The first time I built a freestanding bookshelf, I put a lot of effort into making the feet level and the back straight and at a right angle to the feet. Once I put it up against the wall I'd built it for, I realized I'd solved completely different problem than the problem I really had. I needed crooked bookshelf, since the wall was totally tilted.
In the end I screwed some wall shelves in and called it good enough.
One of my first real DIY projects during a summer in college nearly 20 years ago was replacing the rotted out basement bulkhead doors on the ~120 year old house I grew up in. I took measurements of the old ones, bought some nice tongue-and-groove cedar and high-quality hardware, and built the new doors in the garage. When they were fully assembled, I carried them over to install on the old stone frame. I took off the old ones, put mine in their place...and they didn't fit properly at all.
Momentarily baffled, I realized that, despite appearances, the old frame was actually not square, in fact it was a parallelogram. I'd measured the height and width and assumed it was square. The previous (experienced) carpenter who'd built the doors I was replacing had clearly noticed this, and simply allowed for the misalignment in his design. He built perfectly square-appearing doors that mounted to the not-square frame. I had to go back and rework mine considerably for them to fit without looking ridiculous. They're still there and holding up well, but I also still think of this lesson on a regular basis in my day to day life now.
I feel this in my soul. I thought I could replace a door in a day, but months of fiddling, I discovered by frame is not only a parallelogram but it literally shifts by over an inch between seasons. (~100yr old house 2nd floor)
Sounds about right. I approached this in a different way in my office.
The walls aren't straight, either vertically or horizontally, and they're not even consistently wonky along any given axis.
So I installed uprights vertically, using transparent polycarbonate spacers of different depths at the attachment points[0]. I then installed shelves on the uprights and aligned them horizontally.
The variation is only +/- 6mm or so (for around 12mm variation across the 2.5m x 2.44m wall) but, if I hadn't done this, my shelves wouldn't be level, and wouldn't even be consistently non-level, so would have been awkward to install along the full length of the wall, would all be misaligned with eachother, and would have looked incredibly janky.
[0] In hindsight I wish I'd gone for these in different colours rather than just plain transparent, to make more of a feature of them. The walls are white so I think orange, blue, red, and yellow would have worked well.
Hah, yeah same. I grew up in a house built with hand tools sometime around 1910 (family bought it from an old lady whose father built it), not a single corner was square (though things were generally good vertically by some miracle), but it wasn’t noticeable until whenever were doing major work.
Also learned that lath and plaster needs some special consideration when screwing/nailing things for securement, as the lath (wood strips) could split, causing a subsequent crack in the plaster. Basically for screws or bigger nails, it’s a good idea to drill a small hole first to lessen the pressure, or do a bigger hole and use a spring bolt anchor.
I have been getting a lot more into DIY and that's my experience myself. I keep running back to the store because there is some detail I haven't considered. Iteration time is so much slower than software... kinda bugs me how much you have to plan upfront and think through instead of just YAGNI and agile-ing your way through it.
A related thing that took me a while to accept when I started woodworking is that wood moves, a lot.
If you built the bookshelf in wood, it will be expanding, contracting and shifting over time with temperature and humidity variation throughout the day and season. And asymmetrically depending on the grain.
The straight right angles won't stay that way, and it's better to design such that they change in complementary ways, rather than remain perfect.
The country home I grew up in had uneven floors and door frames.
Then I studied abroad in Italy and realized the phenomenon was global.
Then I moved to San Francisco and realized it was a joke at this point.
Building codes are a joke!
Brunelleschi drank wine all day. Sure he basically invented pulleys and designed the sickest dome in human history, but he was a drunk and I can't blame him because architecture and all of math is a joke if you look close enough.
In ancient city of Rome, it was more expensive to rent a room in multi-story building if the landlords lived in it as well. It showed that landlord was brave enough that the building will not fall one day onto themselves.
same rule applied during housing boom in taiwan in the 1970s, my parents bought an apartment as the construction company had their headquarter in the very same building on the 2nd floor!
Coincidentally just had this realization last night. Leaned a piece of furniture against the wall, realized the ~perfectly straight/level edge didn't lean smoothly against the wall -- the wall is not perfectly straight!! :-O
I recognized this submission from its title but did not remember what it was about. For some reason this anecdote reminded me. Yes now I know it's about the man who built staircases with his father.
When I worked as a camera guy on film sets, this was a typical occurrence. You level out the camera trypod with the magic eye on the tripod. The magic eye being a small amount of liquid with a single bubble inside, pointing always upward.
Soon you realize that an surprising amount of walls are just not straight or level.
This is what molding is for. A lot of people view it as "ornate" or old fashioned, but it served a functional purpose originally and then people started making it fancier.
- you genuinely learn once your assumptions about how a system works break, you realize it, try differently, validate, get a better model of that system
- your interfaces must remain permissive while providing feedback, namely you provide wiggle room then only once it behaves roughly as expected do you tighten then up
> If you’re a programmer, you might think that the fiddliness of programming is a special feature of programming, but really it’s that everything is fiddly, but you only notice the fiddliness when you’re new, and in programming you do new things more often.
I think I'm drawn to programming because the fiddliness is tractable, and fixable.
In which other domain can I:
* introspect the relevant processes/state, step by step
* snapshot/undo
* fix niggles, once and for all, and for everyone; and get their fixes too
* probe and test my inputs and outputs, checking for quality. Get notified if a part changes in a way that breaks me.
And the only tool I need is a commodity general purpose PC.
When I try woodwork, or even electronics, I'm struck by much friction is in even simple tasks: tools, parts, lead time, safety, space, physical effort, cost, ...
Unless you have endless budget, many things can be one-shot. You can't do a test run first, or roll back a cut if the length is too short. You can patch misplaced nail holes, or re-dig a hole (messing up filling a hole with concrete is another matter) and hope you don't kill a tree transplanting it, but the end result isn't clean.
The best I could do with woodworking in the end to approximate programming was live with wasting some timber, leave a lot of margin on the main cuts and size all the pieces as a whole.
Woodworking taught me a lot about planning and design. As a young person, I was like the authors brother. I just wanted to do the thing, not draw a diagram and figure out how much wood I need, or build a fixture to mark the stair lines.
Woodworking (the more constructive, furniture-making kind), rewards a deliberate, controlled process and it savagely penalizes mistakes. Those lessons transfer well to other disciplines. I’d have been a much better student if I’d learned wood working in high school.
Woodworking was part of my first 3 years of high school, but it was mainly about learning safety and tool usage and not planning, estimating, selecting or purchasing timber.
These days I only want to go to the lumberyard once for a project. Learnt the hard way on my first project that you need to take the time to carefully select the timber - checking straightness, matching grain and also colour before I started. Major hassle and waste of time to have to go back to swap boards.
That's also a lesson about what people will sell you. First time I went to a lumberyard, I was (coincidentally) with a friend who did a lot of woodwork. I thought, well, I've just paid for a pack of wood, I'll get it. The worker there was completely happy with that. My friend stopped me, and inspected each piece.
Sure enough, several had cracks at the ends, knots in poor places, and other things that, had I bought it, would have caused me trouble.
I can be a naive person in that I assume good faith. I would never knowingly sell something poor quality to someone else. I had assumed because I was being sold it, it was okay.
They aren’t “knowingly selling you poor quality” as some sort of scam. They are selling you wood to the spec you asked for. If you want higher-grade wood, you either have to spend money getting lumber graded to a higher spec or spend the time going through piles of low-spec boards to find the good ones. Many engineered wood structures are designed to use “poor-quality” wood, and they prefer it because it’s cheaper than using less high-grade wood.
I think this is a very common sentiment among a lot of people, including me.
And also that’s why AI tools create mix reactions. A couple of months ago a post went viral which was really insightful on what I was originally drawn to cs.
I want to thank you for the following reasons:
- that's an amazing, mind-opening essay;
- I've been looking for this toy (Meccano) for my entire adult life. I remember playing with a set that my mom got me when I was a kid, but I didn't know what it was called, and neither did my mom. And when I describe it to people, everyone kept telling me it was Lego (Lego is a much better-known toy brand in this part of the world). I had pretty much given up on ever finding it, thinking it might have just been a niche toy that had ceased production. Words are not enough to describe my joy when I opened the link and saw images of Meccano pieces!
I have read this article already and "reality has a surprising amount of detail" has become a phrase for me. But, I read it again today because the writing is so good. This guy is a gifted writer.
One thing he didn't mention is getting the first and last steps to be the same vertical distance as the others. Nothing will trip you up (literally!) so easily as a final step that is a different height than the other steps.
I thought of this because this morning I was putting a small fence around some plants we want to protect from deer. The fence consisted of 20 sections (bought on Amazon), each about 24 inches wide. Our ground is like rock, and the fence was not that sturdy, so I had to pound a heavy spike into the ground to the depth of the fence posts, then pull the spike out and put each section's legs in, leaving room for the next section's leg to go into the same spike hole. I wanted to be sure I was putting each section in at the right position, lest I end up with a 12 inch gap and have to go back and adjust lots of sections. Long story short, I pretty much succeeded, although when it cools down I may adjust a few sections. But the problem was sort of like the stairs: I wanted an integer number of fence sections, each the same length, to exactly fit around the bushes---just like you want an integer number of vertical steps in a diagonal stair, each of the same (more or less standard) height.
With the complexity of the systems we build it’s humbling and refreshing at the same time to read this. Indeed we are quick to default to what we know. Yet the answer may lie well outside of our expertise. I often build things many consider impossible and find that exactly the attention to detail and seeking contrarian views has helped avoid the worst mistakes.
I think we have all written some code that looks bulletproof to us. We run a set of tests with all the inputs we can think of, and it passes with flying colors (after several iterations of course).
Then we give it to someone else and it fails on their first or second attempt. They simply tried to use it in a way that we did not anticipate. It doesn't mean that we are dumb for not thinking of those possibilities; it just means that we did not think of every one of them.
And it pretty much is most times. For us. But indeed I did run into the "why would they use it like that though" scenario, where it fails. So I have to patch that usecase. At which point I go all "ok now it's really done" until another fringe usecase pops up and so on and so forth. And I did think about how sure I was it was bulletproof, humbling moment.
Hanging a curtain rail that is centered both vertically and horizontally between the window and the ceiling, and it is leveled (with a bubble level), while in a ladder. Mark and then drill. It is harder than it seems at first. Maybe I'm just too a computer guy.
The fact that machine learning can learn highly detailed patterns is the very reason why AI is so useful. So what you’re saying doesn’t really make much sense
> The fact that machine learning can learn highly detailed patterns is the very reason why AI is so useful.
AI doesn't deal with reality, it deals with tokens. This is why all those vibe-coded harnesses, little more than glue between various text IO interfaces, are several hundreds of thousands of source lines of code.
It's why a SOTA model took 100kSLoK to write a C compiler to compile one specific project.
It's why, when I asked for a simple markdown -> ansi escape codes converter (for terminal output) in Python, SOTA Claude and SOTA ChatGPT both give me +- 150 SLoC when my own LUT-based version came to under 10 lines of code + a LUT.
Reality has a surprising amount of detail, but LLMs don't exist in reality, they exist in a virtual world made up off tokens.
The discretization of those tokens can be manipulated to get any result you want. If it meaningfully benefits the AI to have a more fine-grained discretization, then you can do that. AI only compresses as much as we want it to. I understand your sentiment, but the logical conclusion of what you’re saying is that no form of compression is ever valuable. That’s just not a defensible argument.
All information gets compressed. Even your own perception of reality gets compressed.
Do you exist in reality? Or just in a virtual world made up of sensory signals? Do you have access to the Ding an sich any more than a (multimodal) LLM?
How would you know? You have no external frame of reference; a virtual world of sensory signals would be identical from your perspective. (I agree that "reality" is the most parsimonious explanation by far, btw, but that's never been the point of the simulation thought experiment.)
I think the more interesting corollary of this article is that if we're living in a simulation, it's an impossibly, improbably detailed one. I really want some compute time on the HPC that's running it.
> How would you know? You have no external frame of reference; a virtual world of sensory signals would be identical from your perspective.
Okay, lets go with that :-)
I might be living in a virtual reality, correct, I have no way of knowing.
What I do know is that the reality I am in is many thousands of times higher in resolution than the reality of the LLM.
As an analogy, the LLM is seeing a downscaled 32x32 pixel image while I see the original 8k image. Whether there is a larger 1b^2 image that I cannot see is not relevant to the question of whether the LLM can see my reality or not - it can't.
Reality is by definition our physical reality, which is about an infinite number of levels more detailed than the, you know, _virtual_ digital world computers exist in.
Whatever world we construct for LLMs, no matter how detailed we make it, will always be a blocky projection of the real domain onto a virtual one.
It follows then that any insight gained in the virtual world is at best a rough approximation which can be quite useful at times but also utterly faulty on occasion.
How often it is useful vs. wrong is (partially) a function of how complete the real-to-virtual approximation for a given domain.
Certain domains, given their limited degrees of freedom, can be quite accurately modeled, such as a subway map.
But many domains cannot, and it's important to be aware of that inherent limitation in digital models including but not limited to LLM """reasoning"""
>Whatever world we construct for LLMs, no matter how detailed we make it, will always be a blocky projection of the real domain onto a virtual one.
I don't know exactly why but I never really understood this argument. Might be some kind of control thing? Because for me it's pretty simple, it's basically free to give access to reality. Just add "sensory organs" as it were. I can argue you can make them perceive reality even better than we (humans) do, just enlarge the audio/video spectrums. Bam...more reality. The whole point of the argument is we're missing information.
Again, I get the need for controlling the environment for what LLM/AI/AGI/whatever will be, but that will always cost more than giving them access to like...reality. Same reason I don't really believe in the whole simulation argument, it's just more expensive all around, loses resolution, let alone control. I don't doubt there will be some people that would indulge in neverending hedonism but not all people. You need to give up control for that.
> Because for me it's pretty simple, it's basically free to give access to reality. Just add "sensory organs" as it were.
I dunno what you mean by "free". The model is trained on text. To "give" the model sensory organs it would need to be trained on those sensory organs.
Current models can predict text, because that's what the weights represent. Models with sensory organs will need to be trained on the output of those sensory organs.
That sounds close to impossible in the foreseeable future.
In the spirit of the article, what detail in the decision making of layoffs might you be missing?
I expect there's a lot of detail that I'm unaware of relating to running a company (planning; risk; legal; ...) that might make a decision foolish to me, but make sense if given more context.
Contemplating the details of a thing is really satisfying. At times I find myself sitting there and trying to decompose the astonishing amount of work, research, both evolutionary and revolutionary progress that has gone into reaching the current level of something. Buying myself a coffee and stare at the local ferry and acknowledge that someones life's work went into figuring out how to make the paint stick to metal.
Naturally the other point also sticks.. I too often get stuck on the details. :P
> Surprising detail is a near universal property of getting up close and personal with reality.
> As you learn, notice which details actually change how you think.
Lovely article. The older I get the more I appreciate this.
One point worth making: in many cases, after learning to see & appreciate the details, you gain the power to ignore the details that don't matter to you. This can be quite freeing.
Reality does not have amount of details, it is infinite in all directions. Its only that we perceived certain amount of details, some more some less. One can spend their whole life mastering a single aspect and there always will be room to improve.
Yeah that's why all models are wrong and everything ultimtately comes down to intution like supreme court's I know it when I see it.
It's just that intuition doesn't scale and a lot of common cases can be handled with models, rules, definitions etc. People continue to be confused that just adding more rules eventually solves reality but it never has anywhere for anything so continuing to believe it will is wild
i like to think along these lines too. sometimes it's paralyzing. one of the biggest mind blowing facts is that all this complexity started at a point in time: big bang. no one explicitly programmed it. all of it came from a soup of elementary particles 13.8b yrs ago.
Yes but what about AI? (Perhaps the most annoying words written in the last few years mostly on LinkedIn).
But actually in the years since this was written, I do think the world has shifted. Doing things on a computer used to be really hard. Even just installing a framework or getting >python to call the right python on windows. Then install Django and get Django to work with nginx etc. It was just a lot of thankless, frustrating work to get from zero to 1%.
Aside from AI, the tools and packages and culture of computing has gotten better. But AI means you just get all the trivial but difficult stuff for free. And I think a lot of people who would have given up now make it through to see something work and they’ll feel the thrill of building something. It’s just better and easier now.
This is one of my all time favorite blog posts. Why? Because it strikes at something that is both true, and a huge trap for smart people. Specifically, people who are experts in one discipline will often imagine that something in some other discipline is "pretty straight forward." And yet, my experience is that it never is. But that doesn't stop smart people from promising something that turns out to be waaaaaay harder to do than they imagined it could be.
This hits for to me because I'm currently adding on to my house. Or rather, paying professionals to add on to my house, because I actually want it to get finished.
I visit every couple of days. It's REMARKABLE how fast things get done. One day, there were no walls. The next day, almost all of the walls were in place!
... and yet, at the same time, things take a long amount of time because reality has a surprising amount of detail. I haven't taken into account how much you have to do to frame a house. So incredible amounts of work get done, day after day, but 3/4 of them are things I had no idea needed to get done! Gazing up into the roof, the detail is incredible. The PSL beams, the brackets, the joists, the trusses, just.. EVERYTHING!
I thought the structural engineer's plans had an incredible amount of detail on them, and they do, but they also don't really say anything about _how_ to build the thing. How to put up the walls, how to hold them together temporarily, how to lift beams into place. In what order things can and should be done. That all just takes experience.
I agree, that conclusion made me reflect on my existence. What I think I know versus the infinite amount of detail I'm missing. It's good to be reminded that there's so much we don't know that we even don't know.
This has always been the fun part of programming to me. I know most people hate it, but I really don’t mind being on-call (ok I hate being woken up) and fixing weird bugs that users run into. All these small edge cases that people run into because reality is odd. Of course I’m in scientific programming so that probably colors my view.
It’s always a little disappointing to me when I think I’ve run into something unique but it ends up being user error or something.
I echo this. The kind of entropy that real users bring has been refreshing to face as a founder.
Being a founder has a lot of SRE like activities. Fortunately I used to actually like troubleshooting and hence love being a founder but I know a lot of people quit this path because of the "suprising amount of details" in reality!
Based on what is the level of detail to reality suprising? To me suprising means mysteriously or improbably unexpected. Why should we expect reality to be simple. Note complex and simple are somewhat subjective. The human brain evolved to just sufficient baseline level be able to handle the level of complexity of reality. So why would it be unexpected that humans find realty complex when our brains are calibrated just enough to handle it.
a good read indeed! Makes me think about my use of coding agents differently, as the main thing they do is deal with a lot of details that matter to the execution but don't matter to me personally enough to figure them out. Would love to see this author's more recent take as this was written pre-LLMs taking over the world.....
That sort of abstraction has always existed, it's just been a matter of hiring experts or labor from other humans. Reality still has a surprising amount of detail. You deal with it by engaging directly, delegating to someone else to engage with, or just brute force your way to a crooked staircase.
When you hire someone to work on the stairs for you, you /hope/ they know what they're doing, especially if you don't have the skills yourself to judge their work. Same for an agent.
In the end I screwed some wall shelves in and called it good enough.
Momentarily baffled, I realized that, despite appearances, the old frame was actually not square, in fact it was a parallelogram. I'd measured the height and width and assumed it was square. The previous (experienced) carpenter who'd built the doors I was replacing had clearly noticed this, and simply allowed for the misalignment in his design. He built perfectly square-appearing doors that mounted to the not-square frame. I had to go back and rework mine considerably for them to fit without looking ridiculous. They're still there and holding up well, but I also still think of this lesson on a regular basis in my day to day life now.
The walls aren't straight, either vertically or horizontally, and they're not even consistently wonky along any given axis.
So I installed uprights vertically, using transparent polycarbonate spacers of different depths at the attachment points[0]. I then installed shelves on the uprights and aligned them horizontally.
The variation is only +/- 6mm or so (for around 12mm variation across the 2.5m x 2.44m wall) but, if I hadn't done this, my shelves wouldn't be level, and wouldn't even be consistently non-level, so would have been awkward to install along the full length of the wall, would all be misaligned with eachother, and would have looked incredibly janky.
[0] In hindsight I wish I'd gone for these in different colours rather than just plain transparent, to make more of a feature of them. The walls are white so I think orange, blue, red, and yellow would have worked well.
Also learned that lath and plaster needs some special consideration when screwing/nailing things for securement, as the lath (wood strips) could split, causing a subsequent crack in the plaster. Basically for screws or bigger nails, it’s a good idea to drill a small hole first to lessen the pressure, or do a bigger hole and use a spring bolt anchor.
If you built the bookshelf in wood, it will be expanding, contracting and shifting over time with temperature and humidity variation throughout the day and season. And asymmetrically depending on the grain.
The straight right angles won't stay that way, and it's better to design such that they change in complementary ways, rather than remain perfect.
Then I studied abroad in Italy and realized the phenomenon was global.
Then I moved to San Francisco and realized it was a joke at this point.
Building codes are a joke!
Brunelleschi drank wine all day. Sure he basically invented pulleys and designed the sickest dome in human history, but he was a drunk and I can't blame him because architecture and all of math is a joke if you look close enough.
My landlord does not live in my building T_T
Wait, it's not even a person.
According to Zillow my apartment is owned by a cat with 47,000 followers
I can never look at staircases the same.
Soon you realize that an surprising amount of walls are just not straight or level.
- you genuinely learn once your assumptions about how a system works break, you realize it, try differently, validate, get a better model of that system
- your interfaces must remain permissive while providing feedback, namely you provide wiggle room then only once it behaves roughly as expected do you tighten then up
I think I'm drawn to programming because the fiddliness is tractable, and fixable.
In which other domain can I:
* introspect the relevant processes/state, step by step
* snapshot/undo
* fix niggles, once and for all, and for everyone; and get their fixes too
* probe and test my inputs and outputs, checking for quality. Get notified if a part changes in a way that breaks me.
And the only tool I need is a commodity general purpose PC.
When I try woodwork, or even electronics, I'm struck by much friction is in even simple tasks: tools, parts, lead time, safety, space, physical effort, cost, ...
The best I could do with woodworking in the end to approximate programming was live with wasting some timber, leave a lot of margin on the main cuts and size all the pieces as a whole.
Woodworking (the more constructive, furniture-making kind), rewards a deliberate, controlled process and it savagely penalizes mistakes. Those lessons transfer well to other disciplines. I’d have been a much better student if I’d learned wood working in high school.
Woodworking was part of my first 3 years of high school, but it was mainly about learning safety and tool usage and not planning, estimating, selecting or purchasing timber.
These days I only want to go to the lumberyard once for a project. Learnt the hard way on my first project that you need to take the time to carefully select the timber - checking straightness, matching grain and also colour before I started. Major hassle and waste of time to have to go back to swap boards.
Sure enough, several had cracks at the ends, knots in poor places, and other things that, had I bought it, would have caused me trouble.
I can be a naive person in that I assume good faith. I would never knowingly sell something poor quality to someone else. I had assumed because I was being sold it, it was okay.
And also that’s why AI tools create mix reactions. A couple of months ago a post went viral which was really insightful on what I was originally drawn to cs.
https://news.ycombinator.com/item?id=46881264
With wood you are up against nature. With software you are up against corporations and comities.
You're up against your wood vendor. Anyone familiar with Home Depot "fresh from the tree" lumber has discovered this.
https://news.ycombinator.com/item?id=16184255 - Jan 19, 2018
https://news.ycombinator.com/item?id=22020495 - Jan 11, 2020
https://news.ycombinator.com/item?id=29429385 - Dec 3, 2021
https://news.ycombinator.com/item?id=38407851 - Nov 24, 2023
https://news.ycombinator.com/item?id=43087779 - Feb 21, 2025
https://contraptions.venkateshrao.com/p/truth-in-inconvenien...
I thought of this because this morning I was putting a small fence around some plants we want to protect from deer. The fence consisted of 20 sections (bought on Amazon), each about 24 inches wide. Our ground is like rock, and the fence was not that sturdy, so I had to pound a heavy spike into the ground to the depth of the fence posts, then pull the spike out and put each section's legs in, leaving room for the next section's leg to go into the same spike hole. I wanted to be sure I was putting each section in at the right position, lest I end up with a 12 inch gap and have to go back and adjust lots of sections. Long story short, I pretty much succeeded, although when it cools down I may adjust a few sections. But the problem was sort of like the stairs: I wanted an integer number of fence sections, each the same length, to exactly fit around the bushes---just like you want an integer number of vertical steps in a diagonal stair, each of the same (more or less standard) height.
Then we give it to someone else and it fails on their first or second attempt. They simply tried to use it in a way that we did not anticipate. It doesn't mean that we are dumb for not thinking of those possibilities; it just means that we did not think of every one of them.
And it pretty much is most times. For us. But indeed I did run into the "why would they use it like that though" scenario, where it fails. So I have to patch that usecase. At which point I go all "ok now it's really done" until another fringe usecase pops up and so on and so forth. And I did think about how sure I was it was bulletproof, humbling moment.
AI doesn't deal with reality, it deals with tokens. This is why all those vibe-coded harnesses, little more than glue between various text IO interfaces, are several hundreds of thousands of source lines of code.
It's why a SOTA model took 100kSLoK to write a C compiler to compile one specific project.
It's why, when I asked for a simple markdown -> ansi escape codes converter (for terminal output) in Python, SOTA Claude and SOTA ChatGPT both give me +- 150 SLoC when my own LUT-based version came to under 10 lines of code + a LUT.
Reality has a surprising amount of detail, but LLMs don't exist in reality, they exist in a virtual world made up off tokens.
All information gets compressed. Even your own perception of reality gets compressed.
Yes.
> Or just in a virtual world made up of sensory signals?
No, definitely reality. Things affect my thought whether I sense them or not.
I think the more interesting corollary of this article is that if we're living in a simulation, it's an impossibly, improbably detailed one. I really want some compute time on the HPC that's running it.
Okay, lets go with that :-)
I might be living in a virtual reality, correct, I have no way of knowing.
What I do know is that the reality I am in is many thousands of times higher in resolution than the reality of the LLM.
As an analogy, the LLM is seeing a downscaled 32x32 pixel image while I see the original 8k image. Whether there is a larger 1b^2 image that I cannot see is not relevant to the question of whether the LLM can see my reality or not - it can't.
Reality is by definition our physical reality, which is about an infinite number of levels more detailed than the, you know, _virtual_ digital world computers exist in.
Whatever world we construct for LLMs, no matter how detailed we make it, will always be a blocky projection of the real domain onto a virtual one.
It follows then that any insight gained in the virtual world is at best a rough approximation which can be quite useful at times but also utterly faulty on occasion.
How often it is useful vs. wrong is (partially) a function of how complete the real-to-virtual approximation for a given domain.
Certain domains, given their limited degrees of freedom, can be quite accurately modeled, such as a subway map.
But many domains cannot, and it's important to be aware of that inherent limitation in digital models including but not limited to LLM """reasoning"""
I don't know exactly why but I never really understood this argument. Might be some kind of control thing? Because for me it's pretty simple, it's basically free to give access to reality. Just add "sensory organs" as it were. I can argue you can make them perceive reality even better than we (humans) do, just enlarge the audio/video spectrums. Bam...more reality. The whole point of the argument is we're missing information.
Again, I get the need for controlling the environment for what LLM/AI/AGI/whatever will be, but that will always cost more than giving them access to like...reality. Same reason I don't really believe in the whole simulation argument, it's just more expensive all around, loses resolution, let alone control. I don't doubt there will be some people that would indulge in neverending hedonism but not all people. You need to give up control for that.
I dunno what you mean by "free". The model is trained on text. To "give" the model sensory organs it would need to be trained on those sensory organs.
Current models can predict text, because that's what the weights represent. Models with sensory organs will need to be trained on the output of those sensory organs.
That sounds close to impossible in the foreseeable future.
I expect there's a lot of detail that I'm unaware of relating to running a company (planning; risk; legal; ...) that might make a decision foolish to me, but make sense if given more context.
Contemplating the details of a thing is really satisfying. At times I find myself sitting there and trying to decompose the astonishing amount of work, research, both evolutionary and revolutionary progress that has gone into reaching the current level of something. Buying myself a coffee and stare at the local ferry and acknowledge that someones life's work went into figuring out how to make the paint stick to metal.
Naturally the other point also sticks.. I too often get stuck on the details. :P
> As you learn, notice which details actually change how you think.
Lovely article. The older I get the more I appreciate this.
One point worth making: in many cases, after learning to see & appreciate the details, you gain the power to ignore the details that don't matter to you. This can be quite freeing.
Infinite is a very big claim.
It's just that intuition doesn't scale and a lot of common cases can be handled with models, rules, definitions etc. People continue to be confused that just adding more rules eventually solves reality but it never has anywhere for anything so continuing to believe it will is wild
No matter how many data centers are built, it is impossible to accommodate that level of detail.
But actually in the years since this was written, I do think the world has shifted. Doing things on a computer used to be really hard. Even just installing a framework or getting >python to call the right python on windows. Then install Django and get Django to work with nginx etc. It was just a lot of thankless, frustrating work to get from zero to 1%.
Aside from AI, the tools and packages and culture of computing has gotten better. But AI means you just get all the trivial but difficult stuff for free. And I think a lot of people who would have given up now make it through to see something work and they’ll feel the thrill of building something. It’s just better and easier now.
I visit every couple of days. It's REMARKABLE how fast things get done. One day, there were no walls. The next day, almost all of the walls were in place!
... and yet, at the same time, things take a long amount of time because reality has a surprising amount of detail. I haven't taken into account how much you have to do to frame a house. So incredible amounts of work get done, day after day, but 3/4 of them are things I had no idea needed to get done! Gazing up into the roof, the detail is incredible. The PSL beams, the brackets, the joists, the trusses, just.. EVERYTHING!
I thought the structural engineer's plans had an incredible amount of detail on them, and they do, but they also don't really say anything about _how_ to build the thing. How to put up the walls, how to hold them together temporarily, how to lift beams into place. In what order things can and should be done. That all just takes experience.
> you could be intellectually stuck right at this very moment, with the evidence right in front of your face and you just can’t see it.
It’s always a little disappointing to me when I think I’ve run into something unique but it ends up being user error or something.
Being a founder has a lot of SRE like activities. Fortunately I used to actually like troubleshooting and hence love being a founder but I know a lot of people quit this path because of the "suprising amount of details" in reality!
https://xkcd.com/1741/
When you hire someone to work on the stairs for you, you /hope/ they know what they're doing, especially if you don't have the skills yourself to judge their work. Same for an agent.