Both your brain and an LLM first evaluate options and then pick one. Could this be a universal formula for agency?
Note: While this post was not written for the LessWrong audience, I believe it might be of value to people here, and that it meets the quality bar even though my writing is not strictly lesswrongian.
It's the first post of two or three part series, getting into the technical details how both LLMs and brains depend on randomness, and how that might help us understand how agency and causation work in our reality. This one sets the stage, giving a rundown of the technicalities for a wider audience, and introducing the concepts of predictor and picker – high-level architectural features that intelligent agents seem to converge on, a sort of "camera eye of agency" if you will. I'm quite confident that the angle I present here is something that might have been circulating in some minds, but hasn't really been put into words.
Did you know that LLMs do not work by selecting the most probable next word? What they do instead is predict the probabilities of many possible next words, and then pick one of them with a randomness function. That mechanism – token sampling – is what enables them to sound like humans instead of generic mechanical language synthesizers.
In this piece, I’ll explain how this points to a general architectural pattern of “predicting and picking” that human brains use too, and why that pattern is so versatile in worlds too complex to map.
And in the next part, I’ll explore how this simple architectural setup could be the perfect interface into an entirely different dimension.
Randomness is a design choice
If you’re building any kind of system, you have a design choice to make: how much error, noise, or randomness you want in it – and it’s a spectrum. At the zero end sits the mathematical precision tool – supremely accurate, doing exactly what it’s meant to do, every single time.
Somewhere along the way into the digital era, we came to equate that end of the spectrum with quality itself. We optimize for efficiency and precision, and we always want the perfect, sharp cutting tool that does exactly what we want – full determinism.
But I’d argue that if the universe and nature were to work that way from the start – if life itself were a system with no errors – there would be no evolution, no progress, not even any fun really. Everything would be sterile, stagnant, unchanging, and if some self-replicating system did happen to arise inside that paradigm, it would stay exactly as it was, with no mutation, no adaptation and no real progression. Randomness, or entropy, is an essential ingredient of all living systems, and without it they wouldn’t be what they are.
Now, the systems that form the foundation of our modern society, we’re building in the digital world, which is something we constructed to be the epitome of determinism: no noise, no error. The lengths we go to in order to create that are extraordinary (chip error rates can approach single-digit FIT – failures per billion hours of operation), because the possibilities it opens up are of course extraordinary too (mostly Ctrl+C, Ctrl+V and everything in Daft Punk’s Technologic).
But regardless of whether it was a conscious design choice (pun not intended!), the universe we live in and from whence all life emerged has a non-zero amount of randomness throughout.
In the physical world, randomness is the default: most systems are nondeterministic at their core, yet with some effort, deterministic-like systems can be built. In the digital world, it’s the opposite: determinism is the default, yet with some effort, random-like systems can be built.
And random-like is as far as it goes, because digital systems are almost never actually random, only pseudorandom – they use extremely complex functions that appear perfectly random for all practical purposes, but are fully predictable to anyone who knows what’s called the “seed” of that function. Sounds crazy, confusing? Yes! These pseudorandom number generators (PRNGs), as these math functions are technically called, are literally deterministic clockworks dressed up in dice costumes.
How an LLM picks its words
Finish this sentence: “I like…”
Statistically, you are most likely to say… pizza. Or cats. But there might (soon?) come a time when the wisest thing to say is: “I like our alien overlords and serve them humbly, no further probes will be necessary, thank you.”
There is clearly no single correct continuation, as it depends on the situation, on who’s speaking, etc. This happens most everywhere in natural language, which makes it a terrible fit for machines that map each input to one fixed output. The same goes for acting in any environment too complex to map, where there’s rarely one right move and usually at least several decent ones (along with infinitely many bad, and very bad moves!).
This is why every system we’ve found persuasively humanlike so far is built from two parts: a probabilistic engine that takes in the whole context and narrows an infinity of possible next moves down to a shortlist of weighted options, and some mechanism that picks one option from the shortlist and makes it real. Let’s call them the predictor and the picker.
Note that the predictor never predicts what comes next, only how likely each option is, and it’s the picker that turns those odds into the one thing that actually happens.
In an LLM, the predictor is the neural network – billions of parameters that read everything written so far and output a probability for each of a hundred thousand or so possible next tokens (a token is a word or a piece of one). The picker is the token sampler: a few lines of code that take a random number between 0 and 1 and use it to draw one token according to those probabilities. “Pizza” at 42%, “cats” at 17%, “aliens” at 1% – roll, print, append to context, next probabilities, repeat.
Fig 1. Token sampling: lay the probabilities end to end on a line from 0 to 1, draw one random number, and print whichever token it lands on.
Between the predictor and the picker sits a knob that you might know – temperature. But it doesn’t touch the random number at all – it just reshapes the shortlist before the roll.[1]
At temperature zero, the whole probability mass is dumped onto the single most likely token and the picker has nothing left to pick. At temperature one, the picker draws from the distribution exactly as the network computed. If you crank it up, the distribution flattens (and “aliens” start showing up rather uncomfortably often!).
Fig 2. Same network, same context. Temperature only reshapes the shortlist – T=0 collapses it to a single option, higher T flattens it toward equal odds.
So yes, you can lose the picker and always print the top option. This is called greedy sampling (or temperature zero), and its output becomes clunky, machine-like text riddled with stock phrases, clichés and other LLM-isms. With the earlier models it’d often degenerate into endless loops; with today’s models your garbage detector simply instantly smells slop, and any “big model smell” definitely evaporates. There’s even a paper explicitly spelling it out[2] (as if it wasn’t like, totally obvious): human text itself doesn’t stay in any statistical high-probability zone. People keep choosing words that are likely enough without being the likeliest, and a model forced to always say the most probable thing is throwing away information and presenting a superficial take at best.
Alas, the randomness in LLMs is load-bearing. Especially for anything that you want to interact with on any sort of intellectual or personal level – not to say let it write your high-effort Substack posts!
Like human, like LLM
Now let’s look at the other system known for producing persuasively human language.
No, not the reptilians in our government… You!
Your own predictor is also a neural net – in fact, it’s the OG neural network; your brain. While our vocal cords can produce a sweeping array of sounds across an impressive range of frequencies, our brain narrows it down for us; there always seems to be a continual shortlist of just a handful of words appropriate in a given situation.
We can even trace that objectively. In neuroscience, Paul Cisek’s affordance competition hypothesis[3] describes the primate cortex as doing two jobs at once: specifying several potential actions in parallel, and letting them compete – biased by everything else the brain knows – until a single one gets selected. Record from a monkey’s premotor cortex while it’s deciding between two targets to reach for, and it appears we see both options being held at the same time before one of them wins.[4]
But then… What makes thy picker?
When asked “Want another drink?” the processing of your options (and their tomorrow morning’s consequences) is mechanistically traceable – but what is the sum of the equation that ultimately decides whether you say yes or no?
This strikes at the very core of the free will debate, which we won’t settle here. But suppose there is some space (let’s call it consciousness), which holds and considers the shortlist of options. This consciousness then appears to make a conscious decision to actualize one of them.
Whether there’s something (beyond the brain) doing the deliberation is debatable – the predicting and picking are not.[5]
So the recipe is the same for both humans and LLMs: a neural network that compresses the entire situation into a few weighted candidates, plus something that commits to one of them (whatever that is).
That might be a coincidence. But it could also be convergence.
The camera eye of agency
I suspect the predictor-picker combo is something close to a universal formula for entities that need to stay versatile and adaptive in a world too complex to map. It’s how you cash in on the absurd power of neural network math without ending up like Buridan’s ass, the perfectly rational donkey who starved between two identical piles of hay because he had no way to break the tie to decide which to eat from.
In biology, convergent evolution is when unrelated lineages arrive at the same solution on their own. The last ancestor we share with the octopus was some worm-like thing with light-sensitive patches at best, and yet – both of us ended up with a camera eye. Simply because for seeing in detail, there aren’t that many good designs to choose from.
I think predict-and-pick is the camera eye of agency.
It’s mostly simple and obvious to have one system presenting options and another selecting from them. But this specific articulation could be shaped by real-world evolutionary dynamics, such as:
In an open-ended environment the same situation rarely has one right move, so a system that outputs a single answer has to throw away everything it knows about the second- and third-best ones, while a predictor that outputs odds gets to keep all of it, storing its uncertainty together with its knowledge.
Anything that reacts to the same situation the same way every time can be “farmed” by whoever figures out the pattern.
The split keeping everything expensive in one place and everything adjustable in another, as all the learning lives in the predictor, while the picker is a few lines of code with one knob. This means the same network can be made careful (T=0.7) or adventurous (T=1.5) without retraining any weights. Brains appear to have the knob too – the leading account of the noradrenaline system says its purpose is to shift us between exploiting what we know and exploring the alternatives.[6]
Take this with a grain of salt – it’s a mere preamble for further exploration.
So what’s the difference?
The implementations of humans and LLMs of course differ in many ways, mostly because they grew up in different worlds. A human is shaped by physics – one body, one location, one irreversible timeline, no backups. An LLM lives in the digital world where everything is fungible, copyable and modifiable.
But what I want to unpack here are the differences in the picker.
In an LLM, the picker is a few lines of code sitting outside the network, fed by a PRNG, whose numbers were fixed the moment we set the seed.[7]
In a human, we are unable to even pin it down at all – it’s dissolved into the noise of a hundred trillion synapses,[8] fed by physical randomness whose ultimate source nobody understands, and it even extends into the realm of subjective experience, with the stubborn first-person impression that someone who sees the whole picture is doing the picking.[9]
If we strip away all the details, we can narrow this down to two key differences.
The first is consciousness. There is something it’s like to be the one choosing between another beer and an Uber, and nothing about a token sampler suggests there’s something it’s like to be a call to a random function. I think this difference is real, but it’s also one I can’t do anything with, because the only evidence for consciousness anyone has ever had is their own, and everything beyond that is taking someone’s word for it – a courtesy we extend to other humans mostly because they’re similarly clueless. So I’ll leave consciousness sitting right here, acknowledged, and maybe we’ll get to it someday.
The second difference is in the randomness function. The numbers feeding an LLM’s picker come from a defined formula – fix the seed and you can read the model’s whole future like a printed timetable.
The noise feeding your picker bottoms out in quantum events, where physics gives us the odds and then goes completely silent: no formula, no seed, and as far as a century of experimental physics can tell, nothing anywhere in the universe that fixes the outcome before it happens.
You might not be able to tell these apart statistically, but they are different kinds of things – one is a calculation and the other is an event.
So once you subtract the part we can’t measure, the entire gap between a brain and a chatbot comes down to a single input wire. On the chatbot’s side, we know exactly what’s on the other end – a config file.
On ours, the wire runs down past our synapses and microtubules into the one place in physics where the question “Who tf picked that?” has no answer.
Nobody knows what’s on the other end of that wire.
Maybe nothing. But it’s exactly where the universe could have a plug for a controller…
I’m getting ahead of myself. Let’s save the nutritious stuff for part two.
Mathematically, the network’s raw score for each token gets divided by the temperature before the scores are turned into probabilities (with a function called softmax). Temperature zero can’t be a division, so it’s defined as the special case of always taking the top token.
Similarly, even if humans were “greedy samplers” (the name at least seems quite fitting) who always pick the highest-probability option, you can’t dismiss the voice saying “I should really call an Uber and get some sleep” while you proceed to order the next round of beers. This shows that (A) more than one option is generated and presented, and (B) the action taken isn’t always the evolutionarily optimal one.
Yes, in practice an LLM often gives different outputs between runs even with a fixed seed, because floating-point rounding, batching and parallel GPU execution all leak into the result (see Thinking Machines Lab, “Defeating Nondeterminism in LLM Inference,” 2025). These are of course fun sources of randomness to look into, and if you do you’ll find that this rabbit hole goes deep, very deep, and that perhaps it even ends in the same quantum-random gap between worlds, where all the roots, fingers, strands and strings of causation coalesce into one.
Again, if you suppose humans always perform the highest-probability action, then manipulating the randomness underlying the entire processing of the network (quantum entropy) could influence which action gets computed as the most likely or optimal one. That would remove the need for the clean network/sampler split we have in LLMs.
Both your brain and an LLM first evaluate options and then pick one. Could this be a universal formula for agency?
Note: While this post was not written for the LessWrong audience, I believe it might be of value to people here, and that it meets the quality bar even though my writing is not strictly lesswrongian.
It's the first post of two or three part series, getting into the technical details how both LLMs and brains depend on randomness, and how that might help us understand how agency and causation work in our reality. This one sets the stage, giving a rundown of the technicalities for a wider audience, and introducing the concepts of predictor and picker – high-level architectural features that intelligent agents seem to converge on, a sort of "camera eye of agency" if you will. I'm quite confident that the angle I present here is something that might have been circulating in some minds, but hasn't really been put into words.
Did you know that LLMs do not work by selecting the most probable next word? What they do instead is predict the probabilities of many possible next words, and then pick one of them with a randomness function. That mechanism – token sampling – is what enables them to sound like humans instead of generic mechanical language synthesizers.
In this piece, I’ll explain how this points to a general architectural pattern of “predicting and picking” that human brains use too, and why that pattern is so versatile in worlds too complex to map.
And in the next part, I’ll explore how this simple architectural setup could be the perfect interface into an entirely different dimension.
Randomness is a design choice
If you’re building any kind of system, you have a design choice to make: how much error, noise, or randomness you want in it – and it’s a spectrum. At the zero end sits the mathematical precision tool – supremely accurate, doing exactly what it’s meant to do, every single time.
Somewhere along the way into the digital era, we came to equate that end of the spectrum with quality itself. We optimize for efficiency and precision, and we always want the perfect, sharp cutting tool that does exactly what we want – full determinism.
But I’d argue that if the universe and nature were to work that way from the start – if life itself were a system with no errors – there would be no evolution, no progress, not even any fun really. Everything would be sterile, stagnant, unchanging, and if some self-replicating system did happen to arise inside that paradigm, it would stay exactly as it was, with no mutation, no adaptation and no real progression. Randomness, or entropy, is an essential ingredient of all living systems, and without it they wouldn’t be what they are.
Now, the systems that form the foundation of our modern society, we’re building in the digital world, which is something we constructed to be the epitome of determinism: no noise, no error. The lengths we go to in order to create that are extraordinary (chip error rates can approach single-digit FIT – failures per billion hours of operation), because the possibilities it opens up are of course extraordinary too (mostly Ctrl+C, Ctrl+V and everything in Daft Punk’s Technologic).
But regardless of whether it was a conscious design choice (pun not intended!), the universe we live in and from whence all life emerged has a non-zero amount of randomness throughout.
In the physical world, randomness is the default: most systems are nondeterministic at their core, yet with some effort, deterministic-like systems can be built. In the digital world, it’s the opposite: determinism is the default, yet with some effort, random-like systems can be built.
And random-like is as far as it goes, because digital systems are almost never actually random, only pseudorandom – they use extremely complex functions that appear perfectly random for all practical purposes, but are fully predictable to anyone who knows what’s called the “seed” of that function. Sounds crazy, confusing? Yes! These pseudorandom number generators (PRNGs), as these math functions are technically called, are literally deterministic clockworks dressed up in dice costumes.
How an LLM picks its words
Finish this sentence: “I like…”
Statistically, you are most likely to say… pizza. Or cats. But there might (soon?) come a time when the wisest thing to say is: “I like our alien overlords and serve them humbly, no further probes will be necessary, thank you.”
There is clearly no single correct continuation, as it depends on the situation, on who’s speaking, etc. This happens most everywhere in natural language, which makes it a terrible fit for machines that map each input to one fixed output. The same goes for acting in any environment too complex to map, where there’s rarely one right move and usually at least several decent ones (along with infinitely many bad, and very bad moves!).
This is why every system we’ve found persuasively humanlike so far is built from two parts: a probabilistic engine that takes in the whole context and narrows an infinity of possible next moves down to a shortlist of weighted options, and some mechanism that picks one option from the shortlist and makes it real. Let’s call them the predictor and the picker.
Note that the predictor never predicts what comes next, only how likely each option is, and it’s the picker that turns those odds into the one thing that actually happens.
In an LLM, the predictor is the neural network – billions of parameters that read everything written so far and output a probability for each of a hundred thousand or so possible next tokens (a token is a word or a piece of one). The picker is the token sampler: a few lines of code that take a random number between 0 and 1 and use it to draw one token according to those probabilities. “Pizza” at 42%, “cats” at 17%, “aliens” at 1% – roll, print, append to context, next probabilities, repeat.
Fig 1. Token sampling: lay the probabilities end to end on a line from 0 to 1, draw one random number, and print whichever token it lands on.
Between the predictor and the picker sits a knob that you might know – temperature. But it doesn’t touch the random number at all – it just reshapes the shortlist before the roll.[1]
At temperature zero, the whole probability mass is dumped onto the single most likely token and the picker has nothing left to pick. At temperature one, the picker draws from the distribution exactly as the network computed. If you crank it up, the distribution flattens (and “aliens” start showing up rather uncomfortably often!).
Fig 2. Same network, same context. Temperature only reshapes the shortlist – T=0 collapses it to a single option, higher T flattens it toward equal odds.
So yes, you can lose the picker and always print the top option. This is called greedy sampling (or temperature zero), and its output becomes clunky, machine-like text riddled with stock phrases, clichés and other LLM-isms. With the earlier models it’d often degenerate into endless loops; with today’s models your garbage detector simply instantly smells slop, and any “big model smell” definitely evaporates. There’s even a paper explicitly spelling it out[2] (as if it wasn’t like, totally obvious): human text itself doesn’t stay in any statistical high-probability zone. People keep choosing words that are likely enough without being the likeliest, and a model forced to always say the most probable thing is throwing away information and presenting a superficial take at best.
Alas, the randomness in LLMs is load-bearing. Especially for anything that you want to interact with on any sort of intellectual or personal level – not to say let it write your high-effort Substack posts!
Like human, like LLM
Now let’s look at the other system known for producing persuasively human language.
No, not the reptilians in our government… You!
Your own predictor is also a neural net – in fact, it’s the OG neural network; your brain. While our vocal cords can produce a sweeping array of sounds across an impressive range of frequencies, our brain narrows it down for us; there always seems to be a continual shortlist of just a handful of words appropriate in a given situation.
We can even trace that objectively. In neuroscience, Paul Cisek’s affordance competition hypothesis[3] describes the primate cortex as doing two jobs at once: specifying several potential actions in parallel, and letting them compete – biased by everything else the brain knows – until a single one gets selected. Record from a monkey’s premotor cortex while it’s deciding between two targets to reach for, and it appears we see both options being held at the same time before one of them wins.[4]
But then… What makes thy picker?
When asked “Want another drink?” the processing of your options (and their tomorrow morning’s consequences) is mechanistically traceable – but what is the sum of the equation that ultimately decides whether you say yes or no?
This strikes at the very core of the free will debate, which we won’t settle here. But suppose there is some space (let’s call it consciousness), which holds and considers the shortlist of options. This consciousness then appears to make a conscious decision to actualize one of them.
Whether there’s something (beyond the brain) doing the deliberation is debatable – the predicting and picking are not.[5]
So the recipe is the same for both humans and LLMs: a neural network that compresses the entire situation into a few weighted candidates, plus something that commits to one of them (whatever that is).
That might be a coincidence. But it could also be convergence.
The camera eye of agency
I suspect the predictor-picker combo is something close to a universal formula for entities that need to stay versatile and adaptive in a world too complex to map. It’s how you cash in on the absurd power of neural network math without ending up like Buridan’s ass, the perfectly rational donkey who starved between two identical piles of hay because he had no way to break the tie to decide which to eat from.
In biology, convergent evolution is when unrelated lineages arrive at the same solution on their own. The last ancestor we share with the octopus was some worm-like thing with light-sensitive patches at best, and yet – both of us ended up with a camera eye. Simply because for seeing in detail, there aren’t that many good designs to choose from.
I think predict-and-pick is the camera eye of agency.
It’s mostly simple and obvious to have one system presenting options and another selecting from them. But this specific articulation could be shaped by real-world evolutionary dynamics, such as:
Take this with a grain of salt – it’s a mere preamble for further exploration.
So what’s the difference?
The implementations of humans and LLMs of course differ in many ways, mostly because they grew up in different worlds. A human is shaped by physics – one body, one location, one irreversible timeline, no backups. An LLM lives in the digital world where everything is fungible, copyable and modifiable.
But what I want to unpack here are the differences in the picker.
In an LLM, the picker is a few lines of code sitting outside the network, fed by a PRNG, whose numbers were fixed the moment we set the seed.[7]
In a human, we are unable to even pin it down at all – it’s dissolved into the noise of a hundred trillion synapses,[8] fed by physical randomness whose ultimate source nobody understands, and it even extends into the realm of subjective experience, with the stubborn first-person impression that someone who sees the whole picture is doing the picking.[9]
If we strip away all the details, we can narrow this down to two key differences.
The first is consciousness. There is something it’s like to be the one choosing between another beer and an Uber, and nothing about a token sampler suggests there’s something it’s like to be a call to a random function. I think this difference is real, but it’s also one I can’t do anything with, because the only evidence for consciousness anyone has ever had is their own, and everything beyond that is taking someone’s word for it – a courtesy we extend to other humans mostly because they’re similarly clueless. So I’ll leave consciousness sitting right here, acknowledged, and maybe we’ll get to it someday.
The second difference is in the randomness function. The numbers feeding an LLM’s picker come from a defined formula – fix the seed and you can read the model’s whole future like a printed timetable.
The noise feeding your picker bottoms out in quantum events, where physics gives us the odds and then goes completely silent: no formula, no seed, and as far as a century of experimental physics can tell, nothing anywhere in the universe that fixes the outcome before it happens.
You might not be able to tell these apart statistically, but they are different kinds of things – one is a calculation and the other is an event.
So once you subtract the part we can’t measure, the entire gap between a brain and a chatbot comes down to a single input wire. On the chatbot’s side, we know exactly what’s on the other end – a config file.
On ours, the wire runs down past our synapses and microtubules into the one place in physics where the question “Who tf picked that?” has no answer.
Nobody knows what’s on the other end of that wire.
Maybe nothing. But it’s exactly where the universe could have a plug for a controller…
I’m getting ahead of myself. Let’s save the nutritious stuff for part two.
Mathematically, the network’s raw score for each token gets divided by the temperature before the scores are turned into probabilities (with a function called softmax). Temperature zero can’t be a division, so it’s defined as the special case of always taking the top token.
Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes and Yejin Choi, “The Curious Case of Neural Text Degeneration,” ICLR 2020.
Paul Cisek, “Cortical mechanisms of action selection: the affordance competition hypothesis,” Philosophical Transactions of the Royal Society B, 2007.
Paul Cisek and John F. Kalaska, “Neural correlates of reaching decisions in dorsal premotor cortex: specification of multiple direction choices and final selection of action,” Neuron, 2005.
Similarly, even if humans were “greedy samplers” (the name at least seems quite fitting) who always pick the highest-probability option, you can’t dismiss the voice saying “I should really call an Uber and get some sleep” while you proceed to order the next round of beers. This shows that (A) more than one option is generated and presented, and (B) the action taken isn’t always the evolutionarily optimal one.
Gary Aston-Jones and Jonathan D. Cohen, “An integrative theory of locus coeruleus-norepinephrine function: adaptive gain and optimal performance,” Annual Review of Neuroscience, 2005.
Yes, in practice an LLM often gives different outputs between runs even with a fixed seed, because floating-point rounding, batching and parallel GPU execution all leak into the result (see Thinking Machines Lab, “Defeating Nondeterminism in LLM Inference,” 2025). These are of course fun sources of randomness to look into, and if you do you’ll find that this rabbit hole goes deep, very deep, and that perhaps it even ends in the same quantum-random gap between worlds, where all the roots, fingers, strands and strings of causation coalesce into one.
A. Aldo Faisal, Luc P. J. Selen and Daniel M. Wolpert, “Noise in the nervous system,” Nature Reviews Neuroscience, 2008.
Again, if you suppose humans always perform the highest-probability action, then manipulating the randomness underlying the entire processing of the network (quantum entropy) could influence which action gets computed as the most likely or optimal one. That would remove the need for the clean network/sampler split we have in LLMs.