
The Future of Intelligence: A Conversation with Jeff Hawkins (Episode #255)
What this covers
This is a conversation between Sam Harris and Jeff Hawkins that begins in neuroscience and ends in a disagreement about AI risk. Hawkins defines intelligence as the possession of a learned world model that supports recognition, planning, and prediction. Goal achievement, in his view, follows from having good models and is not the definition itself. The neocortex, about 70% of human brain volume, is the organ that builds these models.
The first half is largely mechanism. Hawkins describes Vernon Mountcastle's proposal that every cortical column runs the same algorithm whatever the sensory modality, and he notes that the proposal is controversial. He cites evidence from rewired sensory input, argues that all cortex is sensorimotor because layer 5 projects to movement-controlling structures, and says the human neocortex grew by evolution making more copies of the same column. He also covers prediction (neural)), which he places in the thousands of synapses on each neuron and says is mostly outside conscious awareness. Only prediction errors reach attention. He also explains why theoretical neuroscience is hard to fund.
The conversation then turns to belief. Hawkins says language forces the brain to treat stated propositions as provisionally true, which makes a false belief hard to dislodge. Harris raises the "illusory truth effect" and suggests that understanding a claim may already involve accepting it. For Harris, that points to cognition and affect being more entangled than Hawkins allows. Hawkins also disputes Lisa Barrett and Antonio Damasio on whether emotion is integral to reasoning.
The sharpest dispute is over the [alignment problem](jargon:alignment problem). Hawkins treats the cortex as a value-neutral map, with motives supplied by older brain structures, and says current AI is only pattern classification. He locates the danger in human misuse and in self-replication. Harris counters that genuine general intelligence implies autonomy, and that misaligned minds far outnumber aligned ones. He also argues that understanding intelligence is a different task from knowing how to build a mind that is safe.
Hawkins argues that artificial intelligence risk is primarily a problem of bad actors misusing powerful tools rather than autonomous superintelligence becoming misaligned, because intelligence (as embodied in the neocortex) is fundamentally a modeling system without intrinsic goals or desires, and therefore does not pose an existential alignment problem.
- The neocortex is a value-neutral modeling system analogous to a map, with no inherent motivations or desires independent of other brain structures
- Alignment risk assumes AI will develop autonomous goals, but properly designed AI systems need only implement the cortical modeling algorithm without replicating motivational systems
- The real AI risks come from human misuse of powerful tools, not from superintelligence escaping human control
This asset isn't compiled yet
You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.
The brain face a significant problem of maintaining correct world models when much of the input comes through language, because language inherently requires treating stated propositions as provisionally true without direct sensory verification.
“so much of what we humans know is be a language and we have no idea if it's true when someone says something to you right how do you know and so you it's you have to so i mean i gave an example like i've never been to the city of havana well i believe it's there i believe it's true but i don't know i've never been there i've never actually touched it or smelled it or saw it so maybe it's false”
The neocortex, which comprises about 70% of the human brain's volume, is fundamentally a system for learning internal models of the world that enable recognition, planning, and action.
“the neocortex which is about 70 of the volume of a human brain... what's going on in your cortex is it's learning a model of the world an internal recreation of all the things in the world that you know of”
Sam Harris notes the 'illusory truth effect'—that merely invoking a false claim in the act of debunking it increases belief in the claim—suggesting that cognition inherently contains affective/evaluative components that cannot be cleanly separated from modeling.
“one is the illusory truth effect which is even in the act of disconfirming something you know to be false you know a some specious rumor or a conspiracy theory merely having to invoke it i mean to have people entertain the concept again even in the context of debunking it ramifies a belief in it in many many people it's just oh yeah it becomes harder to discredit things because you have to talk about them in the first place”
Most of our knowledge of the world is built on language, and the default assumption when processing language is that stated propositions are true until evidence contradicts them; this makes false beliefs arising from language especially persistent.
“so much of our knowledge of the world is built up on language and the default assumption under language that if someone says something it's true it's like it's a pattern in the world you're going to accept it if i touch a coffee cup i accept that that's what it feels right and if i look at something i stop that's what it looks like well someone says something my initial acceptance is okay that's what it is”
The root cause of false belief is that repeated language input becomes integrated into the world model simply through exposure, making false beliefs particularly difficult to dislodge even when contrary evidence is presented.
“this is really a problem of language we face and this is this is the root cause of almost all of our false beliefs is that someone just says something enough times and that's good enough and you have to seek out contrary to evidence evidence for it”
Sam Harris raises the concern that understanding a proposition may not be cognitively distinct from initially accepting it as true, requiring a subsequent rejection operation to disbelieve false statements; this suggests that cognition and affect are more entangled than Hawkins acknowledges.
“we think intuitively that understanding a proposition is cognitively quite distinct from believing it right... but there's also evidence to suggest that merely comprehending a statement if i give you a statement and you parse it successfully the parsing itself contains an actual default acceptance of it as true and rejecting it as false is a separate operation added to that”
All parts of the neocortex have motor output neurons in layer 5 that project to lower brain structures controlling movement or eye position; therefore there is no meaningful distinction between sensory and motor cortex—all cortex is sensorimotor.
“the neurons that in these cortical columns there are certain neurons that are the motor output neurons these are in a particular layer five as they're called and so in the motor cortex they were really big and they project to the spinal cord and say oh that's how you move your fingers but if you look at the neurons the columns in the visual cortex the parts that get input from the eyes they have these same layer five cells and these cells project to a part of the brain called the superior colliculus which is what controls eye motion”
The human neocortex is relatively large because evolution discovered it could be expanded simply by making more copies of the same basic cortical column; this allows for parallel processing of many separate inputs without requiring evolutionary innovation of new algorithms.
“the the basic algorithm that goes on in each of these cortical columns the circuitry in the in inside the near cortex is pretty well determined by by genetics and in fact what one of milecast's arguments was that humans the human neocortex got large and we have a very large one relative to our body size just because all i had all the evolution had to do was discover just make more copies of these columns you know you don't have to you don't have to do anything new just make more copies and that's something easy for genes to specify”
The difficulty of funding theoretical neuroscience work is a complex institutional problem that Hawkins has come to understand rather than resent, and it creates incentives for researchers to either work outside academia or pursue applied rather than foundational research.
“it's very difficult to be a scientist doing what we do again with traditional funding sources but that we don't work out we don't work outside of science we know we partner with labs and we go to conferences and we publish papers we do all the regular stuff right right yeah it's amazing how much comes down to funding or lack of funding and the incentives that would dictate whether something gets funded in the first place”
Vernon Mountcastle's proposal that the neocortex is organized into columns that all execute the same fundamental algorithm, regardless of sensory modality, has been supported by the observation that if you rewire sensory input to different cortical areas during development, those areas will process the new input according to the same basic algorithm.
“vernon mount castle made this proposal back in the 70s... the different parts of it like that do different things there's parts that do vision there's parts that do language and parts purchase through hearing and so on but when you if you cut into it and you look at the the structure in any one of these areas it's very complicated there are dozens of different cell types there but they're very prototypically connected and they're arranged in certain patterns”
Most of the brain's predictions—the continuous expectations about sensory input and motor outcomes—occur at the level of individual neurons and are not consciously accessible; only prediction errors that deviate from expectations draw conscious attention.
“if your prediction is correct then a small subset of the neurons become active but if the prediction is incorrect a whole bunch of neurons become active at the same time and then that draws your attention to the problem”
Sam Harris argues that creating a truly autonomous superintelligent system is exceptionally dangerous because we cannot predict or control the goals it might develop, and the space of misaligned superintelligences is vastly larger than the space of aligned ones.
“the intuition is that you don't get a second chance to create a truly autonomous super intelligence right like it seems that in principle this is the kind of thing you have to get right on the first try right and having to get anything right on the first try just seems extraordinarily dangerous”
Sam Harris argues that Hawkins's optimism about AI risk rests on underestimating what intelligence means—true general intelligence would involve autonomy and goal-setting capacity that humans cannot perfectly understand or control, making it fundamentally risky.
“what... you know... is based on is a it's not really taking the prospect of intelligence seriously enough and the autonomy that is intrinsic to it i mean if if we actually built a true general intelligence what that means is that we would suddenly find ourselves in relationship to something that we actually can't perfectly understand it's like it will be analogous to a strange person walking into the room”
Current AI systems, regardless of their practical power and success, are fundamentally pattern classifiers and not intelligent in the sense of building world models; most AI practitioners acknowledge they do not believe their own systems are truly intelligent.
“today's ai as powerful as it is and successful as it is i think most senior ai practitioners will admit and many of them have that they don't really think they're intelligent you know they're they're really wonderful pattern classifiers”
The neocortex is a value-neutral modeling system without intrinsic goals or desires, analogous to a map; it sits atop the rest of the brain (limbic system, brainstem) which supplies motivations, emotions, and drives.
“the neocortex on its own it it sits on top of the rest of your brain and the rest of your brain is really what makes us motivated it gets us you know we have our our our good sides and our bad sides you know our desire to maintain our life and have sex and aggression and all this stuff then your court just is just sitting there it's like a map it says you know i understand the world and you can use me as you want”
The human brain's cortex does not require emotional or affective components to build world models and make predictions; emotions and drives are produced by other brain structures and influence what the cortex learns, but the modeling algorithm itself is independent of emotion.
“you can separate these two and and i can say this because i understand actually what's going on in the new york cortex and i can see what i have a very good sense of what these actual neurons are actually doing when it's modeling the world and so on and and you do not it does not require this emotional component”
Neurons have thousands of synapses, most of which are used for prediction rather than immediate signal transmission; synapses on dendrites recognize patterns locally and cause dendritic spikes that depolarize the cell body, priming it to fire sooner when the predicted condition occurs.
“we posit in that paper and i'm pretty sure this is correct is that you know neurons have these thousands of synapses most of those synapses are being used for prediction”
Hawkins claims that self-replication (the ability to make copies of oneself) is the primary source of danger for any agent, biological or artificial, and this risk exists for biological viruses and bacteria, not uniquely for AI.
“the generic problem of self-replication which you talk about briefly and we you point out we face that on other fronts like with you know with the pandemic where we've been dealing with them so natural viruses and bacteria or computer viruses i mean there's there's anything that can self-replicate can be dangerous”
The Mountcastle hypothesis that all neocortex runs the same fundamental algorithm is controversial but has been correct based on structural evidence; if true, it implies that the principles of cortical computation apply across all sensory modalities and cognitive domains.
“vernon mount castle made this proposal back in the 70s and it's just a dramatic idea and it's an incredible idea and it's so incredible that some people just refuse to believe it but other people really think it's a tremendous discovery”
Sam Harris argues that the problem with Hawkins's position is that it does not adequately account for the difference between understanding intelligence and understanding how to build a mind that will not threaten human welfare.
“what... you know... is based on is a it's not really taking the prospect of intelligence seriously enough and the autonomy that is intrinsic to it”
The path to building artificial general intelligence requires understanding how biological brains work first, then replicating those principles; intuition about what constitutes an intelligent machine has repeatedly failed, making brain-first approach necessary.
“my approach has always been for 40 years has been like well i think we need to first figure out what brains do and how they do them and then we'll know how to build intelligent machines because we just don't seem able to intuit what an intelligent machine is”
Hawkins disagrees with neuroscientific claims (from researchers like Lisa Barrett and Antonio Damasio) that emotion is integral to reasoning; he maintains that the cortex implements reasoning and modeling independently of emotion, though emotions from other brain structures influence what the cortex learns.
“i'll take exception to that i know i know this is a opinion that you had lisa barrett on your program recently yeah like and antonio dimaggio is the person who's yeah about this the most yeah i know and i just disagree i just it's you know you can separate these two”
Intelligence requires having an internal learned model of the world, which enables the organism to recognize objects, plan future actions, and predict outcomes of behaviors.
“intelligence is just about having a model in your head and using that for planning and action it's not about doing anything particular it's about understanding the world”
The human neocortex contains approximately 150,000 cortical columns, each containing about 150,000 neurons, and each column is an independent modeling unit; the brain achieves unified perception and thought through these distributed units voting on interpretations.
“our human ear cortex we have about 150 000 of these columns”
Reference frames are not limited to physical sensorimotor space; they structure all knowledge in the brain, including abstract domains like language, mathematics, and politics, and thinking involves moving through these reference frames by activating locations sequentially.
“how is knowledge about the world is stored in these reference frames and so when uh this we're jumping ahead here many steps but when we think and when we posit when we try to you know reason in our head when even my language right now is is where my the neurons are walking through locations and reference frames recalling the information stored there”
Reference frames are spatial coordinate systems that the brain uses to locate and relate objects; they are not conscious constructs but are logically necessary for the brain to make predictions about sensorimotor interactions, as demonstrated by the example of learning the shape of a coffee cup through touch.
“the idea of a reference frame is a way of noting a location you have to have a location signal you have to have some knowledge about where things are in the world relative to other things in this case where's your finger relative to the object you're trying to touch the coffee cup”
Human intelligence should not be understood merely as goal-achievement or problem-solving under variable conditions, but rather as the capacity to build and use internal world models for understanding reality; the ability to achieve goals follows from having good models.
“if you have an internal model of the world you had to learn it... there's some things we have built in when we're born but the vast majority of what you and i know sam is is learned... if you want to say there might be an internal model that wasn't learned well that's pretty trivial... but i'm talking about models that are learned and you have to interact with the world to learn it”
The primary AI risk comes from bad actors using powerful AI tools for harmful purposes, not from AI systems autonomously developing misaligned goals; this is analogous to any powerful technology.
“there's some of the some of the risks that come about from the people's fears about the alignment problems specifically is that the the intelligent agent will decide on its own or decide for some reason to do things that are in its best interest not in our best interest... i just don't see how that can physically happen”
Building artificial general intelligence requires four key principles: (1) embodiment and the ability to move sensors in the world; (2) the use of reference frames to organize knowledge; (3) distributed intelligence (multiple modeling units voting together); and (4) learning sensorimotor models of the environment.
“they have to have some sort of embodiment they have to have the ability to move their sensors somehow in the world... it requires these reference frames... the final there's a set of things but one of the other big ones... a thousand brains is that the way to think about our near cortex... we have essentially 150 000 separate modeling systems going on in our brain and they work together by voting”
It is logically possible and practically desirable to build artificial general intelligence that implements only the cortical modeling algorithm without replicating the motivational, emotional, or drive systems of the brain.
“when we did when we build intelligent machines we have the option and and i think almost the imperative not to build the old parts of the brain too you know why do that we just just have this thing which is inherently smart but on its own doesn't really want to do anything”
Embodiment and physical movement need not be literal; virtual movement (such as following hyperlinks on the internet) or abstract movement through conceptual space can serve the same function as physical embodiment in enabling learning of environmental structure.
“when i use the word movement or embodiment and i i'm careful to define it in the book because it doesn't have to be physical it you know i example i gave you can imagine an intelligent agent that lives in the internet and right a movement is following links right it's not a physical thing but there's still this conceptual mathematical idea of what it means to move”
If an AI system implements only the cortical algorithm without motivational systems (limbic system, brainstem analogs), it would be intelligent but lack intrinsic goals or desires, analogous to a map which is a powerful tool but has no intentions of its own.
“a map is a model of the world and you can if my map can be very powerful uh tool for some to do good or to do bad but on its own the map doesn't do anything so if you think about the neocortex on its own it it sits on top of the rest of your brain”
Hawkins argues that intuitions about AI risk typically rely on human analogies—assuming AI will develop human-like motivations and desires—but this is not a necessary consequence of building intelligent machines.
“i think what most people and perhaps yourself have fears about is is they they they use humans as an example of how things can go wrong and so we think about the alignment problem or you think about you know motivations of an ai system well okay does the ai system have motivations or not does it have a desire to do anything now as a human an animal we all have desires right”
Traditional funding agencies (NSF, NIH, DARPA) declined to fund Hawkins's large-scale theoretical neuroscience work despite feedback that it was important, creating a market failure in research funding that drove him to self-fund and seek private support.
“when i was at berkeley what i was told over and over again i really came to understand this in fact uh i i went and eventually i be after that when i was running uh the redwood neuroscience institute i went to washington to talk about to the national science foundation and national institute of health and also to darpa who were the funders of neuroscience and everyone thought what we were doing which is sort of big theory large scale theories of neocortical function that this was like the most important problem to work on but everyone said they can't fund it for various reasons”
The process of understanding a novel proposition through language involves imaginative simulation—trying to see if it fits one's world model—rather than immediate uncritical acceptance, though the brain's default is to treat language input as veridical.
“if someone gives you a proposition like two plus two equals five you know i don't know what the evidence that you believe it and then falsify it but i certainly imagine you can imagine it trying to see if it's right it'd be like me saying to you hey you know sam the milk is on the right in your refrigerator and you'd have to think about it for a second you say well let me think i'll open it no last time i saw it was on the left you know no that's wrong but you would walk through the process of trying to imagine it and trying to see does that fit my model”
Hawkins suggests that most people who express fears about AI alignment do not understand the separation between intelligence and motivation and therefore anthropomorphically project human desires onto hypothetical superintelligent systems.
“most people don't understand this separation they just assume that this intelligence is wrapped up in these all these all the things that make us human the intelligence explosion problem is a separate issue”
Hawkins was able to self-fund the Redwood Neuroscience Institute (which he ran for three years) substantially through his earnings from Palm and Handspring, though he also had additional investors.
“did you have to go raise money it was well it was a bit of both it was certainly i was a major contributor i wasn't the only one but i didn't want the funding to be the driver of what we did”
Hawkins has written a chapter in his book specifically addressing how the brain forms and maintains false beliefs, suggesting this is a significant problem that the cortical model must account for.
“i know i have a whole chapter on false beliefs because so much of our knowledge of the world is built up on language”
Cortical columns cannot be directly observed under a microscope; their existence is inferred from the patterns of neural connectivity and the functional correlations between spatially adjacent neurons processing the same information.
“the funny thing about columns is you can't see them they're not visual things you can't look under microscope you won't see it but he pointed out why uh why the why they're there it has to do with what how they're connected”
A cortical column is approximately 1 millimeter in diameter and 2.5 millimeters deep, contains about 150,000 neurons, and all neurons within a column process information from the same part of sensory space or the same conceptual domain.
“if you take a very sort of skinny little one millimeter column out of it that is the processing element”
Within a cortical column, the thousands of kilometers of neural filaments constitute an extraordinarily dense and specific wiring pattern that is not random but determined by genetic and developmental factors.
“in each one of these little columns it's something like 150 000 neurons on average and if you could just unravel all of the connections there you know the the tiny filaments of of uh nerve endings what you would have there is on the order of kilometers in length you know all wound up into that tiny structure”
The brain's handling of learned versus innate knowledge involves a substantial learned component; almost all specific knowledge of objects, categories, and domains must be acquired through experience, even though basic architectural principles are genetically specified.
“there's some things we have built in when we're born but the vast majority of what you and i know sam is is learned you know we didn't know what a computer was when you're born you don't know what a coffee cup is you don't know what you know a building is you know what doors are you don't know what your computer codes are none of this stuff everything that almost everything we interact with in the world today uh in language we don't know any particular language when we're born uh we don't know mathematics so we had to learn all these things”
Hawkins was one of the founders of the field of handheld computing, starting companies including Palm and Handspring, before dedicating the latter part of his career to neuroscience research.
“he was one of the founders of the field of handheld computing starting palm and handspring”
Hawkins is a member of the National Academy of Engineering and author of two books: 'On Intelligence' and 'A Thousand Brains: A New Theory of Intelligence'.
“he's also a member of the national academy of engineering and he's the author of two books the first is on intelligence and the second and most recent is a thousand brains a new theory of intelligence”
Hawkins founded Numenta as a private neuroscience research lab where he has been working for approximately 17 years, conducting pure neuroscience research with the goal of understanding how the cortex works.
“after having all that success in the mobile computing space i started a neuroscience institute... and now i've been running sort of a private lab just doing pure neuroscience for the last 17 years that's that's numenta right that's that's nementa yeah”
The book 'On Intelligence' published approximately 14 years prior to this conversation introduced the thesis that the brain is constantly making predictions at all sensory modalities, and most of these predictions are not consciously accessible.
“my my first book which i published like 14 years ago called on intelligence was just about that right topic it was about how it is the brain is making all these predictions all the time and all your sensory modalities and you're not aware of it”
Numenta published a peer-reviewed paper in 2016 titled 'Why Do Neurons Have So Many Synapses' proposing that most synapses are used for making predictions, with neurons entering a predictive state in anticipation of firing.
“we wrote a paper that came out in 19 2016 excuse me 2016 which is it's called why do neurons have so many synapses”