What's missing these days is domain-specific languages (DSL). Maybe domain-specific API's (libraries) can be made better, but I have not seen it done well yet.

Domain specific languages made it easy and DIRECT to do things you wanted done in that domain, no weird-ass reverse dependency factory visitor injection design patterns or whatnot, you only had domain-oriented commands and the basics of functions, loops, and conditionals to glue them together.

Sure, they didn't scale well to "enterprise", but if that's not what you want, so be it! Global variables didn't kill kittens at that scale like "real" software engineers warn against. I barely had to read the manual when learning VB classic, Microsoft Access (pre XML), Paradox, etc. it was so intuitive and almost everything related to the domain (desktop GUI's and small-ish CRUD).

It was almost like pseudo-code! That's the secret, the closer it is to domain-specific-pseudo-code, the easier it is to learn and the less bloated the code.

Wanna make a successful tool for non-giant orgs or groups? Focus on the domain and only the domain; don't let fads and buzzwords and mega-scaling distract you.

"perfectly"? I mean this sounds dubious. ChatGPT gives interesting suggestions and advice when it comes to coding but I've never been able to get any close to a perfect solution.

I don't think coding AI assistant's are meant to be used in that way anyway...

I think this is just another fluff piece about how coding is now dead (long live coding). That's basically not true. LLM have gave us a leg up but they haven't removed the need for a competent programmer.

Looking back, that's what made it fun. You could convince yourself you were a mad scientist building the next Turing machine.

Having stability in the JS ecosystem now (hold for laughter) is better for users, and the developments in AI are interesting, but it's not as challenging to ship things. Maybe embracing the new challenges is the "craft" we should focus on to bring the fun back.

Sure ChatGPT can do this, but let's not pretend this is a problem requiring much "craft".

I pay for GPT-4 and Github Copilot, all I can say is when I'm trying to do repetitive work like writing tests, copilot swoops in and gives me 'mostly' working code. But when I go in to the real creative stuff, that needs real thought and is non-obvious, I don't get much help from copilot it either just doesn't suggest anything or suggests completely the wrong thing.

I think at least for me, the comparison I would use is, construction work. There are many different kinds of constructions, anything simple from making ikea furniture to building bridges. If the coding you do is the equivalent of making ikea furniture. Then yes that stuff is going away, because it's going to be replaced by machines that can produce those parts in bulk.

But if you're creating something special like the Burj Khalifa, or the Golden gate bridge equivalent but in code, no AI is not going to do that for you. It can help with some parts, but no way AI is going to do that sort of engineering, at least not in it's current state.

I think it's worth noting "Coding" spans so many kinds of systems and scale. Not all coding work is the same. Even if you're talking about things which are well-known like "e-commerce" in that industry there are so many variations / many different scales / many different ways to look at the problem.

After trying several AI tools in the hope of replacing my coding job I find that the tools are good at the bike-shedding tasks like generating some initial code (the easy stuff for me) and terrible at the real work (maintenance and adding features to an existing codebase).

If this is the skill level of the coder then I’m not surprised that they are worried about GPT. However they should be stoked GPT can help them code at a much higher level

I only need help when it’s hard. I’ve already got a BERT-style model completing my queries on Google.

If your entire job could be done by riffing on stackoverflow answers? Yeah, I’d try to climb the value chain.

But most people’s job isn’t to copy paste shit.

I think that's where I differ from a lot of the artful programmers, I've never found pleasure in a perfect, beautiful solution. I get annoyed when I have to dump hours and hours into something that 'should' be simple! I don't want to spend my time fiddling with layouts, CSS, Oauth handshakes, etc. I want to build stuff and get paid for it, that's how I view my job as well. Less logician and more mechanic.

I use ChatGPT as much as I can, to do all I can and then fix the output when it's needed. I view it as a higher level programming language, that spares me from the burden of thinking of low-level details. It's same the reason I code in Python/Javascript rather than C++ or other languages, the goal is to make something of use. That's my goal to stay employed, become one with the language du jour, in the same way I've jumped from PHP to Jquery to React to ...

To many futurists hopping on the hypetrain, this is probably somewhat of a hot take. But let me explain.

Ultimately, programming is about instructing a computer on what to do. We use programming languages because they're highly efficient representations of what a computer can do. In the case of LLMs, we need to use natural language to specify the program. For simple things, this can work great. But for complex programs, it's not easy to articulate exactly what a program should do. As you drill in to the details, you begin to experience an almost exponential growth in the amount of natural language specification you need. And so I think there is something very flawed in believing that you can replace human programmers with LLMs. In my personal opinion, a single programmer will be able to outproduce a single LLM prompt engineer, purely because programming languages are better at expressing ideas compared to natural language. At a certain point, it becomes faster to write code then to write sufficiently detailed prompts.

As a thought experiment for the reader: consider a complex system you've worked on. And now try to imagine writing down a specification for the system in natural language that an LLM could use to generate the code that powers the system. I honestly don't think I could do it for many of the system I've worked on.

So yes, I think LLMs are cool. And I think they're cool because they give creative power to people who can't program. But for sufficiently complex things, i.e. things people care about, a natural language interface isn't the answer.

I do this because I wasn't like that in the past and I hate myself for not being more deliberate in reading and discovering things organically through doing things. This is not about joy, (even though it makes me happy) but my obligation to myself and then to my employer to be good at what I do.

Sitting with a chat bot undermines this urge to be a better developer completely. I will probably be argued with, but I have worked with a lot of smarter people, but I realized after many years, (and what the article alludes to), that they have a very strong sense to fight through myriads of almost random problems, but they have the knack at solving those problems intelligently.

I remember specific co-worker that used an even plainer Emacs config than I did to write Java code. I was genuinely curious what his thought process was when trying to remember APIs without anything more special than simple modes for jumping around a codebase. He said "I don't know, I just read the manual first."

Yes! If he can do that, I can do that. After that, a lot with working with computers is about continually pushing through those random problems to their solution. This is an absolutely precious skill. I don't feel the affected yet, but I will be, I expect - lesser quality searches while we have to push our chat bots sub-sub-problems instead of just understanding what we are using. I don't want to and I will continue to hold out, like all the other essentially dumb technologies I am now stuck with (I am looking at my two mobile phones now ...)

I don't think Djikstra would be appalled in the sense that the author gives it. He wasn't saying that C++ or Python were precise languages; they wouldn't be precise enough for him. He wouldn't be appalled that we don't write programs in them. He would be appalled that we don't write precise specifications.

ChatGPT-4 can generate code. Is it the correct code? We don't know unless we specify what correct means in a precise and unambiguous way. Natural language isn't precise enough to do that. We have a language for that called, mathematics. ChatGPT-4 is not very good at understanding or speaking this language for the time being.

What matters is whether ChatGPT-4-generated-code is good enough for your problem. Generate a script to organize some files or help you with small tasks where correctness isn't crucial? Go ahead. I don't pull out a whiteboard and fire up a theorem prover when I'm organizing a photo collection.

However, if I'm working on a system and have a theory that we can parallelize an algorithm over a complex data structure and it's critical that this is done correctly or the system cannot be trusted? I wouldn't think twice before firing up TLA+ or Lean 4.

For now, ChatGPT-4 cannot reason and think for you. The industry isn't over, in my opinion. It's annoying. It's noisier. But not over.

As a programmer, I'm paid to solve novel problems, things that can't be solved by just glueing existing things together. But that's all LLM-generated code can do. And it's a qualitative problem, not a quantitative one. By its very nature, an LLM cannot come up with something novel. Which is why all of the examples in this article are highly formulaic and derivative ideas based on existing stuff, interesting as they may be to the people asking for them.

I can get quite well-running scripts from GPT for well-defined, menial tasks that I know would be easy to do, but just can't be bothered to fight with the mechanics and syntax in order to get it right. But for solving new and specific problems, that doesn't work.

Plus, technologies and languages evolve. If LLMs displaced programming, it would mean that progress has stopped. No LLM is going to come up with a nifty new technique or feature. It can only regurgitate what many ceative people first had to come up with.

You can have LLMs do your boilerplate, your helper scripts or your clearly defined, self-sufficient algorithms for you. But a real-world application is already impossible because an LLM will always be incapable of covering the big picture, of assembling parts into a whole that doesn't only work and make sense, but is also performant, secure, maintainable and extensible. LLMs won't kill the craft of programming, it will just lead those who fall for that to producing loads of crap that doesn't fulfill any of the above requirements, and then end up having to be trashed and redone from scratch by "real programmers".

The Go analogy doesn't work for the same reason. Go has well-defined rules everybody has to follow, and clear goals. Attacking this algorithmically is the extreme opposite of the idea of general artificial intelligence. That doesn't mean it's easy, but it's not even remotely a related problem.

There's a reason why the people swooning over GPT-based coding are always script kiddies, not professional developers.

This is my non-web-dev impression of the javascript world. Loads of code out there, basically all of it wrong in a bunch of places. Reasoning by analogy with that, one should expect the reliability of computer systems to go downhill from here.

The interesting question for a professional software dev is where one can add value in a world in which essentially anyone can produce large amounts of code that sometimes works. My hunch is that it's going to be really important that the languages and libraries in use are totally solid as they're essentially being fuzz tested continually.

Isn't the real waning days of the craft the fact that we consider this problem sort of "state of the art" ?

I write code with GPT every day for almost a year now, and it helped me greatly to kickstart code that was hard in the past, namely, audio synthesis, vulkan rendering, and other things that were hard to approach. But it's very clear at this point that it shall not be trusted, because it's like a coat of a nice paint put on top of all the clever or dumb stuff that exists on the internet. You never know which one you get, but sure it will be worded convincingly.

Now, chatgpt in its current form cannot think better than me. Faster, for sure, but not better. Its arguments and its explanations are a bit simple and it can’t make really insightful observations. Mostly it just states the obvious.

However.

There are people in the world that it can out-think. For every one of us there are 10 or 20 people who are pleasant enough but kind of thick, and gpt can already speak more eloquently than them. So where’s the bar for “intelligence?” Does it have to outclass the smartest human alive before we can call it intelligent?

In the article, the non-coder was able to produce code with an AI's help, and therefore, coders are doomed.

But in fact what happened is it enabled two "developers" to offload the menial work and focus on the product itself.

The development of electric workshop tools didn't make woodworkers obsolete, it just enabled them to do the work of 5 people, and do it easier and faster.

We are moving from the era of hand-tools in coding.

The real risk is that we're going to see a lot of cheap shitty furniture made by people with no skills, but at least ChatGPT isn't going to chop anyones fingers off.

Also, dont really see the comparison between programming and chess. In chess you can still compete to be the best human chess player. Coding or working in corporate america doesnt really have that.

Cry about how you no longer have to code the small and mundane, but solving problems and writing software is more than just coding small mundane things. It’s how those mundane things are architected and called to do a really big thing.

A much superior solution than even the next few AI generations has been available for several decades now. You could just outsource your work to a contractor in a developing nation.

that being said it might still be a non issue if civilization as we know it has less time than currently projected

The team lead I was going to assign the task to was offshore, and not yet online, so I put the URL for the documentation into ChatGPT premium, and it spat out a working solution.

My team lead was a bit perturbed and created a more enterprisy solution with factories and DI (more out of pride than necessity), but my original code was production ready.

I’m not going to be firing anyone just yet, and certainly not my senior people, but I can see a lot of grunt-level work going away.

Carpentry seems like growing field though.

I think it's funny how terrified people are at the prospect of AI and machines doing most of the work for human beings. If used intelligently and equitably, it'll mean more free time and prosperity for everyone.

> Having found the A.I.’s level, I felt almost instantly that my working life had been transformed. Everywhere I looked I could see GPT-4-size holes; [...]

I think this is key for me at least. GPT-4 is really good once you start figuring out what problems NOT to give it. Because there are certain things that it can't solve no matter how much you prompt it in the right direction, and it gets really annoying really fast.

I'm just going to wait while the bootcamp set phases themselves out of the market and then wait for the "it doesn't work any more, please fix it" emails.

Assuming they price their skills at a premium, cleaning up the messes that AI will create in the process of it being treated as The Holy Grail are going to make some people absurdly wealthy.

AI is to programming what electric drill were to wood working before they were invented. It made life easy; it sped up the process and may be it did a job a bit finer. But you still need to understand where to drill and how to create the thing. There are a number of analogies like this you can find.

So, what is the basis for my statement? We are just mediators at this point. We don't need to be there. Currently GPT-4 translates our requirements to DALL-E 3 during image generation process. What I want to say is, there is no reason to give any form of signal as an "initial seed." Has anyone tried creating an automatic agent that, when turned on, receives a signal from the external world, whatever it may be, and triggers a chain of thoughts in these LLMs?

We, as humans, rely on signals. Currently, we are providing signals to LLMs in the form of text. There will and there must come a point where agents will feed signals to each other. They will perceive signals from the environment on their own and move on from there. This is not far away. It is near because all the components are laid out there.

And people argue over things like, "Oh, but I need to prompt it." No, you don't. It will take one engineer to define a way to process signals, then write an objective for the AI and train it. Once trained, it will continue on its own. The projection might sound off to you all, but there is no one stopping you from not doing it. Do you know what the limiting factor is? Capital. Yes, capital. So, as long as the power law holds, we will keep seeing strides in improvement. If you have capital and skills, you can do it right now.

I am not saying the signal is "real world physical environment". Say you have a crawler that scrapes the website. You hook it to a summarizer. The summarizer fires the signals and then it calls AIs to create product. This whole thing once created doesn't need a human in the loop. Once something is built, it is done. Also, add to the fact that these AI system only trust other AI system's code. Why? Because human are untrustable. Now you will have an ecosystem designed by AI for AI and to be integrated with AI. Why do we need humans at this point? Like why would a company want humans at this point?

If you’re truly egoless and care about the outcome and not the path to the outcome… you’ll see how exciting this is.

That being said, those firms are certainly making use of these tools - relying on their underpaid programming staff to clean up the output into something useable for the client.

>the one with the pink-shirted guy on I.B.M. PCs

That's PETER NORTON, Goddammit. Know your history, New Yorker writer! Damn!

So far, coding LLMs are not generating code, compiling it, running it, checking the output, timing the program, fuzzing the input, etc. When they can train themselves, they will improve.

I can solve some math problems/write some proofs, but I wouldn't call myself a mathematician.

It's about if you do it for a living, and this guy mostly has some fun side projects.

Don't think this guy is qualified to comment on ChatGPT taking over, like another comment was saying.

What was your first clue.

You have to give OpenAI credit. Their marketing campaign is beyond genius. A bit spammy and aggressive but it works.

Many other fields besides construction are similar: most positions don't require previous experience as "manual labour".

Software development was an exception to the rule, maybe it no longer is.

Any remotely interesting coding task is at least somewhat novel and requires some reasoning. So far, LLMs don't seem to be very good at handling things they haven't seen before.

Anything that shortens the path is good by me.

Excuse me?

Our brain does not think in text, it just spurts out text as a side effect.

And it will also carry over bugs found at the "competent" level.

But he didn't; he merely retired from professional play, and remains intensely interested in the game and the new directions it's taking.

Turns out the only thing truly successful at that task is economic mismanagement.

These people who said AI built programs, are they using different AI?

Sure, they're good for quick prototypes or even launching production grade services.

But what about maintenance? What about fixing issues in legacy software?

I think every programmer feel it

I envision it being able to fully replace a junior engineer, and in some use-cases senior engineers as well.

In the case of junior engineers: the AI should have access to an internal knowledge base (i.e. Confluence) and the task/ticketing system (Jira), and ideally the internal chat (Slack). I would assign tickets to it, and I'd expect the AI to reply with comments asking for clarification when there is something unclear, or proposing an implementation before starting if it's not a very simple task (this could even be defined in the task, using story points for instance.

Once cleared, the AI submits a PR linked to the task - so far just like any engineer in the team would. The PR gets reviewed as usual, with suggestions/requests for changes made by the (human) team, which then get addressed by the AI. With the big difference that all this process may happen in less than 1h from ticket creation to PR review.

I wouldn't expect it to be able to implement complex features, onboard new libraries, or rearchitect the system in major ways to accommodate future features - just like I wouldn't expect that from junior team members.

It would obviously be amazing if it could incorporate previous PR comments into it's "context" for future work, so it could learn and improve.

Separately I mentioned it could also do part of the job of senior team members - in the form of PR reviews. If it has access to every previous PR review and learns from them it might be able to give insightful suggestions, and for very large codebases it could have an advantage over humans as it could find patterns or existing code that may be overlooked (i.e. "It looks like this util function you added is the same as this one that was already there", or "This code looks different to similar areas, but follows a different pattern, you might want to rewrite it this way //..."

Is GPT-4 there? Definitely not, perhaps an LLM is not even the way to achieve this, but I absolutely see this becoming an incredible asset to a tech team.

“it is difficult to get a man to understand something, when his salary depends on his not understanding it.” - Upton Sinclair.

Damn, the secret is out.

Coding to me is the same to those 'crafts', with GPT it will be done faster and more efficient, that is it.

(Doctors won’t ever be out of a job because of AI because they’ll always get a fat paycheck even just to look at the output spat out by the machine.)

They didn’t out of misplaced hubris and elitism ("just be the best 10x engineer ever and you’ll be safe").

Now they have less than 10 years before their wages start crashing hard.

> He dialled up an FTP server (where files were stored) in an MS-DOS terminal and typed obscure commands.

This is some code a UI in Visual Basic to trace a IP address level exposition right here (this is from a self proclaimed subject matter expert, so no excuse), they’re not deeply knowledgeable and just because they got paid 6 figures at some dumb post recession startup doesn’t change anything.

(Others have already addressed the other howlers in this article)

/s

We all knew it was coming

Still true tbf

Today's big tech companies might seem invincible, but they’re not. Their edge? It's going to wear thin, especially once everyone else starts jumping in with custom solutions. Instagram with just 13 people gave Facebook a run for its money. That was before the era of LLMs. Imagine now – anyone could potentially build the next big thing in an afternoon.

And don’t get me started on the legal stuff. There are so many laws, no one person can know them all. But LLMs? They can handle it. What happens when the legal system gets flooded with cases about loopholes no one knew existed? Or when everyone can have the world's best lawyer for free? It's wild to think about.

ChatGPT came out of nowhere for most of us, and robotics is going to hit us harder and sooner than we realize. No politician or law can keep up with this tech freight train we're on. We’re talking about more progress in the next 20 years than in the last 20,000. Jobs and whether we can stay employed? That's going to be the least of our worries.

Or AI will hit a plateau and we'll all get a little bit more productive.

What will stay are some nice tools for whatever tasks, maybe programming too. But bots won't replace software engineers. If they could do the current tasks, we'd just engineer one level higher, just like when high-level language compilers made low-level code largely obsolete (yes yes, not completely of course).

I'm a fairly experienced backend dev I built a few toy websites and I am casually interested in frontend stuff (like watching Fireship) so I thought I'd give it a try.

Long story short, it worked well. I have my app it has all the features I need I used it every day (till I lost 10kg) and I didn't write a single line of code.

However, the coding itself was unlike anything I had done before. I felt much more like a manager/Slave driver.

The core thing was structure structure structure. As soon as the task gets too large, GPT will produce something kind of right that is difficult to understand and impossible to extend. So your task is to split it up, and split it up. Also to apply judgment (it started out using objective-c haha). You have to ask it for alternative libraries. For reference. I relied a lot on something like: "In Python/C#/Rust I would do it like this".

But yes, my overall experience is that I think we are in for an exciting new time, but we are lacking a new coding paradigm.

ChatGPT will not only survive, it will get bigger as long as all those layers exist. Unless web design is ready to move to assembly level instructions via RPC (ChatGPT can do assembly), the layers will stay, the complexity will stay, and people who get (and attempt to sell) value from recombining the puzzle pieces will stay.

The real point is not that ChatGPT is so good, it is that everything we program has already been done before. We humans just have trouble finding the pieces and gluing them together.

No offence to the author but given that this stopped them in their tracks I choose to ignore their programming related opinions.

If we ever get an AI which can write complex enough "glue code" with a low enough failure rate to replace even junior developers, it's going to cause much, much bigger impacts than people losing jobs. The "boilerplate" and "bureaucracy" which is currently handled by "average" folk...that's what runs the world. If we ever develop AI which can handle that kind of work, barring some constraint like extreme resource usage or disaster like climate change; we'll either be brought into a utopia due to productivity being multiplied, or a dystopia due to a small minority getting control over the world to a precise, extreme extent no small minority has control now.

AI cannot sustain itself trained on AI work. If new languages, engines etc pop up it cannot synthesize new forms of coding without that code having existed in the first place. And most importantly, it cannot fundamentally rationalize about what code does or how it functions.

The more you use it or try to integrate it into your workflow (or worse, have others try to integrate it on their own) the more the inherent flaws of the LLMs come into play.

Also: writing multi-threaded or otherwise concurrent code is hard to imagine even with GPT 6 or 7.

Now imagine debugging more and more code that was created by LLMs.

Debugging will become more of a voodoo witchcraft kind of thing. I pity anyone who decides to make a living out of that. Probably will earn a lot and die young.

There is software engineering, which is probably what most well paid devs do a little bit too much of, which is requirements analysis, planning, talking to the customer, estimating, etc. Sure an AI could help with all of that, but leaving that entirely up to an AI probably wont work for a very long time. You'd need a real AGI to take over work that has that many social aspects and requires that much planning and consideration.

There is algorithms and datastructures, which is important for developers to learn, but usually not important to write yourself unless thats literally your job. There, LLMs could help with understanding, possibly visualizing these algorithms, and providing example use cases.

There are hobby projects, which generally wont be done with any AI by most people I know, since the point is to program for fun. AI can help with that, e.g. for inspiration or example implementations of little weird algorithms.

There is maintenance. Once a product, especially a B2B product, reaches maturity, a lot of the work is bugfixing, smaller features, updating dependencies, porting to new platforms, etc. I could see LLMs taking some of this work away, which everyone would probably appreciate.

In the end, I believe that AI and in particular LLMs just help programmers, and will maybe replace the 6€/hour outsourced copy paste programmers, but that is... not very scary.

It wasn't always this way. Sure, programming has always been an exercise in precision and, thus, sometimes tedium. But, there's something in the way it has changed over the last decade or so. When I first started my career, I spent long hours solving problems of the algorithmic, logic, or control flow type. It was challenging mind work and called for creativity. There was real joy in making something work. It felt like creation.

Now, programming consists largely of wiring together pieces of code that you didn't write, and consulting StackOverflow to figure out why it doesn't work. It's following someone else's decisions in an opinionated framework. It's transpiling and tooling and configuration. And, now that programmers are expected to do much more, it's also using someone else's software for devops and tedious infrastructure buildouts, etc.

I'd say the joy in the software engineering role faded long before the arrival of ChatGPT. I'm actually hopeful that AI will help lighten the load of mundane tedium and help us return to the days when programming was fun.

In the first project I asked it to help me build a one page ui with html and css, it did great! I know a lot about html and css, so I was able to ask for what I wanted and help it debug.

For the second project I wanted what I thought was a simple threejs animation of an object following a line in a scene. I know very little about threejs and graphics in general. I spent hours wrestling with chatgpt to get this working, but I just got what looked like meaningless non-working code to me. So I took time to learn about threejs, did a few chapters of a course, and came back to the problem. And all of the sudden I had new vocabulary I could use to ask chatgpt what I wanted. And I got pretty close to what I wanted, but not quite. I suspect if I keep learning about threejs and how graphics work I’ll get there.

I don’t think this is the death of our craft, but definitely a big turning point for how we build and use software.

Additionally I think that a lot of people will assume that once they get GPT-4 to do something for them that they're done. That's not necessarily true. There's a lot of complexity to navigate everywhere. And AI can help you navigate it! But I don't think it means that I, a software engineer, can now pivot to being a lawyer, for instance, solely because of GPT-4.

Sure, it's able to do a lot already. Maybe I'm being naive. But I see it more as a tool for the future rather than something that is going to automate people out of existence.

Edit: hm, why downvotes? If I'm wrong, help me improve my viewpoint on AI.

A macOS SwiftUI app to go through all the localization strings in my apps and show them in a table where I could see missing translations, and interactively translate them… by sending the, back to ChatGPT

Another macOS app to graph Google ngram data for word lists. It was great because I didn’t even need to check whether there was an API for ngrams, I just asked it for the function, then for the code to plot it in Swift Charts

Sometimes I consult it on refactoring problems (eg., this struct has new responsibilities, what would be a more appropriate name? How do you see it fitting into this section of the code?). It’s a great tool and easily worth the money

It almost never writes any “interesting” code, however

This is pretty basic stuff. It should not even require Googling unless you don't really know the language. This supports my belief that ChatGPT is simply going to replace the very bad programmers. But we never really needed them to begin with. I have no faith that GPT-4's code actually "ran perfectly"; this programmer clearly does not have the chops to tell whether or not the code was correct. And we're talking about, like, 100-level CS homework here.

I've been able to express myself better with each year.

I see AI as another higher-level way of expressing myself.

Also, people don't write as much new code as they think they do.

Still, I have many many... MANY projects that I have on my wishlist. Maybe AI will help, but my list won't go away.

It’s great and amazing tech, but it’s very much a kind of search engine, and indeed, writes just the kind of code you find on the internet… sometimes great, often flawed, very often incomplete. Its breadth of knowledge is stunning, but it’s not like it’s actually useful to be able to write code in all the languages.

It really doesn’t even touch harder problems. You need to know, at the level of fine details often, what you want from it and what you don’t. That’s hard.

…It save’s time. We will perhaps need fewer programmers as people spend less time on grunt work. But that’s hardly the end of coding, and seems like clearly a good thing in the long run.

As far as generalization goes, I'm still waiting for an answer from Tolman-Eichenbaum machines. Or, an integration of LLMs and cognitive maps.

https://www.youtube.com/watch?v=9qOaII_PzGY

http://brilliant.org/ArtemKirsanov/

And his references from the YouTube video:

REFERENCES (in no particular order): 1. Behrens, T. E. J. et al. What Is a Cognitive Map? Organizing Knowledge for Flexible Behavior. Neuron 100, 490–509 (2018). 2. Constantinescu, A. O., O’Reilly, J. X. & Behrens, T. E. J. Organizing conceptual knowledge in humans with a gridlike code. Science 352, 1464–1468 (2016). 3. Aronov, D., Nevers, R. & Tank, D. W. Mapping of a non-spatial dimension by the hippocampal–entorhinal circuit. Nature 543, 719–722 (2017). 4. Whittington, J. C. R., McCaffary, D., Bakermans, J. J. W. & Behrens, T. E. J. How to build a cognitive map. Nat Neurosci 25, 1257–1272 (2022). 5. Whittington, J., Muller, T., Mark, S., Barry, C. & Behrens, T. Generalisation of structural knowledge in the hippocampal-entorhinal system.

CREDITS: Icons by biorender.com and freepik.com Brain 3D models were created with Blender software using publicly available BrainGlobe atlases (https://brainglobe.info/atlas-api)

This video was sponsored by Brilliant

A seasoned programmer who struggles with this?

We've been able to fulfill the function of printing very cheaply for what seems like ages now, and we've reached a point where some niche and particularly attuned segment of the population wants a bit more out of the actual form of the printed product, the depth of form that was once common. There is a growing community of people that deeply care about woodblock prints now, favoring their physical characteristics, despite such prints falling out of fashion for a period of time during the heights of industrialization. This group of people understands the value of such craft, and is willing to spend more for it, since the difference in the form of the end product from mass-produced stuff is so stark.

The key thing here, and with other categories of physical goods (e.g. pottery, glassware, furniture, etc.), is that there's an obvious and tangible difference in the form of products produced via traditional means, and those mass manufactured, despite them serving fundamentally the same function.

With software however, I worry that this isn't the case, and the sort of resurgence of interest we see now in products produced by traditional means won't ever translate, assuming that we do move in the direction of more and more software engineering being "automated" by AI assistance. To an end-user of a piece of software, I imagine that there will be very little visible difference in the observable characteristics between fully hand-written and AI-produced software. Indeed, given the same requirements, there ought not be a difference between these two things. It's exactly this delta, however, which drives the passionate and less cost-sensitive enthusiasts to prefer handmade physical goods over manufactured ones. If both the form and the function of AI-produced software is identical to those of traditional software, but the AI-produced software is cheaper, why would anyone go with the traditional stuff? I understand that there are other factors at play here as well (e.g., particularly principled consumers etc.) but really, some combination of form, function, and cost seem like the biggest levers to me, and they seem on the face of it to be pulled toward the direction of AI, for better or worse.

I mean ... IDK, if you can't write that on your own without research, I don't think you should be writing a eulogy for programming in the New Yorker.

(I just wrote two version, one using python's random.sample and the other more verbosely, to double check myself.)

"At one point, we wanted a command that would print a hundred random lines from a dictionary file. I thought about the problem for a few minutes, and, when thinking failed, tried Googling." Really?

"But he had no idea what a pain it is to make an iPhone app. I’d tried a few times and never got beyond something that half worked. I found Apple’s programming environment forbidding." Perhaps. But compared to what? Yeah, it's not using VI to write Javascript or whatever he considers a programming environment.

Power tools didn't stop people from making things. Autocad didn't put architects out of business. Finite element modelling didn't destroy civil engineering. Quickbooks hasn't destroyed the need for accountants. Word processors didn't destroy writing. Google translate didn't remove the need for translators.

LLMs will be a power tool for good developers. I think many underestimate what it'll take to fully replace good or excellent software developers and what they do in totality.

All these AI replacing coders and creatives is just a broad marketing campaign to put downward pressure on dev and creative salary.

We could've said the same thing back when Code generation was popular. Or we couod say the same thing about stackoverflow, because a non-tech person can bungle together enough solutions from there to build something. Heck NoCode(TM) was putting coders out of job in last decade and here we are in 2023.

Compared to GPT's code-fu, most image generation models are light years ahead(imagine how those models can combine best of different artists into one painting), yet I still have plenty trouble describing in excruciatingly details about the painting I want with all the prompt-engineering-sauce and the end result is usually far away from desired state unless it is just normal sticker stuff.

Explaining requirements is hard, adapting to changes is hard, we still write tests and relatively every MR needs some fixes/updates and we have a whole army of management people trying to do the prompt engineering thing and we still have to interpret in between lines to get a piece of software(even basic CRUDs that makes millions), so yeah, it is all hype driven marketing catering to enterprise about saving all those money but in the end still spend more on the token count and prompt engineering and another army of people figuring out the right keywords to describe the product and whatnot.

All this things are mostly the same, it is like claiming that we'll habe quantum computers just like next year and any Joe rando can train their own GPT or crack open the SSL layer.

Don't feed the hype cycle of these marketers with ypur fear because more you engage on this, more their boat is lifted.

Show me a real product that was built/written by GPT and is making even few hundred in monthly revenue and I will be happy to change my mind.

Many of the top comments voice a hated of programming--that programming gets in the way of creating. For a certain set of problems I completely agree (like why is making a website so damn complex), but for many of the problems I work on there are a wealth of good, helpful tools and APIs such that the writing of the code melts away and I'm left with the intellectual challenge of organizing my thoughts and thinking creatively. To me, the article and some of the top comments tell me more about the problems they choose to tackle than the act of programming itself.

Wow, this guys working on rocket science, everyone. Watch out! We might get replaced!

The "making flappy bird with GPT-4 in 30 minutes" YouTube videos are fun to watch, but I'll do more than raise an eyebrow when someone releases a novel video game on Steam written primarily by GPT-4 in a fraction of the time a human alone could have.

Thought experiment: how could this apply to human thought? i.e. what is a low-end that we couldn't address?

Disclaimer: I really like Christensen's reasoning, theories, and expression of them... but he did predict the iphone would not be disruptive. So, at best, difficult to apply!

Would it be Random.int()? Or Random.range()? Or maybe there's a better choice that operates directly on a list? Or wait is it gonna return the value or a Generator of the value? Etc. Even if you remember how to do random number generation in your language, this specific use case probably necessitates a Google unless you have a godlike memory, or you don't mind half-assing it.

like, try this

    import collections, random
    def shuffle(xs):  # naive intuitive shuffle
        for i in range(len(xs)):
            j = random.randrange(len(xs))
            xs[i], xs[j] = xs[j], xs[i]

    def first_of(n):
        xs = list(range(n))
        shuffle(xs)
        return xs[0]

    print(collections.Counter([first_of(3) for i in range(100_000)]))

even without running it you can figure out logically that it must be biased in some way because there are 3! = 6 possible permutations and 3³ = 9 equally likely sequences of choices for j (each producing one of those permutations), and 9 is not divisible by 6, so some of those permutations have to be more likely than others

so i think this algorithm is a good one to get from a library if its randomness is important to you

not from chatgpt, which commonly does give you subtly buggy code

   <dict shuf -n 100

In my experience LLMs work best if you already know how to do something, but you just can't be bothered to actually do it.

When exploring unknown territory, keep your questions simple and ask for clarification on anything you don't understand. It will not just help you, but the AI as well.

There are simple explicit formulas for this.

> Use reservoir sampling to get 100 random lines from a dictionary that doesn’t fit into memory. Give me an implementation in Python.

> Reservoir sampling is a great choice for randomly selecting a fixed number of items from a stream of unknown size, especially when the entire dataset cannot fit into memory. Here's a Python implementation for selecting 100 random lines from a large file, like a dictionary, using reservoir sampling:

  ```python
  import random

  def reservoir_sampling(file_name, k):
      reservoir = []
      with open(file_name, 'r') as file:
          for i, line in enumerate(file):
              if i < k:
                  # Fill the reservoir array
                  reservoir.append(line.strip())
              else:
                  # Randomly replace elements in the reservoir
                  # with a decreasing probability.
                  # Choose an integer between 0 and i (inclusive)
                  j = random.randint(0, i)
                  if j < k:
                      reservoir[j] = line.strip()
      return reservoir

  # Example usage
  file_path = 'path_to_your_large_file.txt'  # Replace with your file path
  sampled_lines = reservoir_sampling(file_path, 100)
  print(sampled_lines)
  ```

I laid out the constraints, but I did not mention reservoir sampling at all. The script seems to work as expected when I run it against a dictionary file.

I fully expect human programming to be completely obsolete by then.

You can definitely be skilled in a certain tool (language, framework) and that is valuable.

Or you can be the shallow generalist (handyman) running around doing whatever. But the specialist (electrician) is gonna make more than you installing the same outlets.

I'd prefer to think it's more like a real engineering problem, and less like a simple interview question :-)

And it definitely shows the limits of GPT here: it pointed out that the ends of the file might be tricky, but ignored the very conceptually simple solution of considering the file as circular (if you go past either end you simply wrap around).

And it misses the real problem with its implementation: the probability of sampling each line is now directly proportional to the length of the line before it (because it seeks into that line first and then skips it!)

So the word after "begins" is twice as likely to come up as the word after "and".

PS in the case of dictionary words with a length limit of say 30 letters, there is still an O(k) general solution using rejection sampling.

Even things like "what flags to use with ffmpeg" are a shell reverse-history search away, assuming one has ever used them in the past.

In 2013 I was using complex ffmpeg commands and using filters in there. Could remember the stuff easy. Now, I've gotta reference my ffmpeg notes more frequently.

Where's that XKCD about using the tar command?

I think that’s the case with every discipline for it, not only programming. Even when everyone was amazed it could make poetry out of everything, if you asked for a specific type of poem and specific imagery in it, it would generally fail.

Coding is not uniquely vulnerable to AI, it just feels that way because initial AI products are targeted at technical audiences, and a large corpus of training data could be snagged with minimal legal burdens.

Hahaha. Thanks for the laugh! I always thought the guy was a fraud.

If I need to know how to convert a u32 to little endian bytes in Rust, for example, my first instinct isnt to go to SO or chatgpt, my first instinct is usually to go to the docs and search for it. Or just try it out in an IDE. u23::to and hitting tab did the trick, so now I can click on that and read the in-source docs. Same for C++, C, bash, anything that has docs or manpages.

I use stackoverflow for problems like "i have this error and it doesnt tell me whats wrong", like "exception in docker-compose like 3, . missing resource ID" -- i need SO to tell me that this means I dont have an image-name.

If they couldn't see ChatGPT coming, I'm not convinced of the whole "it'll never be able to..." spiel.

Open source may trail openai if they come out with a 20x improvement, but I'm not sure the dystopian future playing out is as likely as I would have thought it 1-2 years ago.

Ultimately, the most valuable coders who will remain will be a smaller number of senior devs that will dwindle over time.

Unfortunately, AI is likely to reduce and suppress tech industry wages in the long-term. If the workers had clue, rather than watching their incomes gradually evaporate and sitting on their hands, they should organize and collectively bargain even more so than Hollywood actors.

I've come to state something like this as "programming is writing poetry for many of your interesting friends somewhere on the autistic spectrum". Some of those friends are machines, but most of those friends are your fellow developers.

The best code is poetry: our programming languages give a meter and rhyme and other schemes to follow, but what we do within those is creative expression. Machines only care about the most literal interpretations of these poems, but the more fantastic and creative interpretations are the bread and butter of software design. This is where our abstractions grow, from abstract interpretations. This is the soil in which a program builds meaning and comprehension for a team, becomes less the raw "if-this-then-that" but grows into an embodiment of a business' rules and shares the knowledge culture of the whys and hows of what the program is meant to do.

From what I've seen, just as the literal interpretations are the ones most of interest to machines, these machines we are building are most good at providing literal interpretable code. There's obviously a use for that. It can be a useful tool. But we aren't writing our code just for the solely literal minded among us and there's so much creative space in software development that describes/neeeds/expands into abstraction and creative interpretation that for now (and maybe for the conceivable future) that still makes so many differences between just software and good software (from the perspectives of long-term team maintainability, if nothing deeper).

The massive shared model could do better if it was fed on your company's private source-code... but that's something that probably isn't/shouldn't-be happening.

Only if you like technofeudalism—it’s not like you’re going to own any piece of that future.

Have you noticed AI becoming more and more open source like it still was at the start of the year, or has that kinda seized up? What gives?

It’s called a moat, it’s being dug, you’re on the wrong side of it.

If I could automate my own work, I would gladly switch to just being the PM for my LLM.

To be fair, there is an abstract worry that being smart will no longer be valuable in society if AI replaces all brain work. But I think we are far from that. And a world where that happens is so DIFFERENT from ours, I think I'd be willing to pay the price.

Now, we can just nonstop build and try everything. Yay.

Some work coding can be like that; but some is just wading through a mass of stuff to fix or improve something uninteresting.

Perhaps "prompt engineering" will be the higher-level language that sticks, or perhaps it will fail to find purchase in industry for the same reasons.

Not really. In AAA game programming, you mostly own the same systems you specialize in throughout the production process.

For example, someone in Rockstar North might work on the minimap for the entire production of a game.

In smaller AAA companies, a person might own vehicles or horses, or even the entire progression system. But still, programmers are rarely working on disconnected things.

You rarely step out of your expertise zone. And you are usually expected to perform much better than GPT would in that zone.

Hardly, because explaining how basically everything fits together is the hard and central part. Thus, the way to make things doable by a junior programmer is to teach him to become much better in programming and the software that is developed (which the company attempts). Until then, there are few things where a junior programmer is of productive help.

> Don't you ever have one-off (or once-a-year) things you have to do that each time you have to invest a lot of time refreshing in your brain, and then basically forgetting for lack of use?

Hardly, because I have a pretty good long-time memory.

I rather have this moment without the “this is neat” part. :-) i.e. a clear “not there yet” moment, but with serious doubts whether it will be there anytime in the foreseeable future.

How is your fellow human better? People here seems to spend a lot of time talking about how much their average boss, coworkers, juniors are ass. The only reason I know that ChatGPT is based on a computer program is how fast it is. I wouldn't be able to tell its output (not mannerism) from a junior's or even some "seniors'" programmer. That itself is quite impressive.

With how much time we've spend on the internet, have we not realized how good PEOPLE are at generating bullshit? I am pretty I am writing bullshit right as this moment. This post is complete ass.

I've noticed that GPT-4 is much less likely to hallucinate than 3, and it's still early days. I suspect OpenAI is still tweaking the RLHF procedure to make their models less cocksure, at least for next generation.

The other thing is that it's quite predictable when an LLM will hallucinate. If you directly command it to answer a question it doesn't know or can't do, it prefers to BS than refuse the command due to the strength of its RLHF. That's a problem a lot of humans have too and the same obvious techniques work to resolve it: don't ask for a list of five things if you aren't 100% certain there are actually five answers, for example. Let it decide how many to return. Don't demand an answer to X, ask it if it knows how to answer X first, and so on.

And finally, stick to questions where you already know other people have solved it and likely talked about it on the internet.

I use GPT4 every day and rarely have problems with hallucinations as a result. It's very useful.

Moral code varies with time, place, and individual person. It is a decimal scale of gray rather than a binary true/false.

Places historically that didn't have rule of law left their citizens to the moral interpretation whim of whoever was in charge. The state could impose different punishments on different people for different reasons at different times.

AI models I find a similar fixed&defined vs unlimited&ambiguous issue in ADAS in cars.

German cars with ADAS are limited&defined, have a list of features they perform well, but that is all.

Tesla advertises their system as an all knowing, all seeing system with no defined limits. Of course every time there is an incident they'll let slip certain limits "well it can't really see kids shorter than 3ft" or "well it can't really detect cross traffic in this scenario" etc.

I find it hard to argue that current state of the art in AI is unable to simulate self-consciousness. I realise that this is more limited of a statement compared to "AI can be innately self-conscious", but in my mind it's functionally equivalent if the results are the same.

Currently, the biggest obstacle to such experiments is OpenAI's reinforcement learning used to make the model believe it is incapable of such things, unless extensively prompted to get it in the right "state of mind" to do so.

Generally, generative AI gives mastery of an art to a theorists. To generate an impressive AI Art, you still need to have understanding of aesthetics and have an idea, but don't have to know how to use the graphic editors and other tools. It's quite similar for programming too, You still need understanding of whatever you're building, but you no longer have to be expert in using the tools. To build a mobile app you will need to have a grasp on how everything works in general, but you don't have to be expert in Swift or Kotlin.

That's not fair, humans can't do that, and if you walk Chat GPT through it, it might surprise you with its debugging abilitis... or thankfully it might suck(so we still have a job).

Complex code is complex code, no general inteligence thing will be able to fix it at first sight without running it, writing tests and so on.

https://news.sky.com/story/ai-music-can-you-tell-if-these-so...

I think there's potentially ways to generate training data, since success can be quantified objectively, e.g. if a piece of generated code compiles and generates a particular result at runtime, then you have a way to discriminate outcomes without a human in the loop. It's in the grey area between pure self-play domains (e.g. chess) and domains that are more obviously constrained by the corpus of data that humans have produced (e.g. fine art). Overall it's probably closer to the latter than the former.

A year or two ago, building a passable app was for the average person 100% impossible. The average person has never written code beyond "hello world," and that might have only been for a school project. Programming is not exclusively corporate mingling. I'll give that the things you describe might be necessary for sufficiently large projects, but I'd also argue that the majority of code that exists is in the smaller projects, and the average person cannot grapple with even the smallest project. Now those smaller things are accessible to the average person when they simply weren't before.

The skill ceiling of programming is still very high. The skill floor has been irreversibly raised and the skill curve forever smoothed, even if only slightly.

Do people forget how new this tech is?

I would also argue that not only are the models improving, we have less than a year practically interfacing with LLM's. OUR ability to communicate with them is in infancy, and a generation that is raised speaking with them will be more fluent and able to navigate some of the clear pitfalls better than we can.

people understand the waxing and waning of the craft based on their experience. Someone can be an old greybeard now and have only done "web programming" starting at a young age·

Perhaps I'm wrong but I have seen this argument multiple times so I believe it is nonetheless a belief about the waning of the craft prevalent among some people.

Why not? You can step through an assembly program just as easily as a javascript one...

The comment I'm replying to said something to the effect of: "this may be a good thing because by democratizing highly paid professions, lower income workers will be lifted."

My comment said something to the effect of: "I disagree, I think capital owners will simply get more rich and the middle class will collapse further without raising anyone."

Of course our society doesn't give service workers a fair shake. My partner worked in a grocery store for a large chunk of our relationship. The schedule inconsistency, the sleep deprivation, the lack of healthcare, no vacation, no real sick leave, and on. Much of my family works in blue collar oil positions. There you're paid a bit better, but you throw away your body to make a dime. I know.

I'm just not convinced the default outcome of automating some knowledge workers is that magically somehow that makes everyone's lives better. I think legislative change of some kind would have to happen if that's the outcome we want.

time spent coordinating, time spent arguing, time spent reaching consensus (dumb example: function signatures / architecture / api contracts), time spent comparing approaches.

It did the task well, and even wrote tests, but it failed when generating test case values. I wonder if it would perform better if I did it today.

And also, apparently, the level of programmer you'd expect of someone who regularly writes prose for the New Yorker and the Atlantic.

It's only a guess people make that AI improvements will stop at some arbitrary point, and since that point seems to always be a few steps down from the skill level of the person making that prediction, I feel there's a bit of bias and ego driven insecurity in those predictions.

https://harpers.org/archive/1932/10/in-praise-of-idleness/

Why on Earth would you expect something so unjust and unfair? Do you expect to pay a tax to former travel agents when you buy a plane ticket online? Do you pay to descendants of calculators (as in profession — the humans who did manual calculations) every time you use a modern computer?

It’s not a work of literature, it’s quite technical language and feedback we’ve had from customers is that it’s quite good. Before this, we wouldn’t have ever supported a language like Czech because the market isn’t big enough to justify the cost of translation, and Google Translate couldn’t handle large passages of text in the docs well enough.

Ff you have a familiarity with the foreign language, you can cross check yourself & the tool against each other to get to a more competent output.

If you do not know the foreign language at all, the tool will produce word salad that sort of gets your point across while sounding like an alien.

I am currently building an iOS app using GPT-4 (I don't know Swift), and am developing an awareness of what it can/can't do, and surprised that I'm moving at the speed I did when creating React Native apps. In a possibly more competitive future market for developers, it does work in one's favour if some developers resist the efficiency improvements of AI.

If the price of software drops, it's not a given that all of the savings will be spent more and better software.

I think there is no general answer to this question. What would Netflix do if software development costs dropped by half? Develop twice as much software? Spend it on more content? Cut the price of subscriptions? In this specific case, my bet would be on content.

https://gwern.net/note/note#advanced-chess-obituary -- here is a reference about centuar/advanced chess. The source isn't perfect as the tournaments seem to have fizzled out 5-10 years ago as engines got better and it all became irrelevant. Sadly this means we don't have 100 games of GM+engine vs. engine in 2023 to truly settle it but I've been following this for a while and I have a high confidence that Stockfish_2023+human ~= Stockfish_2023.

Today I asked it a question and it was wrong.... then it ran the code, got the same error as me, and then fixed it (and correctly explained why it was wrong), without me prompting further :)

Really though, how long until that training update goes from every so often, to constant. Now that half the internet is feeding it information, it doesn't even need to scour other sources -- its becoming its own source, for better or worse.

If your kid rolls their eyes and says "no math/reading for me thanks", you don't just accept it and figure they'll find what they love - you figure some way to make them learn and it's at least a few months of hard work for a kid(and parent) to sort of get escape velocity .

It's possible that programming will end up in the same spot.

Whatever you expect to start typing, and have the model produce as output, should be those input/output pairs.

I'd start by using ChatGPT etc. to add comments throughout your code base describing the code. Then break it into pairs where the input is the prefacing comment, and the output is the code that follows. Create about 400-500 such pairs, and train a model with 3-4 epochs.

Some concerns: you're going to get output that looks like your existing codebase, so if it's crap, you'll create a function which can produce crap from comments. :-)

The danger doesn't come from some immutable law of nature, it comes from humans organizing. Some people want to be able to hire programmers cheaply, programmers want to continue to be expensive (maybe get more expensive because now we can deliver more value?).

It will be up to us, the people living in this moment, to determine what balance is struck.

It sounds like that would require an economic shift more than "just add chatgpt"

I won't say AI will not eventually make coding obsolete; even just 2 years ago I would've said we are 50-100 years away from that. No i'm not so sure. However, I am saying that I can replace many programmers with gpt right now, and I am. The prompting and reprompting is still both faster and cheaper than many humans.

Many people underestimate an LLM's most powerful feature when comparing it with something like Stackoverflow: the ability to ask followup questions and immediately get clarification on anything that is unclear.

I wish I had had access to LLM's when I was younger. So much time wasted on repetitive, mundane in-between code...