Writing

Do LLMs think like us?

Or do we think like them?

Danish Farid

· 801 words · 3 min read

I’ve been thinking a lot about thinking in this new agentic, LLM-fueled, Jensen-powered, vibe-based era. More specifically, I’ve been thinking about how some inner workings of LLMs mirror, and reveal things about, my own thought patterns and approaches to problems. What can we learn from them about ourselves, and how do we make the best use of this knowledge?

For example, when faced with two similar situations I might take two entirely different approaches to solving them, resulting in vastly different outcomes. This difference can sometimes be attributed just to how I felt in the moment. A psychology major might be able to explain this perfectly, but I’ve only ever taken one psych course in my life. It was only after LLMs and the idea of “context” that I started to question why I would behave differently in the two situations.

Funnily enough, the last time I thought deeply about my own behaviour like this was during COVID, in the distant town of Waterloo, Ontario. A few months into social distancing, I increasingly started seeing and analyzing myself in the 3rd person, like a character in The Sims game, complete with need meters: hunger, sleep, thirst, etc:


"Parched, he should drink water not coke zero!"

"Hungry (again), eat an apple."

Hunger, thirst and sleep fit pretty well into the need meter analogy, but I started doing this with feelings, behaviours and social contact as well, concepts that don’t fit quite as neatly into a ‘need meter’.


"Mental levels in the drain. Maybe he needs to go walk outside?"

"Social interaction slipping to dangerous levels, get him on the phone soon."

Bewildering, but it did explain why I could never think my way out of a funk and usually needed some kind of external adjustment. Turns out this wasn’t just a novel discovery on my part, psychologists call this ‘distanced self-talk’. Referring to yourself in the third person can sometimes make it easier to regulate emotions and reason through stressful situations.1

After noticing similarities, my first question was whether third person ‘pep talks’ are to humans what “priming” (role prompting) was to these language models. For a while, prompting best practice was to assign a role to chatbots up front: “You are a highly experienced C++ performance engineer, I need you to look at…”. Sometimes this would help improve the output.2 With vastly improved models though, it’s not really required anymore, but something similar can still benefit a (mere) human! There is even a version of this tested with kids called the “Batman effect”, where children stuck with a boring task longer when they pretended to be Batman.3

That said, telling myself I can lift a car or solve Navier-Stokes overnight is unlikely to have the desired effect. Where it could potentially have an impact is at the edge of things I can already do: lifting a little more in the gym, running through shin pain a bit longer, finding a bug in latency-critical code faster, etc.

Further similarities:

  1. Our “context windows” fill up the way they do for large language models, and filling them up too much can cause confusion and forgetfulness.
  2. We like to use tools to organize ourselves (calendars, journals, to-do lists), but using too many can add more noise than value and can end up confusing us.

Some quick research, however, reveals entirely different systems underneath these similarities: mind-body connection and overcoming emotional and biological barriers (Alter Ego effect) vs a model managing statistical probabilities. Obviously, we are not transformer based entities, but perhaps this line of thinking can still lead us to how we “game” ourselves into doing better? What can we learn from the industry putting millions of person hours into doing this with LLMs?

One area where there is essentially no overlap is the ease with which context is cleared. When an LLM starts acting up, it’s usually time for a fresh context, I wish I could do this in my head as cleanly. Sleep, a walk or a shower get you part of the way there, but I wanted a better tool, so I built one: an iPhone app called ‘Dump it’. Why do I not actively consider and decide my own context window throughout the day, like I do with my agent sessions? Surely I too would be more focused. This app was an experiment to try something like this out.

Do we function like LLMs? No.
Are we similar in some ways? Maybe.
Could some best practices with language models help a human out as well? Almost certainly.

In any case, the future is mostly uncertain, but one thing is for sure: I’ll keep thinking.

Footnotes

  1. Kross, E. et al. (2014). Self-talk as a regulatory mechanism: How you do it matters. Journal of Personality and Social Psychology, 106(2), 304–324. ↩

  2. Kong, A. et al. (2024). Better Zero-Shot Reasoning with Role-Play Prompting. Proceedings of NAACL 2024. ↩

  3. White, R. E. et al. (2017). The “Batman Effect”: Improving perseverance in young children. Child Development, 88(5), 1563–1571. ↩