LLMs and Programming in the first days of 2024
antirez.com
antirez.com
> Would I have been able to do it without ChatGPT? Certainly yes, but the most interesting thing is not the fact that it would have taken me longer: the truth is that I wouldn't even have tried, because it wouldn't have been worth it.
This is the true enabling power of LLMs for code assistance -- reducing the activation energy of new tasks enough that they are tackled (and finished) when they otherwise would have been left on the pile of future projects indefinitely.
I think the internet and the open source movement had a similar effect, in that if you did not attempt a project that you had some small interest in, it would only be a matter of time before someone else did enough of a similar problem for you to reuse or repurpose their work, and this led to an explosion of (often useful, or at least usable) applications and libraries.
I agree with the author that LLMs are not by themselves very capable but provide a force multiplier for those with the basic skills and motivation.
It's not an exaggeration to say that I now do two weeks of programming a night. Of course a lot of times the result gets thrown away because the fundamental idea was flawed in a non-obvious way. But learning that is also worth while.
No longer do I need to give PR feedback more than a couple times, because we can just ask chatgpt to come up with a lint rule that detects and sometimes auto-fixes the issue. I use it to write or change Jenkins jobs, scaffold out tests, diagram ideas from a monologue brain dump, write alerting and monitoring code, write and clean up documentation.
Most recently I wanted to get some "end of year" stats for the teams, normally it would never have happened because I don't have half a day to dedicate to relearning the git commands, and how to count the lines of code and attribute changes to teams and script the whole process to work across 20 repos.
20 minutes later with chatgpt I had results I could share within the company.
It's just allowed me to skip almost all of the boring and time consuming parts of handling small things like that, and instead turns me into a code reviewer who makes a few changes to make it good enough then pushes it out
ChatGPT does diagrams for you? Writes documentation?
I found a plugin a while back called "AI Diagrams" that generates whimsical.com diagrams for me. Combined with the "speech to text" systems in chatgpt means I can just start babbling about some topic and let it write it all down, collect it into documentation, and even spit out a few diagrams from it.
I generally have to spend like 10 minutes cleaning them up and rearranging them to look a bit more sane, but it's been a godsend!
Similarly I sometimes paste a bunch of code in and tell it to write some starter docs for this code, then I can go from there and clean it up manually, or just tell it what changes need to be made. (technically I tend to use editor plugins these days not copy+paste, but the idea is the same)
Other times I'll paste in docs and have it reformat them into something better. Like I recently took our ~3 year old README in a monorepo that goes over all the build and lint commands and had it rearrange everything into sets of markdown tables which displayed the data in a much easier to understand format.
A local LLM would be preferrable all things equal, but in my experience for this kind of stuff, GPT-4 is just so much better than anything else available, let alone any local LLMs.
You can often get much of the same effect by bouncing ideas of someone, who doesn't necessarily need to know the problem space well enough to solve things but just well enough to give meaningful input. But people with the right skills aren't available at the click of a button 24/7.
In 5 years time you may well be more productive than ever. In 15 years I doubt there'll be many programming jobs in the form they are recognisable today.
There are two guys outside of my window at this very moment getting rid of a huge pile of dirt. One is in an excavator, the other in a truck. There’s a bunch of piles that they’ve taken care of today. Two centuries ago this work would’ve taken a dozen people, a few animals of burden, and a lot more time.
Where are the other ten hypothetical people? I don’t know, but chances are they’ve been absorbed by the rest of the economy doing something else worthwhile that they are getting paid for.
Zoned out somewhere in the backstreets of SF
On an individual level, however, I'd suggest keeping an eye out for opportunities to retrain.
As an analogy, I'd rather be like the coal miners in the 80s who could read the writing on the wall and quietly retrained into something else rather than those who spent their better years striking over cuts to little avail.
It's a very daunting prospect seeing a path to unemployability, though.
Depending how quickly the change happens, it could be a gentle transition, or it could upset a lot of people.
Imagine yourself as CEO. What do you think is your most likely train of thought? A/ "I can fully replace my labor force with bots" or B/ "My employees now have superpowers to do things we couldn't even conceive of two years ago". While there are certainly some scenarios where the first choice is appropriate, the latter sounds far far more likely in most scenarios to me. Why would you contract when there's suddenly so much opportunity to expand?
The other problem is, of course, is that all the historical examples (the data) are too few to generalize from while we do see how these examples are different from each other. As technological evolution progresses, automation gets more and more sophisticated, it can replace jobs that require more and more skills and talent. In other words, jobs that fewer and fewer people were able to do in the first place. This means that the bar for successfully competing in the labor market gets higher and higher and it will get to a point where a substantial number of people will just be plain uncompetitive for any job.
Or, at least that was one of the morels until LLMs were invented. (Mostly everyone thought that automation would take over the opportunities from the bottom up in general.) Now it seems that indeed white collar jobs are more in danger for now. But I digress.
The point here is that past examples are false analogies because AI (and I moslty mean future AI) is funcamentally different from past inventions. It's capabilities seem to improve quickly but we're mostly stuck with what evolution gave us. (We, as a species, are evolving but it's very slow compared to the rate of technological evolution and also we, as individuals, are stuck with whatever we were born with.)
Society had a lot of time to get used to the printing press, the advances of the Industrial Revolution and the internet. This is because the knowledge had to spread, and it's also because a ton of equipment had to be manufactured and/or shipped all over the world. We had to make printing presses, design and build factories, and get a critical mass of people internet-capable computers.
AI is fundamentally different, in that the knowledge of AI can spread instantly because of the internet and because the vast majority of the world already has access to all of the hardware they need to access the most powerful AI models.
Soon we'll have humanoid robots coming, and while they obviously have to be built, we are much more capable now than we were 50 years ago at building giant factories. We also have an efficient system of capital allocation that means that as soon as someone demonstrates a generally useful humanoid robot and only needs to scale production, they'll have access to basically infinite investor money.
We of course wouldn't have to make this explicit calculation if we could incorporate all this knowledge about fossil fuels directly into the prices of goods and services, but this is very difficult thing to do, and so far nobody has managed to do it in a global way.
It may still make material sense to use ChatGPT to create slides for middle management meetings, but that is not at all certain in a world with a significant price on emissions (though, to be fair, almost no human activity from the past fifty years stands up to this test either).
So doing stuff yourself is less carbon efficient than letting ChatGPT do its job.
As for calculating co2 pollution into the prices - we're slowly doing that, e.g. EU is setting up its carbon tax that applies to companies abroad.
The issue is that if we were to instantly include true costs of carbon removal into everything, the economy might collapse at worst, and at the least poor people might not afford food, heating nor other basic necessities any more. It will take time to do it sensibly.
What do you think programming will be like in 15 years and where is the high-value work by human programmers?
Very few jobs now where you put together basic webpages with HTML and CSS (Frontpage, then Wix/Wordpress etc replaced you). Very few jobs where you spend 100% of your time dealing with database backups and replication and managing the hardware (the cloud replaced you). Very few jobs where you spend all your time planning hardware capacity and physically inserting hard disks into racks (the cloud replaced you, too).
Am I just bad at using these tools?
If you ask it about things you don't know it was likely trained on high quality data for and get bad answers, you likely need to improve your writing/prompting.
Its best usecase is when you're not a domain expert, need quickly to run some unknown API/library inside your program inserting code like "write a function for loading X with Y in language Z" when you barely have an idea what is X,Y,Z. Its possible in theory to break-down everything to "write me a function for N" but the quality of such functions is not worth the prompting in most situations and you better ask it to explain how to write a function X,Y,Z step-by-step.
At least for now, you have to treat them like cheap metal detectors and not heat-seeking missiles.
1 - very hit or miss -- I need to fidget with the aws api in some way. I use this roughly every other month, and never remember anything about it between sessions. ChatGPT is very confused by the multiple versions of the APIs that exist, but you can normally talk it into giving you a basic working example that is then much easier to modify into exactly what I want than starting from scratch. Because of the multiple versions of the aws api, it is extremely prone to hallucinating endpoints. But if you persist, it will eventually get it right enough.
2 - I have a ton of bash automations to do various things. just like aws, I touch these infrequently enough that I can never remember the syntax. chatgpt is amazing and replaces piles of time googling and swearing.
3 - snippets of utility python to do various tasks. I could write these, but chatgpt just speeds this up.
4 - working first draft examples of various js libs, rails gems, etc.
What I've found has extremely poor coverage in chatgpt is stuff where there are basically no stackoverflow articles explaining it / github code using it. You're likely to be disappointed by the chatgpt results.
The only way to get there is to spend a ton of time playing with them, trying out new things and building an intuition for what they can do and how best to prompt them.
Here's my most recent example of how I use them for code: https://til.simonwillison.net/github-actions/daily-planner - specifically this transcript: https://gist.github.com/simonw/d189b737911317c2b9f970342e9fa...
1) Shit’s broken. Officially supported thing kinda isn’t and should be regarded as alpha-quality, bugs in libraries, server responses not conforming to spec and I can’t change it, major programming language tooling and/or whatever CI we’re using is simply bad. About the only thing here I can think of that it might help with is generating tooling config files for the bog standard simplest use case, which can sometimes be weirdly hard to track down.
2) Our codebase is bad and trying to do things the “right” way will actually break it.
What did it save me?
First of all the time to discover the do_GET and do_POST methods. I know that I should have read the docs but it's like asking a colleague "how do I do that in Python" and getting the correct answer. It happens all the time, sometimes it's me to ask, sometimes it's me to answer.
Second, the time to write the first working code. It was by no means complete but it worked and it was good enough to be the first prototype. It's easier to build on that code.
What it didn't save me? All the years spent to recognize what the code written by ChatGPT did and to learn how to go on from there. Without those years on my own I would have been lost anyway and maybe I wouldn't been able to ask it the right questions to get the code.
I don't mind taking suggestions about code from an AI because I can immediately verify the AI's suggestion by running the code, make small edits, and testing it.
* IP=PSPACE (you can verify correctness of any PSPACE computation in polynomial time)
* NIP=NEXPTIME (you can verify correctness of any NEXPTIME computation with two non-cooperative provers)
* NP=PCP(1,log(n)) (you can verify correctness of any NP statement with O(log(n)) bits of randomness by sampling just O(1) bits from a proof)
What these means is that a human is indeed able to verify correctness of the output of a machine with stronger computational abilities than the human itself.
From there, I’d ask it to do one modification at a time to the code. I’d be very precise. I’d give it only my definitions and just the function I wanted it to modify. It would screw things up whereby I’d tell it that. It would fix its errors, break working code with hallucinations, and so on. You need to be able to spot these problems to know when to stop asking it about a given function.
I was able to use ChatGPT 3.5 for most development. GPT 4 was better for work needed high creativity or lower hallucinations. I wrote whole programs with it that were immensely useful, including a HN proxy for mobile. Eventually, ChatGPT got really dumb while outputting less and less code. It even told me to hire someone several times (?!). That GPT-3-Davinci helped a lot suggests it’s their fine-tuning and system prompt causing problems (eg for safety).
The original methods I suggested should work, though. You want to use a huge, code-optimized model for creativity or hard stuff, though. Those for iteration, review, etc can be cheaper.
Whenever I tried it on something more common and/on in some stuff I had absolutely zero familiarity it did help me bootstrap quicker than reading some documentation would have
That tells a lot about how hard it is to write/find documentation that is tailored exactly to you and your needs
For me, ChatGPT is doing two things: 1) saving trivial StackOverflow and library code walking to answer specific question, and 2) helping the initial project research stage to grasp feasibility of approaches I may take before starting.
LLM's helping to just start the thing is actually a huge deal.
Being developing before when Mac had b&w screen (system 6?). Had so many good ideas just didn't work on them for whatever reason, and eventually someone did them. Maybe it is just tech people run into the same problem, or collective tech unconscious, the longer you have being around, eventually people in the industry will try to solve the problems. These occurrences goes to show me that there are really aren't many true original ideas out there, a lot of it comes down to implementation, PR, funding, geopolitics, your users and luck.
But another area I have found it extremely helpful in is exploring a new topic entirely, programming or otherwise. Telling it that I dont know what im talking about, don't necessarily need specifics, but here is what I want to talk about and want it to help me think through.
Especially if you are that person who is willing to take what you hear and do more research or ask more question. The entrance to so many fields and subjects is just understanding the basic jargon, listening for the distinctions being made and understanding why, and knowing who the authorities are on the subject.
There are an endless number projects to make a "cleaner, revised X", where the coding itself is rote and has already been done at some point, it's just shoved into slightly different semantics that will be a bit more optimal or secure or configurable. It's something that an LLM feels like it's "tip of the tongue" capable of, and in more trivial cases you really can tell GPT to "rewrite this from JS to Python" and it works. But it's limited by just interpolating what's in the training set, when what you want is "port all these standard libraries to my experimental language, and also make a build system for them".
It would be interesting to train a copilot model that is specifically intended to ask clarifying questions and be a partner in determining a solution, rather than doing its best to generate code for a vague or incorrectly specified question from a junior.
This sounds like shotgun debugging.
If not for the unwarranted confidence in incorrect responses, I could say they were at least not much worse than what I could piece together from what I knew. As it stands, they are OK filling in for a rubber duck and as autocomplete.
No matter how senior of a programmer you are, you're eventually going to encounter a technology you know very little about. You're always going to be a junior at something. Maybe you're the God of Win32, C++ and COM, but you get stuck on obscure NSIS scripts when packaging your software. Maybe you've been writing web apps for the last 25 years and sit on the PHP language committee, but then you're asked to implement some obscure ISO standard for communicating with credit card networks, and you've never communicated with credit card networks on that level before. Maybe you've been writing iOS apps since the first iPhone and Mac apps before that, spent a few years at Apple, know most iOS APIs by heart and designed quite a few yourself, but then you're asked to implement CalDAV support in your app and you don't know what CalDAV is, much less how to use it. An LLM can help you out in these situations. Maybe it won't write all the code for you, but it'll at least put you on the right track.
Or worse, you've filled your head with different tech and now you need to rehash and brush up on stuff you learned prior but swept under the rug for new stuff. It's a strange sensation. Naturally you just go with the median of whatever your company you work for is doing - then find yourself in this situation where it's "been a while" since you worked on CSS. Or it might take you a weekend of study to bring back those Python dataclass skills.
Thankfully I will most likely already be retired.
I am constantly shocked by how many people put up with such a painful workflow. OP is clearly an experienced engineer, not a novice using GPT to code above their knowledge. I assume OP usually cares about ergonomics and efficiency in their coding workflow and tools. But so many folks put up with cutting and pasting code back and forth between GPT and their local files.
This frustrating workflow was what initially led me to create aider. It lets you share your local git repo with GPT, so that new code and edits are applied directly into your files. Aider also shares related code context with GPT, so that it can write code that is integrated with your project. This lets GPT make more sophisticated contributions, not just isolated code that is easy to copy & paste.
The result is a seamless “pair programming” workflow, where you and GPT are editing the files together as you chat.
I use LLMs to chat about pros and cons of various approaches or rubber duck out problems. I need to copy code over for that, but I've not found aider good for these kinds of things, because it's all about applying changes.
I usually have several back and forths about the right way to do things and then maybe apply some change.
Sure, there's a few things you could keep in mind if you just want to chat about code (not modify it):
1. You can tell GPT that at the start of the chat. "I don't want you to change the code, just answer my questions during this conversation."
2. You can run `aider --dry-run` which will prevent any modification of your files. Even if GPT specifies edits, they will just be displayed in the chat and not applied to your files.
3. It's safe to interrupt GPT with CONTROL-C during the chat in aider. If you see GPT is going down a wrong path, or starting to specify an edit that you don't like... just stop it. The conversation history will reflect that you interrupted GPT with ^C, so it will get the implication that you stopped it.
4. You can use the `/undo` command inside the chat to revert the last changes that GPT made to your files. So if you decide it did something wrong, it's easy to undo.
5. You can work on a new git branch, allow GPT to muck with your files during the conversation and then simply discard the branch afterwards.
What I feel like I want is /chat where it still sends the context, but the prompt is maybe changed a little, to be closer to a chatgpt experience.
I haven't dug into the prompt aider is using though, so I could be wrong.
Great tool for refactoring changes though! Keep up the good work.
You think? He's the creator of Redis.
It's similar to aider (which is a great tool btw) in goals, but with a different recipe.
I hope to try aider again but it's in the unfortunate category of "I have to find a problem and a codebase simple enough that aider can handle it" whereas Copilot and ChatGPT come to me where I am. Copilot and ChatGPT help me with my actual job on my real life codebase, warts and all, every day.
Aider helps a lot when your codebase is larger than the GPT context window, but the files that need to be edited do have to fit into the window. This is a fairly common situation, where your whole git repo is quite large but most/all of the individual files are reasonably sized.
Aider summarizes the relevant context of the whole repo [0] and shares it along with the files that need to be edited.
The plan is absolutely to solve the problem you describe, and allow GPT to work with individual files which won't fit into the context window. This is less pressing with 128k context now available in GPT 4 Turbo, but there are other benefits to not "over sharing" with GPT. Selective sharing will decrease token costs and likely help GPT focus on the task at hand and not become distracted/confused by a mountain of irrelevant code. Aider already does this sort of contextually aware selective sharing with the "repo map" [0], so the needed work is to extend that concept to a sub-file granularity.
[0] https://aider.chat/docs/repomap.html#using-a-repo-map-to-pro...
Aider seems super cool, will check it out. What kind if context from the git repo does it share?
I think you've done a fantastic job of covering chat and confirmation use cases with the current features. Comments on here may not reflect the high satisfaction levels of most of your software users :)
Aider helps put into practice the use cases that antirez refers to in their article. Especially as someone get's better at "asking LLMs the right questions" as antirez refers to it.
Large changes are best performed as a sequence of thoughtful bite sized steps, where you plan out the approach and overall design. Walk GPT through changes like you might with a junior dev. Ask for a refactor to prepare, then ask for the actual change. Spend the time to ask for code quality/structure improvements.
I was pretty sure I knew which Objective-C method was broadly responsible for the bug, but I didn't know what that method did, and the decompiled version was a nonsensical mess. I felt like I'd hit a wall.
Then I thought to feed the decompiler babble to GPT-4 and ask for a clean version. The result wasn't perfect, but I was able to clean it up. I swizzled the result into the app, and I'm pretty sure the bug is gone. (I never found reproduction steps, but the problem would usually have occurred by now.)
I never could have done this without GPT-4.
Admittedly a complete rewrite of a piece of code, even without understanding what you are doing (e.g by using an LLM), is unlikely to have the same bugs as the original implementation (but may have different bugs), but hopefully no-one is doing this for code where bugs have any significant consequence (e.g. system downtime, cost to customers).
When I cleaned it up, I took out some complexity which I believe was responsible for the bug, at the cost of some performance. According to GPT-4, the original version was checking file descriptors to decide when to do work. My version just does the work every 5ms.
There was a tradeoff between performance and complexity. The high-performance, high-complexity version was buggy, so I switched to a simpler option at the cost of some performance.
This isn't where the LLM was significant. The LLM was able to make sense of unreadable decompiled code, similar to how the author had ChatGPT translate from compiled assembly code back to C. (Giving GPT-4 the actual assembly never occurred to me, in hindsight I should have tried that first.)
For example, at my current company the developer who introduced a “clever” navigation system didn’t know how HTML forms should be used, and why servers still allowed what they did. It worked. Now, 20+ years later that sole developer’s stupid decision, and lack of HTML best practices will cost my company a few million dollars (and by the way already cost probably a few million). A missing day of learning (and by the way a clear sign, that that developer should’ve never trusted with this task).
Senior developers learn this, and I’ve never seen that better developers would be satisfied and would say “yeah, I fixed it”, when they don’t understand the what and how completely, even when it’s not strictly necessary. They burned themselves enough times.
I think it's an error to bring that up here though, where we're talking about someone patching a closed source app for their personal use. Is it worth the cost/benefit of decompiling and studying the app's code sufficiently long to be highly confident of the fix?
Sloppiness and "good enough" has its place. So does full effort correctness.
The main culprit is that the backend is Java EE, and they used a single form for multiple things. This is a 20+ years old software, so they had to used request parameters and attributes everywhere. It’s impossible to locate where a given parameter is used or where attributes are created and used exactly. That’s the first problem. Second, is this single form per page thing. They use that HTML form for everything. Even page navigation, they just discard unnecessary fields on server side. This means that they change that form from JavaScript all the time. Sometime the generated action is not used at all, it’s overwritten with every event. And a single input can be used for multiple things. Third, they used a very flexible framework, which was outdated 10 years ago, so it definitely needs to be replaced.
Add these together, you have terrible spaghetti code for 350+ pages with 1000+ endpoints. There is no separation of code even on HTML level. I know where this “use only one form” came from. I found the ancient doc of that framework, where it mentioned. The problem is that they meant a form object on server side per page (and not per endpoints), and not on client side. And they fucked up completely, because they started to use multiple form objects per page, and single client side HTML forms.
So now, replacing that old framework starts with a hefty refactoring, creating individual HTML forms for example. It will take at least half a year for a team, just the refactoring, because it’s very difficult to split those forms, since request attributes and parameters used everywhere.
Just as any introductory macroeconomics class teaches, if one island has superior skill in producing widgets A, it doesn't matter how terrible the other island's skill at producing B is, we'll still see specialization where island A leverages island B. So of course antirez's relative ability in systems programming would relegate the LLM to other progamming tasks.
However! We do not exist in isolation. There is a multitude of human beings around us, hungry for technical challenges and food. Many of them have or could obtain skills complimentary to our own. In working together, our cooperative efforts could be more than the sum of their parts.
Perhaps the LLM is better at writing PyTorch code than antirez. Just because we have an old bent screwdriver in the garage doesn't mean we should try to use it. Perhaps we'd be better off heading to the hardware store today.
Furthermore, i'd vastly prefer a workflow most of the time where i don't even have to ask. Or so i imagine. Ie i think i'd prefer a Clippy style "Want me to write a test for this? Want me to write some documentation for this?" etc helpers. I don't want to have to ask, i want it to intuit my needs - just like any programmer could if pair programming with you.
And most of all i want it to have access to all files. To know everything about the code possible. Don't just look at the func name in isolation, attempt to understand how it's used in the project.
If i have to baby sit an LLM for a simple function refactor to give it all files where the function is used or w/e, i'd rather do it myself with tools like AST Grep or even my LSP in many cases.
I'm very interested in LLMs for simple tasks today, but the tooling feels like my primary blocker. Also possibly context length, but i think there's lots of ways around that.
I had great impression of sourcegraph’s cody. https://sourcegraph.com/cody few months ago.At least with the enterprise version of the sourcegraph that had indexed most of the orgs private repos.
The web ui (vscode extension was somehow worse, not sure why) was providing damn good responses, be it code generating or QA/explanation about code spanning through multiple repos. (e.g. terraform modules living in different repos)
Afaiu it was using the sourcegraph index under the hood. But I never really deep dived into the cody’s design internals (not even sure if they are actually public)
That being said, I’ve departed from the org months ago and haven’t used cody since then, so take this with a grain of salt, since the whole comment could be outdated a lot.
He sees a new tool that others have found interesting, and he identifies ways to use that tool that are useful for him, while also acknowledging where it's not useful. He backs it up with plenty of examples of where he found it not-useless. This is not a revolutionary insight, especially for a developer. We constantly use a variety of tools, such as programming languages, that have strengths and weaknesses. Why are LLMs so different? It seems foolish to claim they have zero strengths.
Of course they do have uses, but more related to discovery of APIs and documentation than actually writing code (esp. where bugs matter) for the most part.
I also have to wonder how long until open source code (e.g. GPL'd) regurgitated by LLMs and incorporated into corporate code bases becomes an issue. The C suite dudes seem concerned about employees using LLMs that may be publicly exposing their own company's code base, but illogically unworried about the reverse happening - maybe just due to not fully understanding the tech.
This is, indeed, the core of our disagreement. You seem confident it would be a bother to others to ask for help. I'm confident that there are many who would value the opportunity to collaborate with you. I feel sure that whatever analysis you're doing could benefit from the sounding board of a domain expert, and you'd both benefit from the exchange.
EDIT: to clarify, being "famous" has nothing to do with it. Each of us has worth and we all would gain by working with others.
To me its like saying you shouldn't play solitaire because you are a world class poker player, and there's plenty of people who want to game with you. They are orthogonal concepts - just communicating with people can be more work than just reading on your own.
In the long term, it is beneficial to have experts as your collaborators. From my experience though, true collaboration is unlocked once you have established a personal relationship with someone, which takes time and repeated effort. Until then, the collaboration is no better than searching the internet or asking chatGPT.
Establishing relationships with people is hard and takes a lot of work, and frequently doesn't work out like you hope. ChatGPT is a close enough approximation for smaller tasks like the OP describes
Homo sapiens's superpower is social cooperation. My concern is that these systems will abet the existing social forces which seem to be causing unprecendented levels of isolation of adults, which will continue to drive smart people away from collaboration and towards solitude, at a level far beyond simple prefences would suggest.
We already have enough trouble hearing each other through the noise, and understanding what each other has to say. I don't have the answers but I'm looking for them and I do hope other humans will, too.
- survival
- making money
- sex
- enjoyment via social interactions (like parties, hangouts, etc)
It just so happens that for the majority of our civilization, to get those things, we've had to cooperate, but as we develop technology, our ability to get those things increases and our reliance on others decrease (though in a weird way it increases since technology is complexity so society becomes larger, more complex, and more inter-dependent)
We are still in an unprecedented technological boom of computing so we are adjusting on the fly to it. Like the OP says, AI can greatly accelerate independent learning, but eventually that learning plateaus. Once it does, we have to go back to collaboration, but until we find that limit, I think it's human nature to push on.
I think there is an incorrect worldview that tries to blame human problems on technology.
It's quite true that isolation is an increasing problem. But the idea that instead of using an AI that can spit out a comprehensive answer in seconds, we should all pretend that such tools don't exist, and start constantly asking for help with every idea or request instead, while waiting 3-100 times longer for a less thorough response, is ludicrous.
It's a great idea to collaborate more and try to avoid isolation. But those are societal problems. They are not caused by the latest tools.
Also, as far as humanity's "super-power" as being collaboration, this is quite a shortsighted comment. I believe that well before AI achieves "super" level IQ, it will vastly outperform humans due to other advantages. One of those advantages is speed. Another is the ability to communicate and collaborate much, much more rapidly and effectively than humans.
One type of digital life that may take over control of the planet (possibly within decades rather than centuries) would be a type of swarm intelligence with the ability to actually "rsync" mental models to directly transfer knowledge.
I'm a mediocre programmer who uses GPT for a ton.
Are you volunteering to answer my questions on all the obscure stuff I ask it? Because I don't really know anybody else who will.
Anyway, my email is in my profile, write to me if this is something you're up for!
Edit: Here's a list of the stuff I asked it over the weekend:
- Discuss the pros and cons of using Standardized-audio-context instead of just relying on browser defaults. Consider build size and other issues.
- How to get github actions to cache node_modules (not the npm cache built-in to the actions/setup-node action.
- Howto get the current git hash in a GitHub action?
- Rewrite a react class-based component in functional style
- How to test that certain JSX elements were generated without directly-comparing React elements for my ANSI color to HTML parser?
- Does it make more sense to keep a copy of the original text in an editor, or just hold on to something like a crc32 to mark a document dirty?
- Can you set focus to a window you create with window.open? (You sure can!)
- Rewrite the rollup.config.js for a library of mine to produce a separate rollup config per audio worklet bundle
- Turn this tutorial for backing up a Mastodon instance into a script
- Refactor a standalone class to split it in half so each class manages precisely one thing.
- I have some code I'm writing for turn-by-turn directions. I have the data structures already, let's write code to narrate them.
- What's with this weird type error around custom CSS properties with React?
I wish we had an easier time talking about ideas with a little more detachment.
Did you read the article? Throughout the entire post he clearly says LLMs have a lot of value in his workflow.
There can also be significant overhead in looking elsewhere for a solution. That's a big part of why so many developers reinvent things. This is often dismissed as NIH syndrome, but there's more to it than that.
You raised "introductory macroeconomics". The economic effect that will most strongly apply in the case of LLMs is that of the technology treadmill (Cochrane 1958): when there's a tool that can improve productivity, it will be used competitively so that those who don't use it effectively, to improve their productivity, will be outcompeted.
This seems like an unavoidable result in typical capitalist economies.
Your point about leveraging hungry humans would require strong incentives to overcome the treadmill effect. Most Western countries don't have many ways to implement anything like that. The closest thing might be unions, but of course most software development is not unionized.
At the beginning, when there's 0% of the task done, and you need to start _somewhere_, with a hello world or a CMakeLists file or a Python script or whatever, it takes effort. Before ChatGPT/LLM, I had to pull that effort out from within myself, with my fingertips. Now, I can farm it out to ChatGPT.
It's less efficient, not as powerful as if I truly "sat down and did it myself," but it removes the cost of "deciding to sit down and do it myself." And even then, I'm cribbing and mashing together copy-pasted fragments from GitHub code search, Stackoverflow, random blog posts, reading docs, Discord, etc. After several attempts and retries, I have a "5% beginning" of a project when it finally takes form and I can truly work on it.
I sort of transition from copy-pasting ChatGPT crap to quickly create a bunch of shallow, bullshit proofs-of-concept, eventually gathering enough momentum to dive into it myself.
So, yes, it's slower, and more inefficient, and ChatGPT can't do it better than I can. But it's easier and I don't have to dig as deep. The end result is I have much more endurance in the actual important parts of the project (the middle and end), versus burning myself out on the beginning.
Was I asking the right questions from the beginning, and if not, can I effectively salvage my work?
Sunk costs disappear into a $20 subscription
This is the key insight from using LLMs in my opinion. One thing that makes programming especially well suited for LLMs is that it's often trivial to verify the correctness.
I've been toying around this concept for evaluating whether a LLM is the right tool for the job. Graph out "how important is it that the output is correct" vs "how easy is it to verify the output is correct". Using ChatGPT to make a list of songs featuring female artists who have won an Emmy is time consuming to verify it's correct, but it's also not very important and it's okay if it contains some errors.
Problems where coming up with solution is hard but verifying a possible solution is easy.
And we all know what that class of problems is called.
is this why software never has any bugs?
This is not a comment on your specific example, but on the idea as a whole.
My company has approved copilot but Copilot autocomplete has been an awful experience. company hasn't approved copilot chat ( which is what i need) .
But I would love something similar that can run on my laptop for my code to generate unit tests, code comments ect ( ofcourse with my input and guidance).
I had the same experience, I feel like I must be crazy because so many of my colleagues have been singing its praises. I found it immensely distracting and disabled it again after a couple of days.
It was like having someone trying to finish my sentence while I was still speaking; even when they were right, it was still annoying and knocked me out of my flow (and very often, it wasn’t right).
If you mean the model, I’ve been happy with Deepseek Coder. Mistral is a popular alternative as well.
There's another one called Wingman Copilot as well: https://marketplace.visualstudio.com/items?itemName=WingmanC...
A sibling comment suggested Wingman but I haven't tried it.
> And then, do LLMs have some reasoning abilities, or is it all a bluff? Perhaps at times, they seem to reason only because, as semioticians would say, the "signifier" gives the impression of a meaning that actually does not exist. Those who have worked enough with LLMs, while accepting their limits, know for sure that it cannot be so: their ability to blend what they have seen before goes well beyond randomly regurgitating words. As much as their training was mostly carried out during pre-training, in predicting the next token, this goal forces the model to create some form of abstract model. This model is weak, patchy, and imperfect, but it must exist if we observe what we observe. If our mathematical certainties are doubtful and the greatest experts are often on opposing positions, believing what one sees with their own eyes seems a wise approach.
I'd like to see this evidence, and by that I don't mean someone just writing a blog post or tweeting "hey I asked an LLM to do this, and wow". Is there a numerical measurement, like training loss or perplexity, that quantifies "outside the training set"? Otherwise, I find it difficult to take statements like the above seriously.
LLMs can do some interesting things with text, no doubt. But these models are trained on terabytes of data. Can you really guarantee "there is no part of my query that is in the training set, not even reworded"? Perhaps we can grep through the training set every time one of these claims are made.
The perfect example of that is the tikz unicorn in the Sparks paper. Seemed like a unique task, until someone found a tikz unicorn in an obscure website.
There is plenty of evidence that LLMs struggle as you move out of distribution. Which makes perfect sense as long as you stop trying to attribute what they’re doing to magic.
This doesn’t mean they’re not useful, of course. But it means that we should should be skeptical about wild capability claims until we have better evidence than a tweet, as you put it.
This was the package: https://ctan.org/pkg/tikzlings?lang=en
I mean..yes?
Multi digit arithmetic, translation, summarization. There are many tasks where this is trivial.
Seriously, just don't use Google for search. Google search is just a way to get you to look at their ads.
Use a search engine that is aligned with your best interests, suppresses spammy sites, and lets you customise what you want it to surface.
I've used chatgpt as a coding assistant, with varying results. But my experience is that better search is orders of magnitude more useful.
Can you give an example of such a search engine? Which one(s) do you use and why?
I've tried a couple of searches on the free tier and they gave pretty much the same results. I only have so many free searches to check too.
https://blog.kagi.com/kagi-features
If you prefer LLMs to Googling, then at least consider "phind":
This is definitely a departure because when I subscribed to Kagi a couple months ago, all of my Google results for similar searches were SEO spam blogs filled with Amazon affiliate links that look like they had just sucked some Amazon reviews automatically into some poor facade to generate affiliate revenue.
These results were a surprise to me. Not sure what changed.
I imagine what changed is that Kagi started getting traction on site like here and some managers at google actually did something about it.
My own test "voynich illuminated manuscript" which used to give nothing but pintrest spam on google. Now there is just one result from pintrest in google and pretty much every result in Kagi is from pintrest.
There is an academic tab which seems interesting. I will give it a try later.
(No connection with kagi.com except being a very satisfied user)
Pay for it, so search results are the product, instead of an ad platform sold to advertisers with you as the product.
https://search.brave.com/help/goggles
There is a list (search) of public goggles: https://search.brave.com/goggles
The goggles itself are just text files with basic syntax and can be hosted on e.g. github gist. (though you have to publish it to brave)
https://github.com/brave/goggles-quickstart/blob/main/goggle...
Tbh, I can’t really compare brave search to kagi, since I never used kagi (though I’m using Orion - webkit based browser from the same dev and love it). Afaik, brave search is using its own index, thus making the results somehow limited and inferior to kagis. Just wanted to throw some (free) alternative here that works for me. :)
* Note that Brave search, despite privacy oriented, is still ad funded and there was few controversies about brave’s (browser) privacy in the past. (if that’s relevant for you)
* I’m not affiliated with Brave in any way.
I do quickly run into bumps, where search is necessary (a lot of times it's some variant of a breaking change in a dependent library). Once I find a good enough issue description, I just slap that back into chatgpt. It handles it very well and sticks for the rest of the conversation. Somehow chatgpt is aware that the context information takes precedence over trained data.
I also have the Kagi subscription, which I'm using for above. I'm very happy with both tools working in tandem and am genuinely happy about that kind of time spending
I have never loved learning the details of an obscure communication protocol or the convoluted methods of a library written by someone who wants to show how good they are. It seems like "junk knowledge" to me. LLMs save me from all this more and more every day.
This is depressing or tongue-in-cheek considering who he is -- Redis creator -- and has an older post titled 'In defense of linked lists', so talking about linked lists in Rust is not "junk knowledge" or something an LLM can analyze circles around any human.
It's the best coding nihilism as a profession post I have read though.
Today we have a great example in Kubernetes, and all the other synthetic complexity out there. I'm in, instead, to learn important ML concepts, new data structures, new abstractions. Not the result of some poor design activity. LLMs allow you to offload this memorization out of your mind, to make space for distilled ideas.
I, too, find LLMs a balm for this pain. They have kind-of-basic level of knowledge, but about everything.
In short, it allows for a more efficient expenditure of mental and emotional energy!
Much of programming, coding and developing is done by a person who is a knowledge worker and writes code. A good proportion of code to be written, will be written just once and never again. The one-off code snippet will stay in a file collecting dust forever. There is no point in trying to remember it in the first place, because without constant repetition of using it, it will be forgotten.
LLMs can help us focus our knowledge where it really matters, and discard a lot of the ephemeral stuff. That means that we can be more of knowledge workers and less of coders. I will push it even further and state that we will become more of knowledge workers and less of coders until we will be, eventually and gradually, just knowledge workers. We will need to know about algorithms, algorithmic complexity, abstractions and stuff like that.
We will need to know subjects like that Rust book [1] writes about.
Yes, but I have to double-check every answer. And that, for me, greatly mitigates or entirely negates their utility. Of what value is a pocket calculator that only gets the right answer 75% if the time, and you don't ex ante know what 75%?
But, I have to disagree on this point, since many programs written in ie. C have security issues that takes a long time to discover.
I would agree that it is a primary goal of software engineering to move as much as possible into the category of automatic verification, but we're a long, long way from 99%.
I think that antirez is technically correct in that there is a vast amount of code that will not compile compared to the amount of code that will compile. So saying '99%' sort of makes sense.
But that doesn't capture the fact that of the code that compiles there is a vast amount of code that doesn't do what we want to happen at runtime compared to the code that does do what we want to happen.
And after that there is a vast amount of code that doesn't do what we want to happen 100% of the time at runtime compared to the code that only most of the times does what we want to happen at runtime.
The interesting thought experiment that came to me when thinking about this was that I would be more likely to trust LLM code in C# or Rust than I would be to trust LLM code in assembly or Ruby.
Which makes me wonder ... can LLMs write working Idris or ATS code?
I've seen some refer to non existant APIs while discussing migration to a new library major version. "Sure that's easy, we should just replace this function with this new one".
Imagine all those more subtle bugs that are harder to spot.
Or, from a different angle - all models are wrong, some are useful.
As it happens, LLMs are useful even if they're sometimes wrong.
Perhaps so. I guess it depends on how long it takes to code up property-based tests.
- I can also tell the llm to write tests for the code it wrote and i can validate that the tests are valid.
- LLMs are also valuable in introducing me to concepts and techniques I would never had had exposure to. For example, I have a problem and explain my problem, it will bring up technologies or terms I never considered because I just didn't know about them. I can then do research into those technologies to decide if they are actually the right approach.
If I don't trust the generated code, why should I trust the generated code that tests the generated code?
I appreciate the author writing this article. Whenever I read about future of field, I get anxiety and confusion but then again I think other options too which were available to me was less interest of me.
I am now at the place that I still have the opportunity to pivot and focus on pure/applied mathematics than being in software field.
Honestly I wanted to make money through this career but I don't know what carrer to choose now.
I keep working on myself and don't compare myself to others but if argument is top 1% programmers will be required in the future then I doubt myself because I have still learn lot of things and then how about competing with both experienced & knowledgeable.
I was thinking about pin-pointing a target then becoming expert at it (by 10000 hrs rule)
I'm sorry to ask but today or in-general I am very confused which path/carrer to Target related to computing, Mathematics. Please suggest and give me your valuable advice. Thank you
Write code for the wise cracking door AI instead.
If you're looking at a 5-10 year timeline then even pure or applied mathematics may well heavily use AI models.
We're always going to need architects that build the scaffolding together with LLMs. Programmers + LLMs will be able to outcompete programmers without. If one programmer can do more it just means projects will become more ambitious not less programming needed.
I've never worked for a company that had too little work for their software engineers. Rather many projects are on long timelines because there are only so many hours available per month.
Another analogy: with a high level programming language you can do what previously needed 10x the lines of code in assembly. I don't think they caused job losses for software engineers.
Putting that aside, based on your question and willingness to put it out there… I would say this: just surrender to what charms you right now. Do you feel drawn toward programming? Follow that. Or math? Follow that. They may not be mutually exclusive.
As you go, stay tuned in to how you feel about the activity in the moment. Not your anxiety about what you think about the future prospects, but just how it feels right now to be doing the thing. That feeling may change over time, and it will guide you if you stay tuned in.
Some one, a person with a sense of responsibility, has to sign off on changes to the code. LLMs have shown to come with answers that make no sense or contain bugs. A person (for now) needs to decide is the LLM's suggestion is acceptable, if we need more tests, if we want to maintain it.
I think programmers will be needed for that, they will just be made more productive (as what happened with the introduction of garbage collection, strong typed languages, powerful IDEs, StackExchange, ...)
It appears no better than the mixtral example that it's supposedly an improvement on.
Yours is the first blog which matches my experiences with the code side of things, but I've found them even more useful in the learning side of things.
Hmm, this suggests to me that in a better world, the systems problems would have been solved with code, and the sorts of one-off problems which current LLMs do handle well would have been solved with formulae in a (shell-like? not necessarily turing-complete?) DSL.
or, the other stuff GPT was producing is just as bad, but that he's not experienced enough in the domain to see it, where as the stuff he is experienced with looks immediately sus or subpar.
At the same time, GPT apparently doubles programming productivity. (Though obviously this depends on the task.)
I've long wished to have the best of both worlds. It seems I may soon get my wish: local LLMs will probably catch up with GPT-4 this year, or even outpace it!
That's a tradeoff I already make when using relativly new third party libraries/services to accelerate experimentation.
In one off tasks where someone is not enough of an expert to know its flaws, and such expertise is not required, "the marvel is not that the bear dances well, but that the bear dances at all".
Does HN have any favorite local LLMs for coding-related tasks?
But that's basically what engineering (, and medicine, and law, and all sorts of professions out there,) has always been about. Engineers build railways and bridges based on the same proven principles in slightly different forms, adapting to the specific needs of each project. Their job is not to come up with groundbreaking inventions every day.
> And now Google is unusable: using LLMs even just as a compressed form of documentation is a good idea.
Beyond all the hype, it'd undeniable that LLMs are good at matching your query about a programming problem to an answer without inundating you with ads and blog spam. LLMs are, at the very least, just better at answering your questions than putting your question into to google and searching Stack Overflow.
About two years ago I got so sick of how awful Google was for any serious technical questions that I started building up a collection of reference books again just because it was quickly becoming the only way to get answers about many topics I cared about. I still find these are helpful since even GPT-4 struggles with more nuanced topics, but at least I have a fantastic solution for all those mundane problems that come up.
Thinking about it, it's not surprising that Google completely dropped the ball on AI since their business model has become bad search (i.e. they derive all their profit from adding things you don't want to your search experience). At their most basic, LLMs are just really powerful search engines, it would take some cleverness to make them bad in the way Google benefits from.
How much has changed?
It's just a scam to keep you scared and stop you from empathizing with your fellow workers.
I am quite unconvinced this is the reason. Seems rather conspiratorial.
This is what happened to America's manufacturing industry. Shouldn't emphasizing with fellow workers mean recognizing the pattern instead of dismissing it as FUD?
What has happened is GPT-4 came out (which is certainly better in some domains but not everywhere), but mainly the models have become much cheaper and slightly easier to run, and people are pairing LLMs with other things rather than using them as a single solution for all possible tasks — which they probably could do in principle if scaled up sufficiently, but there may well not be enough training data and there certainly aren't computers with enough RAM.
And, like with the self-driving cars, we've learned a lot of surprising failure modes.
(As I'm currently job-hunting, I hope what I wrote here is true and not just… is "copium" the appropriate neologism?)
As for demand, that's difficult to predict. I'd argue a lot of software being written today doesn't really need to be written. Lots of weird ideas were being tried because the money was there, pursuing ever new hypes, with an entire sub industry building ever more specialised tools fueling all this. And with all that growth, ever more programmers have been thrown at dysfunctional organisations to get a little more work done. My gut tells me that we'll see less of that in the next years, but I feel even less competent to predict where the market will go than where the tech will go.
So long story short, I guess we'll still need programmers until there's a major leap towards GAI, but less than today.
This is the holy grail of low-code products.
See all the SaaS products, without any access to their implementation, programable via graphical tooling, or orchestrated via Web API integration tools, e.g. Boomi.
Where does the trust in a binary spit out by an LLM come from? The binary is likely unique and therefore your trust can't be based on other users' experience, there likely isn't any financial incentive or risk on the part of the LLM should the binary have bugs or vulnerabilities, and you can't audit it if you wanted to.
QA, acceptance testing whatever, no different from buying closed source software.
Only those that never observed the replacement of factory workers by complete robot based chains can think this will never happen to them.
Here is a taste of the future,
https://www.microsoft.com/en-us/power-platform/products/powe...
An assembly line robot is programmed with a very specific repeatable task that can easily be quality tested to ensure that there aren't manufacturing defects. An LLM generating binaries is doing this one off, meaning it isn't repeatable, and the logic of the binary isn't human auditable meaning we have to trust that it does what was asked of it and nothing more.
There are ACM papers about it.
It didn't hold on.
Do you really inspect the machine code generated by your AOT or JIT compilers, in every single execution of the compiler?
Do you manually inspect every single binary installed into the computer?
Would you trust a compiler's byte code if it spit out slightly different instructions every time you gave it the same input? Would you feel confident in the reliability and performance of the output? How can you meaningfully debug or performance profile your program when you don't know what the LLM did and can't reproduce the issue locally short of running the exact copy of the deployed binary?
Comparing compilers and LLMs really is apples and oranges. That doesn't mean LLMs aren't sometimes helpful or that they should never be used in any situation, but LLM fundamentally are a bad fit for the requirements of a compiler.
Instead of warm bodies somewhere on the other side of the planet, it is a LLM.
> I regret to say it, but it's true: most of today's programming consists of regurgitating the same things in slightly different forms. High levels of reasoning are not required. LLMs are quite good at doing this, although they remain strongly limited by the maximum size of their context. This should really make programmers think. Is it worth writing programs of this kind? Sure, you get paid, and quite handsomely, but if an LLM can do part of it, maybe it's not the best place to be in five or ten years.
I wonder how different this would be if software was not hindered by "intellectual property" laws.
I've found autocomplete via these systems to be improving rapidly. For some work, it's already a big boost, and it's close to a difference in kind from the original IntelliSense. Amusingly though, I primarily write in an editor without any autocomplete, so I don't experience this often. But I do, precisely for the throwaway code and lower-value changes.
Finally, it's not clear to me that the distinction is between systems programming and scripting. My sense is that Chat GPT and similar are (a) heavily influenced by the large corpus of Python, so it's better at it than C and (b) the examples here involved more clever bit manipulation than most software engineers ever interact with.
Perlis once quoth:
> 18. A program without a loop and a structured variable isn't worth writing.
After 5 minutes of thought, I'd update that, for my hacking, to:
"A program without some convergence reasoning and a non-injective change of representation isn't worth writing."
(iow, I'd be happy to let LLMs, or at least other people, wrangle glue and parsley code, according to the taxonomy of: https://news.ycombinator.com/item?id=32498382 )
I do suspect though that both the hashing and 6-bit weight examples are just extremely rare in the corpus. It wasn't confused about loops, or hashing generally, but just didn't do as well as antirez would have liked. The description of the 6-bit to "why don't I just cast this to 8-bits" thing is definitely a problem. And worse, it's a problem a more junior engineer might not understand. But I suspect that a model trained on a corpus with lots more bit manipulation would have been fine, as it wasn't complex.
Clearly we just need a fine tuned one :).
Haven’t had much luck with code completion thus far.
LLMs are like stupid savants who know a lot of things.
LLM's today for me are the equivalent of a large scale human memory for code or for faster augmented retrieval - do they hallucinate details, quite often, but do I find it utilitarian versus dragging myself over documentation details - more often than not.
For me they help when time is short and when I want to maximize creative exploration.
> this goal forces the model to create some form of abstract model. This model is weak, patchy, and imperfect, but it must exist if we observe what we observe.
Is a completely fallacious line of reasoning, and I'm surprised that he draws this conclusion. The whole reason the "problem of other minds" is still a problem in philosophy is precisely because we cannot be certain that some "abstract model" exists in someone's head (man or machine, do you argue it does? show it to me) simply because an output meeting certain constraints exists. This is exactly the problem of education. A student that studies to answer questions correctly on a test may not have an abstract model of the subject area at all. Even they may not be conducting what we call reasoning. If a student aces a test, can you confidently say they actually understand a domain? Or did they simply ace a test?
Furthermore, LLM's lack of consistency and inability to answer basic mathematical questions, and their limitation to purely text based areas of concern and representation are all much stronger arguments for siding with the notion that they really are just sophisticated, stochastic, machines, incapable of what we'd normally call reason in a human context. If LLM's "reason" it is a much different form of reasoning than that which human beings are capable of, and I'm highly skeptical that any such network will achieve parity to human reason until it can "grow up" and learn embodied in a rich, multi sensory environment, just like human beings. For machines to achieve reason, they will need to break out of the text-only/digital-only box first.
Exactly! This is why I removed this fundamental questions from my post: in this moment they don't have any clear reply and will basically make an already complex landscape even more complex. I believe that right now, whatever is happening inside LLMs, we need to focus on investigating the practical level of their "reasoning" abilities. They are very different objects than human brains, but they can do certain limited tasks that before LLMs we thought to be completely in the domain of humans.
We know that LLMs are just very complex functions interpolating their inputs, but this functions are so convoluted, that in practical ways they can solve problems that were, before LLMs, completely outside the reach of automatic systems. Whatever is happening inside those systems is not really important for the way they can or can't reshape our society.
Is it just me or did anyone smile at this sentence? The first paragraph sounds like the academic way of saying "we invented huge neural networks but we couldn't understand it".
Reader view in Safari preserves the monospace font... /facepalm
In order to protect your delicate sensibilities
I would further suggest to avoid consulting most
research output from before the mid 1980s.