How does autopilot interact with multitasking for you? Do you remember what you did on autopilot?
My memory is generally worse, though I'll still remember broadstrokes.
I'm unsure what multitasking would look like outside of certain time-pressured jobs. I think I do not usually multitask, regardless of mental state. But it also looks like other people are like this, too?
I'd suspect that multitasking should not be made much worse by going on autopilot (beyond the way that individual tasks are), as task-switching is already going to require dumping my working memory. I'd be slower at each step, though, and so if the reason why I'm multitasking is because I need to attend to multiple types of events that need to be processed quickly, then I'd expect to be much worse.
I mean, if your autopilot is like when I tell Claude to spawn a background subagent, then you should be able to do something else at the same time. Can you?
I seem to basically have one verbal 'stream', so that I mostly cannot think in words while talking or think in words about two different things. If talking about something that I know werl and have talked about before, I will sometimes find myself doing a little bit of accessory planning (sometimes even in words!) while my mouth moves in coherent ways, but I sorta feel like if I focused on it that the spell would break. It's an interesting question, though, so I'll try to notice next time.
I can walk and talk, or drive and talk, easily. The only examples that come to mind of me talking and visualizing were times where I was describing what I was visualizing or reporting a conclusion from my visual thinking, as opposed to talking about something else.
I think my autopilot thinking is more like if Claude only had a context window spanning a dozen tokens.
TL;DR: Generating thoughts can be done on autopilot even while cognitively impaired (say, tired or drunk), as long as the serial mental depth is low. I speculate that this is basically because System 2 is System 1 + working memory. You may find it useful to try tasks on autopilot instead of assuming you cannot.
You may want to skip directly to the "In Practice" section. In short, tell yourself this: "you can do a lot more on autopilot than you think, even if it feels like you couldn't possibly get anywhere right now". Next, try to think on autopilot anyways.
If it seems not to work for you, you should stop.
Epistemic Status: Many weak lines of evidence, with overall medium certainty.
Basic Theory
System 2 is System 1 with Working Memory. System 1 is atomic action.
There's a theory that there are two reasoning systems in the brain:
Type 1 processing is “efficient, unintentional, uncontrollable, and unconscious”
Type 2 processing is “inefficient, intentional, controllable and conscious”.
There's also a theory that the second system is really just the first system augmented with working memory. At each step, the working memory is read, a System 1 operation occurs, and a result is written to working memory.
As an example, when I play chess, I have some immediate sense of what makes for a good move. I will also "calculate" in advance, by imagining making a promising move, then imagining various promising responses by my opponent and seeing how good they seem (or iterating, to judge their responses by how promising my responses to their responses look, ad infinitum until I run out of ability to keep track of stuff in my head or until I feel like it's not worth thinking for longer).
As another example, when I write a couple paragraphs of argument in a text to a friend, it feels like I am constructing sentences in my head, writing them in my mind or to my phone, and then I either:
Each decision is made intuitively, in what feels like an 'atomic' step. I can introspect on the overall process, or reason out plans that I then execute on - but I have much less insight into the atomic processes that generate the next thought or judge previous thoughts by some criterion.
What does this mean? It means that atomic babbles and prunes can be done on autopilot. (Autopilot pruning feels more obvious, but it surprised me to realize that it's also true of babbling).
In Practice
Even when cognitively impaired (e.g. sleepy or drunk), I can still have thoughts on autopilot. They will even be surprisingly good thoughts! Such afflictions tend to make it feel like my working memory was tossed out the window, and so can give me the impression that I am entirely incapable of useful thoughts - yet, if I just try saying shit anyways I'll somehow get good stuff out.
One example that I've trained myself to do: when down, ask "what should I do next?", start strategizing about whatever problem I'm having, consider whether I should take a nap or if some food and water will make me feel better in an hour, and similar. It will feel like I cannot keep up with my chain of reasoning, and so should be wary - but I have learned that it will still be remarkably trustworthy. Apparently, concepts like opportunity cost and the sunk cost fallacy are sufficiently ingrained in me that they're invoked on autopilot.
Sometimes I find tasks aversive, and struggle immensely with starting them. It feels like I involuntarily flinch away from loading the necessary parts into my working memory.[1]
However, I tend to not find it as hard to direct my 'autopilot' thoughts towards whatever I want done, and just use whatever comes out of my atomic generative processes. An improv sketch, instead of a proof-read script.
Ideally, the task will stop being aversive after a few minutes, and I'll be able to bring the full brunt of my mind to bear... but that only actually worked a couple times.
Personally, I've found it helpful to tell myself the following:
"You can do a lot more on autopilot than you think, even if it feels like you couldn't possibly get anywhere right now"
What's easy and what's hard?
The easiest thoughts are conceptual cache lookups that don't require rethinking. As an example, I can recall mathematical definitions and theorems that I know well enough, or arguments that I've thought through enough before. However, even simple novel applications can be difficult.
This suggests that the intellectual benefit others get from talking to me is preserved better (relative to the intellectual benefit of my thoughts to myself): if I have something cached that is novel to them, then I can easily tell them about it.
Truly new ideas are much harder. I have so little insight into how these happen that I don't know how they are affected by thinking on autopilot.
Secondly, it's important that I'm still calm. If I was, say, having a panic attack, then good reasoning is much much harder. I suspect this is because it involuntarily directs my attention to the wrong thing - instead of strategizing, I end up spiraling about some exaggerated catastrophe, which also reduces my ability to notice what's happening and break out of it. For this reason, I find it easier to handle feeling apathetic or empty, because then I can notice that I'm in an unusual mental state and think through recovery actions. Anger, fear, or despair can force attention to the wrong places.[2]
Thirdly, my atomic thoughts are still clearly affected by how tired I am, or various contextual factors. I am more likely to think of things associated with what I have recently encountered or thought about. How vulnerable my intuition is to, say, the planning fallacy, strongly depends on how energetic I'm feeling this moment and whether I've just had an exciting conversation with a friend.
Fourthly, the theory I put forth implies that anything that requires working memory is harder or out of the question. When tired, I often find it hard/impossible to visualize[3], to keep a full sentence in my head, or to do arithmetic for Fermi estimates. I have had limited success mitigating this by writing more of my thoughts down[4], or prompting whoever I'm talking to to manage the conversation.
Lastly, everything requires more effort when tired, and I find it harder to direct my attention anywhere in particular.
More Theory
Internal Monologue
I happen to have an internal monologue.[5] I used to think most of my thinking happened in it, that language was near synonymous with thought.
Now, however, I think that what's mostly happening is that I have a concepts-to-language translator that writes and reads from the internal monologue. When in conversation, I don't have an internal monologue![6] It feels like I am 'writing' to my mouth, instead of to my phonological loop. While I still think that much of the (evolutionary, cultural, practical) value of language is for thought, I think it mostly serves to make thoughts legible and temporarily store them.[7]
Dreams
For a few years, math has occasionally shown up in my dreams. Usually, it's utter nonsense that my dream self thought was coherent.[8]
More recently, coherent math has shown up. I think I once correctly recounted the definition of a sigma-algebra when asked by a professor I was trying to impress in my dream, and I remember a times where my waking self judged my dreaming self to have recalled true facts.
But one time... one time, I learned math in my dream. In what is still one of the most surreal events in my life, I had a dream involving a magical goddess that was demonstrating an idea with holographic visuals while explaining it.
If you're interested in the math, or external sources that served as confirmation that the concept was coherent see the footnote.[9]
I say I 'learned math' instead of 'came up with' because it did not feel like I thought of the idea. If I weren't so confident in my atheistic materialism, I could see how I might've interpreted it as a religious experience. After all, I literally saw a goddess show me math I had never seen before!
Perhaps it is not a coincidence that Ramanujan once saw a hand writing elliptic integrals on a screen of flowing blood in a dream?[10]
I suspect that I mostly don't have access to my usual working memory abilities in my dreams. Maybe this explains why the insight feels external?
If true, then the cognition in dreams might give us evidence about what waking autopilot cognition can do. At the very least, my weird experience suggests that my unconscious can do a lot of sophisticated reasoning!
LLM Chain of Thought
I suspect that LLMs are, on a moment-to-moment basis, not that impressive. However, they have a lot of crystallized knowledge, and you can scale up their speed of thought and run parallel instances to do things like code up a webapp or search the internet for an obscure blog post.
Thinking on autopilot is sorta like being an LLM that's fed only a few of the last parts of the chain of thought to generate the next step.
Am I Deluding Myself?
An alternative explanation is that I am merely deluding myself - perhaps my reasoning is much worse but my ability to evaluate my thoughts is also degrading, leading me to not see the ways in which I've degraded.
I think the previously mentioned theory provides many lines of evidence. Each one is fairly weak, but they are somewhat independent and so aggregate to moderate certainty in my mind.
I have yet to accumulate much hard empirical evidence. I know I've correctly recalled mathematical definitions while cognitively impaired before, and my thoughts seem to hold up upon later reflection.
Conclusion
Guess how I wrote this post?
That's right, half-tired on semi-autopilot. Part of why I could do it is that I had already thought through the core ideas before.
If you liked reading, then you have evidence for the method. Try it yourself!
If you have any of your own tricks for thinking when tired, or get different results, I'd like to hear what you have to say. For example: do you also feel like your working memory is the first thing to lose capability? Do you find it harder to (say) visualize, but find verbal tasks just as easy? How good do you think your working memory is in your dreams?
Lastly, my main criticism of Tuning your Cognitive Strategies is that it mostly doesn't do a good job at explaining how to actually tune your cognitive strategies. I see that post as about training atomic operations. Unfortunately, I also don't have any explicit ideas about how to do that!
Instead, I am pointing out a way you can use cognitive strategies that are already good enough. The more uncannily accurate intuitive motions you've already trained, the better I expect you'll do on autopilot.
I'd say it's currently the biggest problem in my life. If you happen to have experienced this in the past and found that, say, reciting Shakespeare on one foot fixes it for you, then I'd really appreciate it if you told me.
They can, of course, direct attention correctly! I think this is best done when they merely nag/remind me, and inform me, instead of hijacking my entire thought process.
Normally, I have no problems visualizing. It's like I drop down a couple levels on the sliding scale of aphantasia.
I suspect the overhead of manipulating the physical object, as well as the ways that I often cannot draw what I would otherwise imagine, limits the success. Also, ultimately you still need enough RAM to store the input to the function you're calling at each step.
Not everyone does!
Unless I've paused to think something through first, or are searching for a better phrasing before speaking.
And, of course, to communicate them to others.
Once, there were colored billiard balls used as symbols in what vaguely resembled algebra, with no more sense to it beyond that.
Imagine taking a picture of an object in a curved spacetime (just imagine a curved surface). Light travels along geodesics, and so instead of projecting along straight lines I should project along the geodesics.
Here are some slides that work with a formalization of that concept, and a paper dealing with some version of it for de Sitter spaces.
(I skimmed the slides and read a couple paragraphs of the paper - so I can't say much about the math in them).
From Wikipedia: