I've looked through Lesswrong’s archives for mentions of the thalamus and its role in consciousness and I'm a little surprised that it is something which hasn't been discussed here before. In recent years it has become very clear that the thalamus is vitally important in consciousness and very arguably is the seat of consciousness ^[1].

The main discussant of the thalamus on Lesswrong is Steven Byrnes and the concepts I'm laying out here generally align very well with Byrnes’s writing with perhaps the minor addition of asserting the thalamus as the chief originator of the groups of options that are then run through cortical circuits. I agree with the central importance Byrnes places on the flow of thought through cortico-basal ganglia-thalamo-cortical loops, though I place special emphasis on the thalamus.

Integrated Information Theory proponent Giulio Tononi’s mentor, Gerard Edelman, also recognized the central importance of the thalamus. In a short clip on YouTube we can hear Edelman discuss the role of thalamus and thalamo-cortical relay in consciousness.

In the clip Edelman discusses thalmo-cortical loops which are what give rise to the excitation patterns we detect as brain waves, to a large extent (alpha, beta, gamma, and delta). These wave events are theorized to cause what has been labeled “synchrony” or “temporal binding.” Here in his review of Gyorgy Buzsaki’s book on brain waves Scott Alexander describes one popular hypothesis on synchrony:

Brain waves provide “synchrony”, allowing a smallest granular unit of time and essentially converting life into a turn-based game. Suppose that a snake bites your foot. You see the snake with your eyes, and also get a pain signal from your foot. The pain signal has to travel a long way, nerves have conduction delays, and so it reaches your brain well after the visual signal. But your brain needs to be able to combine the visual and pain signals into a single story (snake bit my foot). Brain waves separate experience into short granular “turns” so that the brain can attribute both stimuli to the same “turn” and connect them. It’s also possible I’m totally misunderstanding this part, sorry.

I think he understands it. This is similar to this description of current theory on synchrony in thalamo-cortical loops:

Recurrent thalamo-cortical resonance or thalamocortical oscillation is an observed phenomenon of oscillatory neural activity between the thalamus and various cortical regions of the brain. It is proposed by Rodolfo Llinas and others as a theory for the integration of sensory information into the whole of perception in the brain.[1][2] Thalamocortical oscillation is proposed to be a mechanism of synchronization between different cortical regions of the brain, a process known as temporal binding.[3]

https://en.wikipedia.org/wiki/Recurrent_thalamo-cortical_resonance

The slideshow of moments of which we are conscious plays by the beat of thalamo-cortical loops.

And while it may be hard to argue where the “start” of something that is inherently looped is, the thalamus is perhaps the only brain region which has been shown to be completely necessary for consciousness. No thalamus no, consciousness; you're in a coma. Here Ned Block explains in an interview that thalamo-cortical loops are necessary for subjective experience, and when challenged that this only means wakefulness and perception he points out these loops are indicative of the subjective experience of REM sleep when subjects are dreaming. Gerard Edelman puts it very succintly:

It's known that if you have a stroke in the thalamus you can become completely vegetative and not have any consciousness ever again. If you have a stroke in the cortex – for instance, the V1 where you handle vision primarily, on the path from the thalamus, V1 – you're blind, okay, but you're not losing consciousness. To get the cortex to have you lose consciousness you'd have to obliterate most of the cortex. But the thalamus is another story.

And not only is the thalamus necessary for consciousness, it also seems to be basically sufficient. Thalamic electro stimulation can turn conscious awareness on and off like a flicking lightswitch in fully sedated lab animals (YouTube discussion of the paper). That is to say they knocked out a monkey with anesthetics and subsequently were able to precisely manipulate its state of consciousness or unconsciousness by electrically stimulating only the thalamus. This has been used clinically in completely unresponsive patients, returning some to full awareness.

Very much like the heart pumps blood through the circulatory system, the wave generation of the thalamus pumps electrochemical cascades through the brain. Turn off or remove the pump and things stop moving; and inversely if nothing is flowing you can stimulate the pump to get everything moving again.

But more than just pushing thought around, the thalamus is important in creating neural consensus between brain regions and option alternatives. There is evidence that consensus generation happens in the thalamus through mutually inhibiting connections in the thalamic architecture and the nearby thalamic reticular nucleus (TRN). Here in an interesting general talk about thalamo-cortical loops Murray Sherman of U Chicago explains that these sort of mutually inhibitory connections are dominant in the thalamus's architecture and a few moments later specifically mentions that exciting one of the neurons in such a structure will inhibit nearby similar neurons.

To be clear: the dominant wiring of the thalamus is such that when a given neuron fires it inhibits a cluster of nearby neurons. This is funnelling these electrochemical waves into a narrower channel, eliminating other options and focusing on the chosen alternative.

So in addition to the heart the thalamus is also the senate chamber of the brain. This sort of consensus generation is precisely what happens to create the unitary individual from the disparate recommendations of the cortex ^[2]. 

To briefly reiterate some of what Steven Byrnes has written, in essence: he also conceives of a number of options which are run through loops in the brain; mechanisms in the brain cause one of these options to be reinforced and amplified (through valence) while others are inhibited, consolidating into some kind of decision over the options. I largely agree and am pointing out the central role of the thalamus (and the TRN which surrounds it) in this process.

What I propose: some group of neurons inside the thalamus is excited through either a continuing internal loop or an external stimuli; that sends a group of excitations to the cortex; these excitations are then poured through the complex interconnections of some region of the cortex and return to the thalamus, sometimes by way of the basal ganglia; whichever of the original thalamic cascade returns to the thalamus with the strongest or quickest signal inhibits the other options that were fired in that group while the winner continues on reinforcing itself and/or some related subsequent discrimination target/choice (which itself is some kind of thalamic excitation projecting back to the cortex).

In the case of a loop with the occipital or parietal part of the cortex this would be a task to discriminate between different interpretations of some visual stimulus. Is that bright vertical strip that appears somewhat in motion actually the pant leg of a person walking into the road or a detached fence post that only appears to move because of the motion of the car’s headlights? That is the kind of discrimination task I would be talking about in such a thalamo-cortical loop. You could conceive that these two options along with possibly others are fired off into various parts of the dorsal and ventral visual areas to combine with other visual information to act as an input vector to the network of existing learned cortical connections to see whether the neurons connected to the fence post in the ventral area or the neurons connected to the the pant leg object in the ventral area are more excited. If the fence post wins this race through the cortex's maze back to the thalamus then the driver's thoughts likely drift off into other topics as the fence post possibility doesn’t exert strong inhibition on those thoughts. But if the person’s-pant-leg object wins the cortex maze race the driver's attention will be hijacked as other thoughts are inhibited, he will continue to think about people-walking-into-the-street objects, and the loop between the thalamus and the person part of the ventral area will have reinforced activity, possibly leading to longer loops that suggest motor commands to swerve around the person's projected location.

I would speculate that an evolutionary story is that the thalamus moved from a role of filtering incoming sensory information for attentional purposes to filtering internal thoughts for similar attentional and reinforcement purposes. The thalamus extended from the task of deciding to which sensory patterns the mind attends to more general decision making. This evolutionary story is a strong argument to conceive of decision making as happening in the thalamus despite being influenced by the cortex and the decision-making process being inherently looped; the thalamus or thalamus-like structures have been making decisions since before there was a neocortex.

Brain waves are known to generally travel from the posterior portion of the cortex to the anterior portion. This reflects a somewhat intuitive and natural process of first object discrimination in the occipital, temporal, and parietal lobes followed by action recommendation in the motor cortex followed further by higher level complex decision considerations in the humans’ oversized forebrain before sweeping back posteriorly on the ventral side of the brain, passing through structures like the amygdala, hippocampus, and the mammillary bodies to ultimately provide strong valence signals through the limbic system back to the thalamus. Very much the observe, orient, and decide elements of the OODA loop in basically the same order (arguably the “act” part requires the rest of the peripheral nervous system and the musculoskeletal system) ^[3].

Of course, this large grand circuit doesn't always trigger all at once (although something like suddenly swerving to avoid a pedestrian might do it, possibly recruiting forebrain areas that manage stuff like morality, laws, and insurance coverage in addition to motor recommendations and object identification) and it is perhaps best to analogize these structures as like a wheel with spokes with the thalamus as the hub of the wheel and the exterior surface being the neocortex. The spokes of the wheel, representing white matter thalamo cortical connections, allow the hub to send a signal to some specific portion of that wheel, run across a short section of wheel exterior, and return in an abbreviated loop through one of the spokes. Likely in many cases these short abbreviated loops are serially triggered, warming up relevant sections of the grand loop before a somewhat stronger electrochemical cascade is sent through the pre-warmed path travelling far across the cortex, representing a complex decision.

And while the basal ganglia are recruited to provide especially strong signals, I think much of the valence is provided purely through the cortex's interactions with the inhibitory neurons of the thalamic reticular nucleus or the thalamus itself. That said, the basal ganglia would be used to establish more general goals. For example: the hypothalamus may weigh food related thoughts more heavily while hungry, the amygdala may inhibit thoughts/loops that relate to a cortical pattern associated with a previous negative experience, or the hippocampus tailoring valence reaction relative to a series of planned sub-goals which can be followed sequentially to achieve a larger goal.

I think this model addresses one of Scott Alexander’s questions at the end of his review of Rhythms of the Brain:

If brain waves are really responsible for things like attention and cross-talk between brain regions, then might a lack of brain-wave-equivalents make AIs worse at these things?

A brainwave, or single thalamo-cortico-thalamic loop, seems very analogous to a single pass through a neural AI system’s weight matrices. These moments of conscious experience where we make decisions on what we are observing or what we are doing 10 to 40 times per second are the brain sending a pulse of signal through the weights of our cortex to usually perform some kind of decision or discrimination task.

I'd speculate that SOTA AI systems, particularly LLMs, probably don't have a lot of the episodic memory related elements of consciousness that arise from the basal ganglia, particularly the hippocampus but also certain elements of the amygdala's role in making emotional associations to sensory patterns. While large contemporary models can execute complicated multi-step actions I speculate that much of the time it is more akin to memorizing a complicated set of dance moves that you flow through semi-automatically rather than a sequentially planned episode with sub-goals. A mind with a severely damaged hippocampus is an amnesiac; they also have difficulty with planned tasks. The way certain agentic systems which do generate subgoals accomplish this subgoal generation probably very differently from the human brain; they likely are doing thalamic-level "amnesiac-choice-tasks,^[4]" choosing next steps based on what sounds appropriate at a gut level, rather than recalling a successful episode of the task from which it extracts sub-goals.

This hub-and-spokes model of the thalamus’s role is suggestive of an AI architecture which can selectively activate small parts of a larger model/weighted connection network. In some ways this bears resemblance to some agentic system designs where a central component (thalamus and TRN-like) delegates tasks to specialized subcomponents (cortex areas-like).

While I do acknowledge that other areas of the brain contribute important aspects of conscious experience, like the individual basal ganglia’s contributions to valence qualia or how the precuneus creates our sense of space and spatial qualia, the thalamus has a distinctive central role in what may be the highest function of consciousness: making decisions. Given this central role in orchestrating the flow of thought through the circulatory system of the cortex’s pathways, creating the unified consensus of the individual mind, and the pulsed nature of consciousness, it is perhaps appropriate to label the thalamus as the brain's heart and view it as the seat of consciousness.

1. ^{^}

   I want to mention that the claustrum, Crick and Koch’s favored structure, is highly interconnected with the thalamus. This close interconnection may be the basis for the sometimes correlation of the claustrum with indications of consciousness. That said, there is not a similar structure in the claustrum to the TRN which may perform the consensus generating elimination of unchosen alternatives that we observe in the thalamus, at least as far as we currently are aware; more discussion of this consensus function later in the post. 

2. ^{^}

   Of course it does seem that some of this mutually inhibitory consensus generation between the two halves of the thalamus is mediated by some cortical structure like the corpus callosum; the massa intermedia which connects the two halves of the thalamus is absent in 20% of people and probably doesn't play much of a role in this mutually inhibitory consensus generation.

3. ^{^}

   That said, there are certainly processes that could go in the other direction, like deciding to imagine some specific kind of object, like a chair, which may activate the “chair” object associated neurons in the ventral stream temporal lobe that project back into an excitation pattern for low level individual pixels in our occipital lobe’s V1.

4. ^{^}

   I think a good example of the sort of complicated tasks and discriminations performed almost entirely at the thalamic level with little intervention from the hippocampus may be something like riding a bike; a skill retained by amnesiacs.