacturing without 3D printing. Some of those engines might even be integrated into supersonic aircraft. But maximizing progress requires optimizing the ways that we use this manufacturing technology. It might be overoptimistic to try to 3D print an entire jet engine all in one go. The most powerful approach might be in identifying the core engine components that we can 3D print individually, and then assembling these components into a well-tuned engine with the aid of other manufacturing processes.

Similarly, one of the most exciting and transformative usages of ML science is in the design of AI systems and in the pursuit of AGI. And, in fact, the contemporary AI revolution has been powered by ML. But effective AI design still requires optimizing the way that we use ML. The most high-performing AI language model may not be learned as a single block component. It may be assembled from smaller, purpose-built components learned via ML techniques and integrated via other algorithmic approaches. So when people disagree about the potential of ML in AI, polarizing into camps of ML "believers" and "disbelievers", I think those people are viewing the problem from a perspective that is too constrained. Absolute optimism and pessimism should probably be balanced somewhere in the middle. I'm confident that the first AGI systems will rely heavily on ML components. I also suspect that these systems will integrate a multiplicity of algorithmic components designed or learned through a variety of processes.

As an ML scientist, I like to think of myself as a domain expert in 3D-printing technologies. Like a scientist researching techniques to print using new materials, or at larger scale, or with atomic-level precision, I'm interested in fundamental advances in ML methods. Like a medical researcher investigating techniques for printing live tissues, I'm interested in general applications of ML outside AI. Like an aerospace engineer using 3D-printing technologies to manufacture jet-engine components for supersonic aircraft, I'm interested in AI-purposed ML applications and in using ML to build AGI. And just like an effective usage of 3D printing requires an understanding of how to integrate a variety of 3D-printed components using the whole spectrum of available manufacturing techniques, I think that a true mastery of ML science involves understanding how to integrate a variety of learned and designed components using the whole spectrum of available algorithmic methods.