AI coding and the peanut butter and jelly problem(iamcharliegraham.substack.com) |
AI coding and the peanut butter and jelly problem(iamcharliegraham.substack.com) |
If anything, LLMs have surprised at much better they are than humans in understanding instructions for text based activities. But they are MUCH worse than humans when it comes to creating images/videos.
That's demonstrateably false, as proven by both OpenAI's own research [1] and endless independent studies by now.
What is fascinating is how some people cling on false ideas about what LLM is and isnt.
Its a recurring fallacy that's bound to get it's own name any time soon.
Which it isn't, just like pressing keys isn't the difficult part of being a pianist.
If they invented a machine to press piano keys with superhuman speed and precision that wouldn't make you a musician.
Put it this way — I’m going to give you a text based question to solve and you have a choice to get another human to solve it (randomly selected from adults in the US) or ChatGPT, and both will be given 30 minutes to read and solve the problem — which would you choose?
If you do have that skill to communicate clearly and describe the requirements of a novel problem, why is the AI still useful? Actually writing the code should be relatively trivial from there. If it isn't, that points to a problem with your tools/architecture/etc. Programmers IMX are, on average, far too tolerant of boilerplate.
If the LLM gets something wrong, you have to be more exact to get it to make the program do the thing you want. And when that isn’t perfect, you have to tell it exactly what you want to to do in THAT situation. And the next one. And the next one.
At that point you’re programming. It may not be the same as coding in a traditional language, but isn’t it effectively the same process? You’re having to lay out all the exact steps to take when different things happen.
So in the end have you replaced programmers or decreased the amount of programming needed? Or have you just changed the shape of the activity so it doesn’t look like what we’re used to calling programming today?
John Siracusa (one of the hosts) compared it to the idea of a fourth generation language.
From Wikipedia:
“The concept of 4GL was developed from the 1970s through the 1990s, overlapping most of the development of 3GL, with 4GLs identified as ‘non-procedural’ or ‘program-generating’ languages”.
Program generating language sounds an awful lot like what people are trying to use AI for. And these claims that we don’t need programmers anymore also sound a lot like the claims from when people were trying to make flowchart based languages. Or COBOL.
“You don’t need programmers! The managers can write their own reports”.
In fact “the term 4GL was first used formally by James Martin in his 1981 book Application Development Without Programmers” (Wikipedia again).
They keep trying. But it all ends up still being programming.
So to me, even in an ideal world the dream of AI coding is backwards. It's more verbose, it's harder to conceptualize, it's less precise, and it's going to be more of a pain to get right even if it worked perfectly.
That's not to say it'll never work. But the interface has to change a lot. Instead of a UX where you have to think about and specify all the details up front, a useful assistant would be more conversational, analyze the existing codebase, clarify the change you're asking about, propose some options, ask which layer of the system, which design patterns to use, whether the level of coupling makes sense, what extensions of the functionality you're thinking about in the future, pros and cons of each approach, and also help point out conflicts or vague requirements, etc. But it seems like we've got quite a way to go before we get there.
Its abstraction may leak sometimes, but most people using it are incredibly productive without needing to learn what a spool operator or bitmap does.
Even though the GUI and natural language aspects of 4GL failed, declarative programming was worth it.
No, it isn't.
Programming is thinking deeply about
- the invariants that your code obeys
- which huge implications a small, innocent change in one part of the program will have for other, seemingly unrelated parts of the program
- in which sense the current architecture is (still) the best possible for what the program does, and if not, what the best route is to get there
- ...
It also remembers CLI tool args far better than I do. Before LLMs, I would often have to sit and just read a manpage in its entirety to see if a certain command-line tool could do a certain thing. (For example: do you know off-hand if you can get ls(1) to format file mtimes as ISO8601 or POSIX timestamps? Or — do you know how to make find(1) prune a specific subdirectory, so that it doesn't have to iterate-over-and-ignore the millions of tiny files inside it?) But now, I just ask the LLM for the flags that will make the tool do the thing; it spits them out (if they exist); and then I can go and look at the manpage and jump directly to that flag to learn about it — using the manpage as a reference, the way it was intended.
Actually, speaking of CLI tools, it also just knows about tools that I don't. You have to be very good with your google-fu to go from the mental question of "how do I get disk IO queue saturation metrics in Linux?" to learning about e.g. the sar(1) command. Or you can just ask an LLM that actual literal question.
LLMs are both more general and more useful than those tools. They're more flexible and composable, and can replace those tools with a small wrapper script. Part of the reason why the LLMs can do that though is because it has those other tools as datasets to train off of.
Or the time to review the code - whether by manual fixes, or iterating with the prompt, or both?
For you, maybe. This statement assumes years of grueling training to become bilingual in a foreign programming language. And I can't type at 1000 tokens/s personally - sometimes I just want to press the voice dictate key and blab for five seconds and move on to something actually interesting.
...So, less experienced programmers are supposed to be happy that they can save time with the same technology that will convince their employers that a human isn't necessary for the position?
(And, frankly, I've overall quite enjoyed the many years I've put into the craft.)
- documentation (reference, tutorials, overviews) - tools - logging and log analyzers - monitoring - configurability - unit tests - fuzzers - UIs - and not least: lots and lots of prototypes and iterating on ideas
All of these are "trivial" once you have the main code, but they are incredibly valuable, and LLMs do a fantastic job.
"I picked up an extraordinary violin once. It sounded awful!"
There's an art here. Managerial genius is recognizing everyone's strengths and weaknesses, and maximizing impact. Coding with AI is no different.
Of course I have to understand the code well enough to have written it. Usually much of the time is spent proposing improvements.
I'm a few months in, learning to code with AI for my math research. After a career as a professor, I'm not sure I could explain to anyone what I'm starting to get right, but I'm working several times more efficiently than I ever could by hand.
Some people will get the hang of this faster than others, and it's bull to think this can be taught.
If this can't work, the program abstraction is insufficient to the task. This insufficiency is not a surprise.
That an ordinary 5-year can make a sandwich after only ever seeing someone make one, and that the sandwich so-made is a component within a life sustaining matrix which inevitably leads to new 5 year-olds making their own sandwiches and serenading the world about the joys of peanut butter and jelly is the crucial distinction between AI and intelligence.
The rest of the stuff about a Harvard professor ripping a hole in a bag and pouring jelly on a clump of bread on the floor is a kooky semantic game that reveals something about the limits of human intelligence among the academic elite.
We might wonder why some people have to get to university before encountering such basic epistemological conundrum as what constitutes clarity in exposition... But maybe that's what teaching to the test in U.S. K-12 gets you.
Alan Kay is known a riff on a simple study where Harvard students were asked what causes the earth's seasons: almost all of them give the wrong explanation, but many of them are very confident about the correctness of their wrong explanations.
Given that the measure of every AI chat program's performance is how agreeable its response is to a human, is there a clear distinction between a the human and the AI?
If this HN discussion was among AI chat programs considering their own situations and formulating understanding of their own problems; maybe waxing about the ineffable, for them, joy of eating a peanut butter and jelly sandwich...
But it isn't.
https://m.youtube.com/watch?v=RmbFJq2jADY&t=3m25s
Also seems like great advice, feels like a good description of what Ive been gravitating towards / having more luck with lately proompting.
Here we are - we've arrived at the next level.
The emphasis in my prompts is specification : clear and concise, defining terms as they are introduced, and I've had good results with that. I expect that we'll see specification/prompt languages evolve, in the same way that MCP has become a defacto standard API for connecting LLMs to other applications and servers. We could use a lot of the ideas from existing specification languages, and there has been a lot of work done on this over 40+ years, but my impression is they are largely fairly strict, because their motivation was provably-correct code. The ideas can be used in a more relaxed way, because prompting fits well with rapid application development (RAD) and prototyping - I think there is a sweet spot of high productivity in a kind of REPL (read/evaluate/print loop) with symbolic references and structure embedded in free-form text.
Other comments have mentioned the importance of specification and requirements analysis, and dahlfox menions being able to patch new elements into the structure in subsequent prompts (via BASIC line number insertion).
I got the sense the professor liked it a ton too. In the videos of this on YouTube you can see the professor really enjoying watching it all go down.
It still is so memorable.
Why BASIC? It's a lot harder to mix English and structured programming concepts. Plus it's nice if you forget a step between 20 and 30 you can just say `25 print 'halfway'` from the chat.
I specifically mean pseudocode intended to clarify and communicate - as opposed to be run by a computer.
E.g.: https://chatgpt.com/share/67f9fee9-07dc-8003-a272-ca05d91282...
Blast fax kudos all around
The disappointment always stayed with me that my instructions were not chosen, as I really had been far more precise than the fun examples she did choose. I recall even explaining which side of the knife to use when taking peanut butter from the jar.
Of course, she would still have found plenty of bugs in my instructions, which I wish I still had.
Thanks for that, and also the pet rats, Ms. Clouser!
Bzzt, nope!
If the sandwich does not have an obvious recipe --- an app similar to many that have been written before or familiar, conventional UX, or boring features found in countless existing apps --- LLMs struggle.
Fixed it for ya!
If this is how you're modelling the problem, then I don't think you learned the right lesson from the PB&J "parable."
Here's a timeless bit of wisdom, several decades old at this point:
Managers think that if you can just replace code with something else that isn't text with formal syntax, then all the sudden "regular people" (like them, maybe?) will be able to "program" a system. But it never works. And the reason it never works is fundamental to how humans relate to computers.
Hucksters continually reinvent the concept of "business rules engines" to sell to naive CTOs. As a manager, you might think it's a great idea to encode logic/constraints into some kind of database — maybe one you even "program" visually like UML or something! — and to then have some tool run through and interpret those. You can update business rules "live and on the fly", without calling a programmer!
They think it's a great idea... until the first time they try to actually use such a system in anger to encode a real business process. Then they hit the PB&J problem. And, in the end, they must get programmers to interface with the business rules engine for them.
What's going on there? What's missing in the interaction between a manager and a business rules engine, that gets fixed by inserting a programmer?
There are actually two things:
1. Mechanical sympathy. The programmer knows the solution domain — and so the programmer can act as an advocate for the solution domain (in the same way that a compiler does, but much more human-friendly and long-sighted/predictive/10k-ft-view-architectural). The programmer knows enough about the machine and about how programs should be built to know what just won't work — and so will push back on a half-assed design, rather than carrying the manager through on a shared delusion that what they're trying to do is going to work out.
2. Iterative formalization. The programmer knows what information is needed by a versatile union/superset of possible solution architectures in the solution space — not only to design a particular solution, but also to "work backward", comparing/contrasting which solution architectures might be a better fit given the design's parameters. And when the manager hasn't provided this information — the programmer knows to ask questions.
Asking the right questions to get the information needed to determine the right architecture and design a solution — that's called requirements analysis.
And no matter what fancy automatic "do what I mean" system you put in place between a manager and a machine — no matter how "smart" it might be — if it isn't playing the role of a programmer, both in guiding the manager through the requirements analysis process, and in pushing back through knowledge of mechanical sympathy... then you get PB&J.
That being said: LLMs aren't fundamentally incapable of "doing what programmers do", I don't think. The current generation of LLMs is just seemingly
1. highly sycophantic and constitutionally scared of speaking as an authority / pushing back / telling the user they're wrong; and
2. trained to always try to solve the problem as stated, rather than asking questions "until satisfied."
Think about it, a LLM is an autocompleter. It will give you the most probable next word each time. It doesn't mean it doesn't understand high level concepts, but in the end, it just writes stuff that is similar to its training dataset.
For example, ask it to multiply two numbers. If the number are small enough, you will get the right answer. Now ask it to explain how it did it, it will probably tell you the process as commonly taught in school, but it not actually how it did it. What it did is much weirder for us humans, and the only way to see how it actually works is to look at the internals of the neural network. The LLM can't describe it, it doesn't see inside itself, however, it has many textbooks in its training dataset, so it will grab an answer from these textbooks because that's how people answer.
Seeing how it correctly describes the multiplication process and how it can multiply small number correctly, you would assume it can also multiply large numbers (as we do), but nope, it can't, unless it has access to a separate math module, traditionally made (i.e. not a neural net).
EDIT: I see now that you were referring to the answers it uses to justify the result, not the underlying computations. Sorry! You can disregard the actual comment. Leaving for completeness.
ORIGINAL COMMENT:
That's not how it works. Addition in LLMs is believed to function through different mechanisms depending on model size and architecture, but the single consistent finding across different models is that they generalize beyond the training data for at least those simple arithmetic operations.
For example: "Language models use trigonometry to do addition" https://arxiv.org/abs/2502.00873
For a different "emergent" algorithm, see Anthropic's observations: https://transformer-circuits.pub/2025/attribution-graphs/met...
You can in-context-learn an LLM into being a domain expert in a specific domain — at which point it'll start challenging you within that domain.
But — AFAIK — you can't get current LLMs to do the thing that experienced programmers do, where they can "know you're wrong, even though they don't know why yet" — where the response isn't "no, that's wrong, and here's what's right:" but rather "I don't know about that... one minute, let me check something" — followed by motivated googling, consulting docs, etc.
And yes, the "motivated googling" part is something current models (DeepResearch) are capable of. But the layer above that is missing. You need a model with:
1. trained-in reflective awareness — "knowing what you know [and what you don't]" — such that there's a constant signal within the model representing "how confident I am in the knowledge / sources that I'm basing what I'm saying upon", discriminated as a synthesis/reduction over the set of "memories" the model is relying upon;
2. and a trained-in capability to evaluate the seeming authoritativeness and domain experience of the user, through their statements (or assertions-from-god in the system prompt about the user) — in order for the model to decide whether to trust a statement you think sounds "surprising", vs. when to say "uhhhhh lemme check that."
I personally have found I can often create equivalent code in less English than typing.
Also it works very well where the scope is well defined like implementing interfaces or porting a library from one language to another.
Not much different from image generation really. Sometimes AI is fine, but there's always going to be a need to drop down into photoshop when you really care about some detail. Even if you could do the same thing thing with very detailed AI prompts and some trial and error, doing the thing in photoshop will be easier.
I only took exception to the original statement - that coding is trivial, and the questioning if AI is even useful. So many people are finally able to create things they were never able to. That's something to celebrate. Coding isn't trivial to most people, it's more of an insurmountable barrier to entry. English works - that's why a clear-minded project manager can delegate programming to someone fluent in it, without knowing how to code themselves. We don't end up with them dumping a jar of jam on the floor, because intelligent beings can communicate in the context of a lot of prior knowledge they were trained on. That's how AI is overcoming the peanut butter and jelly problem of English. It doesn't need solutions defined for it, a word to the wise is sufficient.
In the last year I’ve built scripts and even full fledged apps with GUIs to solve a number of problems and automate a bunch of routine tasks at work and in my hobbies. It feels like I woke up with a superpower.
And I’ve learned a ton too, about how the plumbing works, but I still can’t write code on my own. That makes me useful but dependent on the AI.
You wouldn't randomly selected an arbitrary adult from the USA to do a brain surgery on you, so this argument is rabulistic.
But I do expect an arbitrary adult to be able to follow instructions.
Ok. How about you give me a text based task where you would pick the random adult over the LLM?
Do I want to use an LLM to do it from business owner perspective? Yeah probably it is cheaper and more convenient. Which one I want to use, depending the problem we are solving here right?
I'm more concern about the integrity of the current digital infrastructure. In that sense I would NOT trust ANY thing really important to anything digital, much less to LLM. Can I use it for exploration then require an actually human expert approval/edit. Absolutely!
As long as the digital doesn't result in significant physical or financial damage.
Edit: and for HN ppl, of course the LLM will have have to be open weight and all and run locally in a air gaped GPU, preferably in a Faraday cage.
This is rather a red-tape problem. :-)
But if you were to go back to 2020 and ask if your take a random human over a the state of the art AI to answer a text question you’d take the random human every time except for arithmetic (and you’d have to write it in math notation and not plain English).
And if you were to ask AI experts when would you chose an AI they’d say at least not for a decade or two, if ever.
This is key. It works because of previous work. People have shared context because they develop it over time, when we are raised - shared context is passed on the the new generation and it grows.
LLMs consume the context recorded in the training data, but they don't give it back. They diminish it because people don't need to learn the shared context when using this tools. It appears to work in some use cases, but it will degrade our collective shared context over time as people engage with and use these tools that consume past shared context and at the same time atrophy our ability to maintain and increase the shared context. Because the shared context is reproduced and grows when it is learned by people. If a tool just takes it and precludes people learning it, there is a delayed effect where over time there will be less shared context and when the performance of the tool degrades the ability to maintain and extend the shared context will also have degraded. We might get to an irrecoverable state and spiral.
Only if your problem is already uploaded on Github.