This has not been true in my experience. The bar of what is possible has been moved dramatically. So my daily quest is not to continue to achieve our old objectives, but rethink worthy objectives for our new tools.
I think this still leaves much of the concern the article points towards open. If we come across an objective that remains tedious or annoying, does AI create pressure towards thinking that it must not be a good objective and some AI-enabled objective would be better?
- I (https://compile-xc.org) built a compiler for a language I call ‘xc’, similar to Objective-C but designed to be cross-platform and have less square brackets :). It can run on any of {Windows, Mac, Linux, Zynq} and produce code for any of {Windows, Mac, Linux, Zynq, WASM, iOS, Android, m68k, 6502}. It needs no host-platform tools to compile code, so for example I have a signed binary on my iPhone, written and signed entirely on Linux. The compiler currently produces code that is between 2x and 0.5x the speed of clang, which is pretty good for a nascent compiler. It also compiles the same language for execution on a GPU, and makes it trivial to designate parts of your code as data-parallel (https://compile-xc.org/compiler/language/par/). One statement turns a code block into a GPU kernel
- I built (more accurately: am building) a (https://compile-xc.org/compiler/api/uxkit) cross-platform UI framework, so you can write code once, even Table-views, Outline-views, Collection-views, etc., etc and your application will bind to the platform’s native toolkit through a common programming API. The next step here is an “Interface Builder” like application I call RoCkS (for historical reasons), which lets you design the UI for {desktop, tablet, phone} to bind to the same core code, dragging connections between that core code and the UI elements for the different environments
- I built (https://atari-xt.com/os) XTOS for the Zynq - an OS that started with FreeRTOS and added loadable processes, dynamic linking, shared memory, paging, a boot environment that can be scripted like Linux, networking, a graphical UI based on GEM, with a hardware blitter, and HDMI output with Audio islands mixed in. The blitter owns the video memory and uses shared-memory pools to give each “virtual workstation” (GEM-speak for “view of the display”) its own retained-mode and composited graphics memory. The whole OS boots from power-on to desktop in a couple of seconds, which is the reason for making it, rather than just using Linux, I wanted that “instant-on” feeling. It runs on a Z-turn board from MyIR.
- I built an Atari (6502, ANTIC, POKEY, various other chips) (https://atari-xt.com/hardware/x/)emulator, which passes the (https://forums.atariage.com/topic/171296-acid800-an-atari-te...) Acid 800 emulation torture test, and can run games such as Ballblazer, Elektraglide and (https://0x0000ff.co.uk/mov/xt/xtos-aug-11.mov)Despatch Rider. It’s integrated into XTOS, so the OS boots into the GEM-based Desktop.app (written in xc) and you can double-click an icon to launch a game. The plan is to extend that to the ST once I get past the current task, which is …
- I built a Reddit-like website that I plan to launch in a couple of months, written entirely in xc, both the Linux-based server and the client WASM code; the server-side is just xc code, no Apache, no scripting language, just lean-and-mean xc code handling everything from the TLS handshake to the Postgres database back-end. The main difference between (https://blewit.net/)blewit.net and Reddit is that I don’t collect any personal information, don’t track your habits and sell your data, don’t use adverts...
Yes, it’s very similar to that, except the ic has no imagination, vision, or drive.
I wrote my first compiler about 35 years ago, and my first OS around the same time, well before LLMs were on the scene. I know exactly just how much work is involved with doing it well, which is why I’d never normally take on a project like this now I’m retired.
But we do not live in normal times. The language “I” have created (and it’s not as though I just said to Claude “hey, write me a compiler”) is (even if I say so myself) pretty cool. Here’s the full source of a Mandelbrot program:
#import "Stdio.xc"
#import "Par.xc"
#import "../src/include/bench_time.xc"
#define W 2048
#define SIZE (W * W)
i32 main(void)
{
// Eight runs: the first carries one-off costs (a GPU builds its kernel
// then), so the figure is the best run; the first is printed too. Auto
// decides after its fourth (two on each device), and uses the rest.
u32 total = 0;
i64 best = 0;
i64 first = 0;
for (u32 rep in 0..8)
{
total = 0;
i64 t0 = bench_now_us();
par mandel :reduce(+ total)
{
for (u32 i in 0..SIZE)
{
float cx = (float)(i % W) * (3.0f / 2048.0f) - 2.0f;
float cy = (float)(i / W) * (3.0f / 2048.0f) - 1.5f;
float x = 0.0f;
float y = 0.0f;
u32 k = 0;
while (k < 256 && x * x + y * y < 4.0f)
{
float xt = x * x - y * y + cx;
y = 2.0f * x * y + cy;
x = xt;
k = k + 1;
}
total = total + k;
}
}
i64 t1 = bench_now_us();
if (rep == 0)
first = t1 - t0;
if (rep == 0 || t1 - t0 < best)
best = t1 - t0;
}
Stdio.printf("%u %lld %lld\n", (total + 500000) / 1000000, best, first);
return 0;
}
Just one line in that source code is what turns it from a bog-standard C-like program into a GPU kernel, automatically tracked, scheduled, and all data hazards transparently managed.
On my M4 max, that one line of code takes it from 481ms to 1.6ms. All ‘par’ does is say “this next block of code, call it “mandel”, is data-parallel, and for the ‘total’ variable all execution units should combine their output to that variable using +
No huge GPU boilerplate, no stream management for data transport from cpu to gpu and back, no memory barriers, no hassle. Just “hey, do this on the gpu”.
GPU blocks are bound at runtime using the best design for the hardware on the machine, obtained by measuring it if it hasn’t been measured before. If it has, the cache is read. The user can always override, which is why ‘par’ blocks can be named.
Next up is an overlapping scheduler, which keeps 3 work units in play {pending, being-processed, being-transferred-once-done} for a processing device (right now, just a local gpu), which is also the multi-gpu story, and even the networked-machines story.
Schrödinger's AI: can't come up with ideas on its own and we'll all be doomed once we come to rely on it, but we shouldn't credit you with what you use it for; it's all just AI slop.
I think this is great work; these are strange times, i personally don't feel like i'm building stuff anymore, it is more like thinking it into existence
Maybe more like an architect saying "I built" when actually construction workers did all the building. Even that's stretching it though because construction workers tend to follow the plans fairly well. AI has you constantly tweaking the beams and cement yourself mid-construction to get a usable product.
:D added reddit clone to personal watch list. You are braver than me sharing this stuff on HN ( they can be ruthless when it comes to this stuff ). Nice work.
I’m satisfied that my actual engineering is up to snuff, done some pretty cool things over my career. Now I’m just having fun, and LLMs make it easier to have said fun :)
Now that is funny, because I will die on this hill: I wrote every last word of that message - and to my eyes it doesn’t even read like an LLM wrote it.
People are way to quick to try to bring others down these days, unless there’s some subtle humour here I’m not getting.
If anything, I would claim the opposite ( and agree with you )
Not only has AI allowed me to start projects without knowing all the steps ahead of time, it has helped me proceed through each of those steps to completion.
"For the first experiment with 354 participants, one group was tasked with solving 15 basic fraction problems without help, while another had ChatGPT open on the side to use for assistance and to even ask for the answer. The group with AI access started off more accurate than the group without it. Then, after 12 problems, researchers took away the AI tool. Almost immediately, the people in that group stopped solving questions accurately."
Looking at the design of the experiment, it seems everyones brain was primed to think "I have AI at my disposal" ... and then had the framework shattered when it was taken away, which could have affected the test.
Perhaps this study exposes "conversational mindset" versus "mathematical mindset" switching times ?
Theres only one thing I'm certain of ... Claude loves me. :D
The article is not about the output but on the process. Sure you can produce more and more quickly with AI, but at the price of losing the ability to do the same without it. And that's been true in my experience
Or is it more like creating machine code but using a tool called a compiler? If you've ever made anything with AI you know it by default generates things that aren't ideal. You still have to go in and debug, make the UI better, think of more elegant ways to present things. The fact that you're focused on those issues instead of tediously finding the block of code and maybe tediously looking up ways to fix / make it better is probably an overall good thing in the long run. We've lost the ability to code in assembly I'm not sure that's made software or life worse even if coders as a group are in all likelihood less intelligent these days.
In your analogy maybe it's more like creating a pizza except you didn't have to make the cheese, grind up the flour, go find the water in a stream somewhere, grow the tomatoes yourself, butcher the animal for the meat, etc.
You can throw away the compiler and the developer who created the software will still understand how it works. You can't say the same about someone who used an LLM who would need an LLM to continue working on the project and understand how it works. If the tool understands the code better than you and can't continue without it, you have essentially given up ownership.
The only caveat with my statement is that there are different degrees in which you can involve an LLM in your project but I mean if you are using it to write the majority of your code.
A more apt comparison would be the producer role (not the executive type, I mean the Project Manager type of roles). And to be honest, I've seen some amazing music A&R being the reason an album is a masterpiece without ever pushing a button or playing an instrument. But none of them would call themselves musicians.
>Honestly I don’t expect these kind of low quality comment in HN. Where’s your thinking?
A comment that opens a discussion is not low quality just because you disagree with it Mr 7 karma. There is always a degree of irony with such statements.
You choosing the toppings for your pizza and whether you want cheese in your crust does not mean you made the pizza. You did not make it, you merely ordered it.
Wrong metaphor, because making a pizza involves other people.
A better metaphor would be:
A photocopier won’t copy documents by itself. If someone uses a machine to make a copy, they still made that exact copy. Whether that copy constitutes copyright infringement is a separate question.
Authorship still belongs to the user for AI outputs; their originality is debatable, though. These are two distinct concepts that people easily confuse.
Again, apply this reasoning to illegal AI-generated content: if humans are responsible for AI outputs, then they also have authorship of those outputs.
You’re denying another human’s authorship simply because they use a different tool from you.
Another angle you might argue is that you aren’t really “making” something unless you craft it directly. But then what about film directors, architects, or CEOs? Did they really make anything?
Again, you’re denying the role of human intention by gatekeeping the idea of “making” behind direct craft. That’s fine, but it’s a limited worldview.
At which level of instruction abstraction would you deem it to be someone's creation?
You could argue anything high level is not someone's creation (e.g, the use of other people's libraries, compiler, machine language etc.)
I consider it at the point I didn't build it with my hands. If I'm placing the girders, I built the building. If I'm instructing someone to place the girders, I didn't. In that case, the best I can get is "we" built the building.
You are not just going up an abstraction. Whether you are using Python, C, Assembly, you are still engineering because you are writing the specific instructions that make the software work. In this case, you are more like a manager than an engineer that is managing something to engineer for you.
Calling scipy.spatial.Delaunay feels about the same level of abstraction as telling an LLM "write me a routine to calculate Delauny triangulation", except the LLM can do it in any language, and I can trivially tweak the algorithm if I need to. If anything, using LLMs to generate code feels like going down a level, because I can now be bothered implementing things from scratch where previously I would simply have used a library.
You can't convince me that shepherding an LLM to re-implement a library function involves less "engineering" than just... calling the function.
Exactly! The sheer effort of having AI produce good results, from writing, to planning, to coding is immense. I have consistently gather all my energies and focus to guide agents to the ideal, maintainable, and usable solution. Often it is me that makes the crucial design contribution.
Yes, I could be lazy and just vibe piles of slop. But anyone who spends any meaningful amount of time using AI (more than the articles 10 minutes), knows that quite extreme concentration and effort is required to build ambitiously with agents.
I feel like I work harder and and with more difficult problems than before AI, just at a much faster pace and scope.
Yes, thank you! This is exactly my experience. It requires immense effort and concentration, maybe more than before AI, where
- the next ideal step progressively manifested itself to me during coding
- I actually had perfect knowledge of the codebase and didn't have to work as hard to reconstruct it in my head each time
- coding itself is half a mechanical job. AI does the mechanical job itself well, leaving you only the hard thinking.
- AI prose isn't easy to parse, at least to ne.
- Prompting AI until it gets the optimal solution / implementation takes time and revisions.
Something like that.
I keep thinking that it's an issue of my approach, though. I've been trying to improve my AI workflow as much as possible for months, every day, but still, I'm not content with it. I might have gotten stuck on a local optimum. Maybe I'm writing prompts needlessly detailed. Maybe I should read more of the code, because I don't, really, making me not know the actual codebase and how things work.
It's not easy, this whole thing. And honestly, sometimes I wonder - after accounting for the total time, including revisions, mental reconstructions etc - if it wouldn't have been done faster by me in the pre-AI world.
When faced with AI some companies are looking for employees to fire to achieve same goals cheaper. Other companies are looking how to uplift their employees to expand their capability and excelence.
Maybe the same thing happens at the level of individual persons. I think the difference is whether you felt limited or competent. If you felt competent you shrink. If you felt limited, you grow.
The blasé attitude with which people talk about “moving up the stack” as if it’s a foregone conclusion that it’s the right thing to do or even an accurate description of what’s happening is easily one of the most disheartening aspects of this whole movement.
> The blasé attitude with which people talk about “moving up the stack” as if it’s a foregone conclusion that it’s the right thing to do or even an accurate description of what’s happening is easily one of the most disheartening aspects of this whole movement
They need to walk the walk. Or accept that they themselves haven't stopped to think before they "moved up the stack".
I honestly don't care anymore. No matter how hard I try nothing ever gets better. If AI eases the burden even a little I'll take it. Persistenting at hard things hasn't made my life any better yet.
I don't understand this. What ever was the purpose of your life in the first place? What are you measuring to say "hasn't made my life better"?
Everything interesting enough to be worth doing is hard, and persisting at it is my most rewarding way to spend time.
If you took it away and made me loaf on a beach, I would enjoy it for about a day.
And then my life is empty and worthless and definitely not better. Even on that beach I would shortly be trying to make something interesting out of the palm leaves.
And then someon would say we should have a machine to do that because it's hard and "persisting at hard things hasn't made my life any better".
It's the most depressing sentiment I can imagine to say that.
And the IQs of those born after 1975 via the reverse Flynn Affect, those who had these devices of consumption and easy answers such as TV, and later Social Media, AI, etc has been decreasing.
I can only imagine a further and speedier decline will happen, and is happening.
Is it the domestication of humankind? The average person is no longer able to hunt for themselves. Like ant colonies, people are no longer capable of living alone.
Where would they do it? There is not remotely enough available land near where most people live that isn't privately controlled.
What would they eat? Getting enough calories from hunting for it to become a frequent part of everyone's life is currently a thermodynamic impossibility. On top of that, food webs are actively collapsing or have already.
Put a person from 1870s London in current-day London. At first they will be amazed by the technological advancements. Then they will be amazed by how dumb the average person has become. Idiocracy is a documentary, not a comedy.
Wouldn't it have been more useful/conclusive if the study had another control group or two with non-AI based assistance (access to math books, access to calculator, access to human experts, etc)?
This is exactly the reason the studios garbage, and was clearly tailored to prove a point and sell books. He literally should have asked got5 or Claude how to design a useful study of the problem.
FWIW, well designed study probably would still support his premise. He at least needed a group with no help, a group with no help but books and a calculator (tools), a group with human assistance, and a group with AI assistance, and a group that does an unrelated activity during the priming period.
I find that this is noticeable in situations where you're working on something a bit more complex and are initially using an LLM as code assistant to make progress. Once it hits a wall and you find yourself needing to just figure it out yourself, man does the debt get paid with interest. It's not a permanent cognitive impairment though. But there is a bit of a recovery period needed. On the whole I think there are some productivity gains, in spite of setbacks (assuming you know how to solve the problems yourself, to begin with)
It reduces tolerance for dealing with bullshit.
Who wants to start a stackoverflow humiliation ritual and breadcrumb following
random websites to find reliable stuff(90% of which is AI generated),
when prompting gets you directly to the answer?
Before that, the "Google rots your brain" articles prophesied the fall of humanity to
search engines: now its mark of intellect to able to manually
curate and search within websites for relevant info.
https://en.wikipedia.org/wiki/Is_Google_Making_Us_Stupid%3F
Using a table saw for just 10 minutes eroded my ability to persist at hand sawing lumber.
The time difference was measured in hours, not minutes, and and effort required to hand-saw is enormous. Also, the accuracy difference is ridiculous, even with a lot of practice, and maintaining that accuracy continues to be a struggle with hand-sawing.
I now own 6 different electric saws for different specific purposes. Table saw, chop saw, circular saw, track saw, jig saw, scroll saw... I actually probably have more. I got rid of my chainsaw because it lost its chain in a way that really scared me one day. I might get another some day, but not that one again. Does a Dremel count? I'm going to say 'no' so I don't have to count 2 of those.
And yet, I still own a 2 Japanese pull saws and a normal American hand saw. I don't use them as much, but some times they're still worth the effort.
When using table saw, you still need to make a lot of decision: The depth of cut, the angle, the length. If you don't make them, your project will not progress.
With AI, you can stop decide, tell the chat bot to do whatever they want and make any decision they like. Your ability to analyze problem and make decision is not practiced anymore and begin corroding.
I'm very capable of getting sloppy cuts to fit into a project, whether I made them by hand or with power tools, or whether I paid someone else to make those cuts for me using whatever tools they use.
I mean: In cabinetry, for instance, that's the kind of slop that dado joints help to hide and nobody ever knows about it after it's glued up.
And I'm not Norm Abram over here: I can can cheat even harder than that. I can get a sloppy six-sided wooden box to look nice and square with neat joints after basic assembly is done, using just a few minutes of time and a router with an appropriate bit.
It won't pass muster in even the most simple of metrological examinations, but I don't have to care about that. (And my work doesn't detract from that of Norm, either.)
First, you know it's sloppy and can be improved and you know what to improve. An AI source code can look very high quality, until you finally realize it's not after diving deeper.
Second, you make decision to be sloppy. That means you can also make decision to do it right. With AI, no matter how much "make no mistake" you scream toward it, no guarantee whatever it produces is high quality.
I agree that it's not completely a fair comparison. But here's why I think that:
With AI, you can offload almost everything. With the saw, all I'm offloading is the cutting.
I don't actually have to make any decisions when doing woodworking. I can just buy some plans and follow them. The only time I need to make decisions is when I make something unique.
TBF, that's almost all the time for me.
But also TBF, it's the same with AI.
Sure, I can let it make all the decisions, but then I'm not going to end up with something uniquely mine. It's going to keep pulling back towards its training. To get what I want, I need to keep course-correcting the AI while building.
Some people are going to be fine with that, and will let is just make whatever it wants, and call it good enough. That's just like buying the plans and following them blindly.
The difference between tablesaws and AI is, like I said above, the range of things it can do. It's like having all the saws I listed above, plus electric screw drivers, drills, laser levels, and the whole rest of the toolkit, all in 1 box. You even get a chatbot that will tell all those things what to do to follow the plans.
Making good plans has always been one of the hardest parts of both woodworking and software development.
When you have half-arsed plans (in either discipline) you need to fix things up as you go, and that hasn't changed with AI.
AI removed a lot of the other hard things, but left the hard planning step, and the hard "hide the screwups" steps.
I think that's why many high-regard people in tech start using LLM(Linus for example).
LLM let you focus on higher level, but LLM, to some degree, can also do higher level. Ask it "Design the architect of this software" and it will be happy to comply.
I've followed people using AI for a while now. At the beginning they are like you, write me a function that does this. They review every line of code. Then later they ask the LLM to write them a file, write them a module, etc. With the amount of code the LLM output, there is no practical way they review all of the code line by line. It's a slippery slope that require strict discipline and many people lack of that.
> The findings backed up what anecdotes and surveys have increasingly found in recent years: An overreliance on AI for answers erodes our cognitive fitness.
What is our cognitive fitness without our tools? How capable would we be at intellectual tasks if you remove computers, search engines and other things we now take for granted?
It's clear to me our overall output is better with those tools, and better with AI. So the comparison the paper is trying to make is disingenuous, because it tries to compare a baseline without any modern tools (ie, solving fractions by hand) to a result where the best tools are suddenly removed. But these tools are here to stay, and my cognitive fitness has adapted to their continued existence.
How well would these people doing fractions by hand perform if math notation could not be used? Or if the most common method for solving fractions on paper was suddenly disallowed?
Fear-mongering clickbait title signals complete nonsense.
Since I started using AI seriously in my (mathematics) research about 6 weeks ago I've advanced more and been more productive than in years. I'm more motivated and persistent than before because I have at my disposal a tool that helps me get past technical obstacles that previously took up a lot of my time.
Writers disagree on what effect word processing will have on the quality of our written language.
Some writers are concerned that computer assistance may promote dry bland writing.
Learned Helplessness. Happens alot all around. I see it a lot in disability care. SO much that there is a saying: "Suplemental disability caregiver" (Zusatzbehinderung Eltern). Might be a schock to the more socially oriented people. But the fact is, the more help you allow to be provided to you, the more you forget/unlearn how to handle things on your own. Which in the end, will make you hyperdependant on others.
This is just our energy conservation system in practice. It happens every time we experience and easier/better way of doing something. Another version of this headline could have been "Experiencing VIP treatment for just 10 minutes erodes your ability to tolerate the normal shit you have to put up with"
I can’t tell if you’re defending this as good or just hypothesizing about the underlying cause but “appeal to nature” is one of the many freshman-level logical fallacies that seem ubiquitous in AI hype circles. Just because it’s what our animal brains do, doesn’t mean it’s going to be good or advantageous to our current evolutionary context.
Our energy conservation system also encourages us to eat as much high fat, high sugar food as we can get down our gullets, but that turns out to be a big problem for modern humans.
This behavior is not very surprising to me, and I think the conclusion is a bit off. Here is an analogous study:
Participants are asked to drive 20 nails into a plank of wood. Group A is given a hammer and an unripe banana. Group B is given only a banana. Halfway through the experiment, the hammers are taken away.
Hammers erode your ability to persist at hard things…
My point being that AI is a new tool that allows us to work on harder things at a higher level. It’s like having a team of 100 skilled workers with nail guns (and all manner of other tools) that never tire, and you get to direct them. Once you experience that, it’s a bit disheartening to go back to building alone with your bare hands… even if you lament the decline in your hand-eye coordination.
for those that have spent 20 years of their career, perfecting the art of coding / writing / research / all forms of knowledge work, it sure is depressing to see the career moats we tried to shore up just diminish practically overnight. The ability to operate at a higher level of abstraction is real, but so is the sudden devaluation of skills that previously required years to acquire.
>Once you experience that, it’s a bit disheartening to go back to building alone with your bare hands.
Art provides a counterpoint. The invention of cameras allowed you to capture reality in ways never seen before, and yet this lead to art having a _really weird_ few decades/century where artists rejected realism and instead embraced abstraction.
The difference between AI and no AI is like the difference between a hammer and a nail gun not a hammer and a banana.
If you have a nail gun you have to pay attention to what you are nailing together because it might rip through or the nail might stick out on the other side before you get a chance to pull it out and pick a shorter nail.
I've sort of come around to the other side, a few months ago I was pumping out as many project in my lab as I could manage thinking I was suddenly "able to do all the things I never got around to", now, idk. Give me a terminal, Ubuntu 26, and the simplest text editor in the world. I'm not making money off producing slop.
This all reminds me of keruigs for some reason, like a quick easy way to get coffee that produces a lot of waste for a forgettable drink. I'm sort of advocating digital pour overs.
I remember when people would say that if a person took LSD 10 times or more that they would go insane. After about my 30th or 40th trip I knew it was nonsense.
This feels similar. 10 minutes? So, after sever thousand hours of using AI to build projects, I can't solve hard problems? More nonsense.
I have to go now, the nurse here at the psychiatric ward is about to strap me down and give me my meds.
The evidence for reduced persistence is kinda weak.
The strongest experimental result in the paper is: if people use AI to help solve a series of problems, and then the AI is suddenly taken away, they perform worse on subsequent problems than people who hadn't been using AI for the earlier questions.
But:
- They're looking at really really short-term effects, like seconds or minutes. That's not really interesting if we care about long term effects (which is what the abstract says is important).
- What they found is totally expected and isn't something I'd worry about: it's kind of expected that if one guy warms up with 10 questions done himself, and the other guy warms up with 10 questions delegated to AI, then the second guy is going to do worse on the final 10 questions (because he didn't warm up).
More like I have done the work, we all have. Then it was stolen and sold back to us, with the openly stated goal our devaluing our labor. I can understand a bit of entitlement here.
Nothing to excuse because thats not what i am asking humans to do, i will never in my life, and i mean it i would rather a clanker kills me, accept a tool telling me what if "feels (what a joke) is correct.
i suppose you are a human so no issue there, if you are a clanker, get out of society talk
> - What they found is totally expected and isn't something I'd worry about: it's kind of expected that if one guy warms up with 10 questions done himself, and the other guy warms up with 10 questions delegated to AI, then the second guy is going to do worse on the final 10 questions (because he didn't warm up).
I agree with your general conclusion, but with this point, the solve rate for the AI group once AI is removed is lower than the initial solve rate for the other group in all the experiments (or at least definitely in 2 of 3), so I don't think it's explained by a warm-up.
Sort of like if you're sawing boards with a handsaw, and then they give you an electric saw but that fails coincidentally, and you don't expend effort because why not wait for the electric saw to be fixed?
126 comments
[ 0.30 ms ] story [ 51.0 ms ] thread- I (https://compile-xc.org) built a compiler for a language I call ‘xc’, similar to Objective-C but designed to be cross-platform and have less square brackets :). It can run on any of {Windows, Mac, Linux, Zynq} and produce code for any of {Windows, Mac, Linux, Zynq, WASM, iOS, Android, m68k, 6502}. It needs no host-platform tools to compile code, so for example I have a signed binary on my iPhone, written and signed entirely on Linux. The compiler currently produces code that is between 2x and 0.5x the speed of clang, which is pretty good for a nascent compiler. It also compiles the same language for execution on a GPU, and makes it trivial to designate parts of your code as data-parallel (https://compile-xc.org/compiler/language/par/). One statement turns a code block into a GPU kernel
- I built (more accurately: am building) a (https://compile-xc.org/compiler/api/uxkit) cross-platform UI framework, so you can write code once, even Table-views, Outline-views, Collection-views, etc., etc and your application will bind to the platform’s native toolkit through a common programming API. The next step here is an “Interface Builder” like application I call RoCkS (for historical reasons), which lets you design the UI for {desktop, tablet, phone} to bind to the same core code, dragging connections between that core code and the UI elements for the different environments
- I built (https://atari-xt.com/os) XTOS for the Zynq - an OS that started with FreeRTOS and added loadable processes, dynamic linking, shared memory, paging, a boot environment that can be scripted like Linux, networking, a graphical UI based on GEM, with a hardware blitter, and HDMI output with Audio islands mixed in. The blitter owns the video memory and uses shared-memory pools to give each “virtual workstation” (GEM-speak for “view of the display”) its own retained-mode and composited graphics memory. The whole OS boots from power-on to desktop in a couple of seconds, which is the reason for making it, rather than just using Linux, I wanted that “instant-on” feeling. It runs on a Z-turn board from MyIR.
- I built an Atari (6502, ANTIC, POKEY, various other chips) (https://atari-xt.com/hardware/x/)emulator, which passes the (https://forums.atariage.com/topic/171296-acid800-an-atari-te...) Acid 800 emulation torture test, and can run games such as Ballblazer, Elektraglide and (https://0x0000ff.co.uk/mov/xt/xtos-aug-11.mov)Despatch Rider. It’s integrated into XTOS, so the OS boots into the GEM-based Desktop.app (written in xc) and you can double-click an icon to launch a game. The plan is to extend that to the ST once I get past the current task, which is …
- I built a Reddit-like website that I plan to launch in a couple of months, written entirely in xc, both the Linux-based server and the client WASM code; the server-side is just xc code, no Apache, no scripting language, just lean-and-mean xc code handling everything from the TLS handshake to the Postgres database back-end. The main difference between (https://blewit.net/)blewit.net and Reddit is that I don’t collect any personal information, don’t track your habits and sell your data, don’t use adverts...
E.g.: I painted the Mona Lisa.
“I” meaning “I with the help of a JPEG on a computer and a color printer”.
I wrote my first compiler about 35 years ago, and my first OS around the same time, well before LLMs were on the scene. I know exactly just how much work is involved with doing it well, which is why I’d never normally take on a project like this now I’m retired.
But we do not live in normal times. The language “I” have created (and it’s not as though I just said to Claude “hey, write me a compiler”) is (even if I say so myself) pretty cool. Here’s the full source of a Mandelbrot program:
Just one line in that source code is what turns it from a bog-standard C-like program into a GPU kernel, automatically tracked, scheduled, and all data hazards transparently managed.On my M4 max, that one line of code takes it from 481ms to 1.6ms. All ‘par’ does is say “this next block of code, call it “mandel”, is data-parallel, and for the ‘total’ variable all execution units should combine their output to that variable using +
No huge GPU boilerplate, no stream management for data transport from cpu to gpu and back, no memory barriers, no hassle. Just “hey, do this on the gpu”.
GPU blocks are bound at runtime using the best design for the hardware on the machine, obtained by measuring it if it hasn’t been measured before. If it has, the cache is read. The user can always override, which is why ‘par’ blocks can be named.
Next up is an overlapping scheduler, which keeps 3 work units in play {pending, being-processed, being-transferred-once-done} for a processing device (right now, just a local gpu), which is also the multi-gpu story, and even the networked-machines story.
I’m satisfied that my actual engineering is up to snuff, done some pretty cool things over my career. Now I’m just having fun, and LLMs make it easier to have said fun :)
People are way to quick to try to bring others down these days, unless there’s some subtle humour here I’m not getting.
Not only has AI allowed me to start projects without knowing all the steps ahead of time, it has helped me proceed through each of those steps to completion.
"For the first experiment with 354 participants, one group was tasked with solving 15 basic fraction problems without help, while another had ChatGPT open on the side to use for assistance and to even ask for the answer. The group with AI access started off more accurate than the group without it. Then, after 12 problems, researchers took away the AI tool. Almost immediately, the people in that group stopped solving questions accurately."
Looking at the design of the experiment, it seems everyones brain was primed to think "I have AI at my disposal" ... and then had the framework shattered when it was taken away, which could have affected the test.
Perhaps this study exposes "conversational mindset" versus "mathematical mindset" switching times ?
Theres only one thing I'm certain of ... Claude loves me. :D
In your analogy maybe it's more like creating a pizza except you didn't have to make the cheese, grind up the flour, go find the water in a stream somewhere, grow the tomatoes yourself, butcher the animal for the meat, etc.
The only caveat with my statement is that there are different degrees in which you can involve an LLM in your project but I mean if you are using it to write the majority of your code.
It's the same way a technical director is not exactly a software engineer.
If someone used AI and created something illegal(deepfake or whatever), can they claim it’s not their creation?
A comment that opens a discussion is not low quality just because you disagree with it Mr 7 karma. There is always a degree of irony with such statements.
You choosing the toppings for your pizza and whether you want cheese in your crust does not mean you made the pizza. You did not make it, you merely ordered it.
A better metaphor would be:
A photocopier won’t copy documents by itself. If someone uses a machine to make a copy, they still made that exact copy. Whether that copy constitutes copyright infringement is a separate question.
Authorship still belongs to the user for AI outputs; their originality is debatable, though. These are two distinct concepts that people easily confuse.
Again, apply this reasoning to illegal AI-generated content: if humans are responsible for AI outputs, then they also have authorship of those outputs.
You’re denying another human’s authorship simply because they use a different tool from you.
Another angle you might argue is that you aren’t really “making” something unless you craft it directly. But then what about film directors, architects, or CEOs? Did they really make anything?
Again, you’re denying the role of human intention by gatekeeping the idea of “making” behind direct craft. That’s fine, but it’s a limited worldview.
You can't convince me that shepherding an LLM to re-implement a library function involves less "engineering" than just... calling the function.
Yes, I could be lazy and just vibe piles of slop. But anyone who spends any meaningful amount of time using AI (more than the articles 10 minutes), knows that quite extreme concentration and effort is required to build ambitiously with agents.
I feel like I work harder and and with more difficult problems than before AI, just at a much faster pace and scope.
- the next ideal step progressively manifested itself to me during coding
- I actually had perfect knowledge of the codebase and didn't have to work as hard to reconstruct it in my head each time
- coding itself is half a mechanical job. AI does the mechanical job itself well, leaving you only the hard thinking.
- AI prose isn't easy to parse, at least to ne.
- Prompting AI until it gets the optimal solution / implementation takes time and revisions.
Something like that.
I keep thinking that it's an issue of my approach, though. I've been trying to improve my AI workflow as much as possible for months, every day, but still, I'm not content with it. I might have gotten stuck on a local optimum. Maybe I'm writing prompts needlessly detailed. Maybe I should read more of the code, because I don't, really, making me not know the actual codebase and how things work.
It's not easy, this whole thing. And honestly, sometimes I wonder - after accounting for the total time, including revisions, mental reconstructions etc - if it wouldn't have been done faster by me in the pre-AI world.
Maybe the same thing happens at the level of individual persons. I think the difference is whether you felt limited or competent. If you felt competent you shrink. If you felt limited, you grow.
> The blasé attitude with which people talk about “moving up the stack” as if it’s a foregone conclusion that it’s the right thing to do or even an accurate description of what’s happening is easily one of the most disheartening aspects of this whole movement
They need to walk the walk. Or accept that they themselves haven't stopped to think before they "moved up the stack".
We aren't all dead yet. (And we ain't just hunter-gatherers no more, neither!)
Or maybe instead: Someone has already done that and that is why we're not presently fucked.
Or, more realistically: The sky is not actually falling, just as it has also not fallen with any previous introduction of automation.
This is like thinking a gun just shoots people; nah it's human use of the gun
How do you know trying hard isn’t stopping things from getting worse?
> If AI eases the burden even a little I'll take it.
It does ease the burden for a little bit but then makes the burden much harder when you don’t have it. That’s the point.
> Persistenting at hard things hasn't made my life any better yet.
Again, how do you know it’s not stopping it from getting worse? Also, are you sure you’re worrying about the right hard things?
Everything interesting enough to be worth doing is hard, and persisting at it is my most rewarding way to spend time.
If you took it away and made me loaf on a beach, I would enjoy it for about a day.
And then my life is empty and worthless and definitely not better. Even on that beach I would shortly be trying to make something interesting out of the palm leaves.
And then someon would say we should have a machine to do that because it's hard and "persisting at hard things hasn't made my life any better".
It's the most depressing sentiment I can imagine to say that.
https://www.stuartmcmillen.com/comic/town-without-television...
And the IQs of those born after 1975 via the reverse Flynn Affect, those who had these devices of consumption and easy answers such as TV, and later Social Media, AI, etc has been decreasing.
I can only imagine a further and speedier decline will happen, and is happening.
What would they eat? Getting enough calories from hunting for it to become a frequent part of everyone's life is currently a thermodynamic impossibility. On top of that, food webs are actively collapsing or have already.
The intelligence doesn't shine out of all the faces.
FWIW, well designed study probably would still support his premise. He at least needed a group with no help, a group with no help but books and a calculator (tools), a group with human assistance, and a group with AI assistance, and a group that does an unrelated activity during the priming period.
Or when your weekly limit hit. More fun if it's Tuesday. Say goodbye to the rest of week.
The time difference was measured in hours, not minutes, and and effort required to hand-saw is enormous. Also, the accuracy difference is ridiculous, even with a lot of practice, and maintaining that accuracy continues to be a struggle with hand-sawing.
I now own 6 different electric saws for different specific purposes. Table saw, chop saw, circular saw, track saw, jig saw, scroll saw... I actually probably have more. I got rid of my chainsaw because it lost its chain in a way that really scared me one day. I might get another some day, but not that one again. Does a Dremel count? I'm going to say 'no' so I don't have to count 2 of those.
And yet, I still own a 2 Japanese pull saws and a normal American hand saw. I don't use them as much, but some times they're still worth the effort.
When using table saw, you still need to make a lot of decision: The depth of cut, the angle, the length. If you don't make them, your project will not progress.
With AI, you can stop decide, tell the chat bot to do whatever they want and make any decision they like. Your ability to analyze problem and make decision is not practiced anymore and begin corroding.
I mean: In cabinetry, for instance, that's the kind of slop that dado joints help to hide and nobody ever knows about it after it's glued up.
And I'm not Norm Abram over here: I can can cheat even harder than that. I can get a sloppy six-sided wooden box to look nice and square with neat joints after basic assembly is done, using just a few minutes of time and a router with an appropriate bit.
It won't pass muster in even the most simple of metrological examinations, but I don't have to care about that. (And my work doesn't detract from that of Norm, either.)
Second, you make decision to be sloppy. That means you can also make decision to do it right. With AI, no matter how much "make no mistake" you scream toward it, no guarantee whatever it produces is high quality.
With AI, you can offload almost everything. With the saw, all I'm offloading is the cutting.
I don't actually have to make any decisions when doing woodworking. I can just buy some plans and follow them. The only time I need to make decisions is when I make something unique.
TBF, that's almost all the time for me.
But also TBF, it's the same with AI.
Sure, I can let it make all the decisions, but then I'm not going to end up with something uniquely mine. It's going to keep pulling back towards its training. To get what I want, I need to keep course-correcting the AI while building.
Some people are going to be fine with that, and will let is just make whatever it wants, and call it good enough. That's just like buying the plans and following them blindly.
The difference between tablesaws and AI is, like I said above, the range of things it can do. It's like having all the saws I listed above, plus electric screw drivers, drills, laser levels, and the whole rest of the toolkit, all in 1 box. You even get a chatbot that will tell all those things what to do to follow the plans.
Making good plans has always been one of the hardest parts of both woodworking and software development.
When you have half-arsed plans (in either discipline) you need to fix things up as you go, and that hasn't changed with AI.
AI removed a lot of the other hard things, but left the hard planning step, and the hard "hide the screwups" steps.
LLM let you focus on higher level, but LLM, to some degree, can also do higher level. Ask it "Design the architect of this software" and it will be happy to comply.
I've followed people using AI for a while now. At the beginning they are like you, write me a function that does this. They review every line of code. Then later they ask the LLM to write them a file, write them a module, etc. With the amount of code the LLM output, there is no practical way they review all of the code line by line. It's a slippery slope that require strict discipline and many people lack of that.
What is our cognitive fitness without our tools? How capable would we be at intellectual tasks if you remove computers, search engines and other things we now take for granted?
It's clear to me our overall output is better with those tools, and better with AI. So the comparison the paper is trying to make is disingenuous, because it tries to compare a baseline without any modern tools (ie, solving fractions by hand) to a result where the best tools are suddenly removed. But these tools are here to stay, and my cognitive fitness has adapted to their continued existence.
How well would these people doing fractions by hand perform if math notation could not be used? Or if the most common method for solving fractions on paper was suddenly disallowed?
Since I started using AI seriously in my (mathematics) research about 6 weeks ago I've advanced more and been more productive than in years. I'm more motivated and persistent than before because I have at my disposal a tool that helps me get past technical obstacles that previously took up a lot of my time.
https://www.youtube.com/watch?v=Jt0OoXluC8g @ 4:08:
That's from 1983Eg. I’m glad AI built my startups landing page in (almost) one prompt, because it’s allowed me to do the hard things that matter and I enjoy.
Our energy conservation system also encourages us to eat as much high fat, high sugar food as we can get down our gullets, but that turns out to be a big problem for modern humans.
Participants are asked to drive 20 nails into a plank of wood. Group A is given a hammer and an unripe banana. Group B is given only a banana. Halfway through the experiment, the hammers are taken away.
Hammers erode your ability to persist at hard things…
My point being that AI is a new tool that allows us to work on harder things at a higher level. It’s like having a team of 100 skilled workers with nail guns (and all manner of other tools) that never tire, and you get to direct them. Once you experience that, it’s a bit disheartening to go back to building alone with your bare hands… even if you lament the decline in your hand-eye coordination.
Art provides a counterpoint. The invention of cameras allowed you to capture reality in ways never seen before, and yet this lead to art having a _really weird_ few decades/century where artists rejected realism and instead embraced abstraction.
The difference between AI and no AI is like the difference between a hammer and a nail gun not a hammer and a banana.
If you have a nail gun you have to pay attention to what you are nailing together because it might rip through or the nail might stick out on the other side before you get a chance to pull it out and pick a shorter nail.
This all reminds me of keruigs for some reason, like a quick easy way to get coffee that produces a lot of waste for a forgettable drink. I'm sort of advocating digital pour overs.
This feels similar. 10 minutes? So, after sever thousand hours of using AI to build projects, I can't solve hard problems? More nonsense.
I have to go now, the nurse here at the psychiatric ward is about to strap me down and give me my meds.
The evidence for reduced persistence is kinda weak.
The strongest experimental result in the paper is: if people use AI to help solve a series of problems, and then the AI is suddenly taken away, they perform worse on subsequent problems than people who hadn't been using AI for the earlier questions.
But:
- They're looking at really really short-term effects, like seconds or minutes. That's not really interesting if we care about long term effects (which is what the abstract says is important).
- What they found is totally expected and isn't something I'd worry about: it's kind of expected that if one guy warms up with 10 questions done himself, and the other guy warms up with 10 questions delegated to AI, then the second guy is going to do worse on the final 10 questions (because he didn't warm up).
Eventually you think you deserve to have the answer just handed to you without having done any of the work.
I agree with your general conclusion, but with this point, the solve rate for the AI group once AI is removed is lower than the initial solve rate for the other group in all the experiments (or at least definitely in 2 of 3), so I don't think it's explained by a warm-up.
From page 4:
From page 6: The fact that the AI group did worse on the harder problems than the non-AI group did on the easier ones doesn't tell us anything.