Sure, Microsoft is happy to let their AIs scan everyone else’s code., but is anyone aware of any software houses letting AIs scan their in-house code?
Any lawyer worth their salt won’t let AIs anywhere near their company’s proprietary code intil they are positive that AI isn’t going to be blabbing the code out to every one of their competitors.
But of course, IANAL.
The LLMs they train on their code will only be accessible internally. They won’t leak their own intellectual property.
Will that not be more experiensive than having developers?
Possibly. It’s hard to know without seeing the numbers and assessing output quality and volume.
Also it’s not unheard of that some bigwig wastes millions of company €€ for some project they fancy. (Billions if they happen to be Elon)
Of course not. It will be more expensive and they’ll still have to pay developers to figure out what’s wrong with their AI code.
Depends on the use case. Training local llms is a lot cheaper after Galore and there are ways to get useful local models with only a moderate amount of effort, see e.g. augmentoolkit.
This may or may not be practical in many use cases.
24 months is pretty generous but no doubt there will be significantly less demand for junior developers in the near future.
Yeah which is why this is a dumb statement from Amazon. But then again I don’t expect C-suite managers to really understand the intricacies of their own companies.
If only we had an overarching structure that everyone in society has agreed exists for the purposes of enforcing laws and regulating things. Something that governs people living in a region… Maybe then they could be compelled to show exactly what they’re using, and what those models are being trained with.
Oh well.
Seriously how can these CEOs of a GPU company not talk to a developer. You have loads of them to interview
They obviously have to sell something. I doubt that they honestly think that this will happen.
I guess the programmers should start learning how to mine coal…
We will all be given old school Casio calculators a d sent to crunch numbers in the bitcoin mines.
A company I used to work for outsourced most of their coding to a company in India. I say most because when the code came back the internal teams anways had to put a bunch of work in to fix it and integrate it with existing systems. I imagine that, if anything, LLMs will just take the place of that overseas coding farm. The code they spit out will still need to be fixed and modified so it works with your existing systems and that work is going to require programmers.
So instead of spending 1 day writing good code, we’ll be spending a week debugging shitty code. Great.
Translation: “We’re going to make the suite for building, testing, and deploying so obnoxiously difficult to integrate with your work environment that in two years nobody in your DevOps team will be able to get anything to a release state.”
Me, fiddling with a config file for a legacy Perl script that’s been holding up the ass-end of the business since 1996: “Uh, yeah that’s great.”
!remindme in 24 months
I will put down a solid grand that this exact same article will be printed by the exact same website 24 months from now and it will receive the exact same reception.
Nah, if it doesn’t pan out then all these folks will pretend they never said this, but in 24 months programming will be obsoleted by <insert fresh buzz here>
If that’s true, how come there isn’t a single serious project written exclusively or mostly by an LLM? There isn’t a single library or remotely original application made with Claude or Gemini. Not one.
Lets wait for any LLM do a single sucessful MR on Github first before starting a project on its own. Not aware of any.
My last employer had many internal tools that were fine.
They had only a moderate amount of oversight.
I had to find a new job, I’m actually thinking of walking away from software development now that there are so few jobs :(
It sucks but there’s no sense pretending this won’t have a large impact on the job landscape.
What did these tools do? I don’t see any LLm being used for creating anything working from scratch, without the human propmter doing most of the heavy lifting.
Mostly internal data cleaning stuff, close etc, which I accept is less in scope than you’re original comment.
The things you are describing sound like if-statement levels of automation, GitHub Actions with preprogrammed responses rather than LLM whatever.
If you’re worrying about being replaced by that… Go find the code, read it, and feel better.
The code was non trivial and relatively sophisticated. It performed statistical analysis on ingested data and the approach taken was statistically sound.
I was replaced by that. So was my colleague.
The job market is exceptionally tough right now and a large part of that is certainly llms.
I think taking people with statistical training out of the equation is quite dangerous, but it’s happening. In my area, everybody doing applied mathematics, statistics or analysis has been laid off.
In saying that, the produced program was quite good.
Certainly sounds more interesting than my original read of it! Sorry about that, I was grumpy.
All good man.
I think the point is that LLMs can replace people and they are quite good.
But they absolutely shouldn’t replace people, yet, or possibly ever.
But that’s what’s happening and it’s a massive problem because it’s leading to mediocre code in important spaces.
there isn’t a single serious project written exclusively or mostly by an LLM? There isn’t a single library or remotely original application
IMHO “original” here is the key. Finding yet another clone of a Web framework ported from one language to another in order to push online a basic CMS slightly faster, I can imagine this. In fact I even bet that LLM, because they manipulate words in languages and that code can be safely (even thought not cheaply) tested within containers, could be an interesting solution for that.
… but that is NOT really creating value for anyone, unless that person is technically very savvy and thus able to leverage why a framework in a language over another creates new opportunities (say safety, performances, etc). So… for somebody who is not that savvy, “just” relying on the numerous existing already existing open-source providing exactly the value they expect, there is no incentive to re-invent.
For anything that is genuinely original, i.e something that is not a port to another architecture, a translation to another language, a slight optimization, but rather something that need just a bit of reasoning and evaluating against the value created, I’m very skeptical, even less so while pouring less resources EVEN with a radical drop in costs.
Yeah, that’s not going to happen.
Yeah writing the code isn’t really the hard part. It’s knowing what code to write and how to structure it to work with your existing code or potential future code. Knowing where things might break so you can add the correct tests or alerts. Giving time estimates on how long it will take to build the parts of the system and building in phases to meet your teams needs.
I’ve always thought that design and maintenance are the difficult and gruelling parts, and writing code is when you get to relax for a bit. Most of the time you’re in maintenance mode, and it’s harder than writing new code.
This. I’m learning a new skill right now & hardly any of it is actual writing— it’s how to arrange the pieces someone else wrote (& which sometimes AI can decently reproduce.)
When you use a computer you don’t start by mining iron, because the thing is already built
Can I join anyone’s band of AI server farm raiders 24 months from now? Anyone forming a group? I will bring my meat bicycle.
Nonsense. But then CEOs rarely know what the hell they’re talking about.
If you go forward 12 months the AI bubble will have burst. If not sooner.
Most companies who bought into the hype are now (or will be soon) realizing it’s nowhere near the ROI they hoped for, that the projects they’ve been financing are not working out, that forcing their people to use Copilot did not bring significant efficiency gains, and more and more are realizing they’ve been exchanging private and/or confidential data with Microsoft and boy there’s a shitstorm gathering on that front.
If you have the ability to build an AI app in house - holy shit shit that can improve productivity. Copilot itself for office use… Meh so far.
The most successful ML in-house projects I’ve seen took at least 3 times as long than initially projected to become usable, and the results were underwhelming.
You have to keep in mind that most of the corporate ML undertakings are fundamentally flawed because they don’t use ML specialists. They use eager beavers who are enthusiastic about ML and entirely self-taught and will move on in 1 year and want to have “AI” on their resume when they leave.
Meanwhile, any software architect worth their salt will diplomatically avoid to give you any clear estimate for anything having to do with ML – because it’s basically a black box full of hopes and dreams. They’ll happily give you estimates and build infrastructure around the box but refuse to touch the actual thing with a ten foot pole.
There aren’t enough AI specialists. More are being created by picking up these projects.
The problem is that AI is too hyped and people are trying to solve things it probably can’t solve. The projects I have seen work are basically fancy data ingress/parsing/summarisation apps. That’s where the current AI tech can really shine.
I’ll take “things business people dont understand” for 100$.
No one hires software engineers to code. You’re hired to solve problems. All of this AI bullshit has 0 capability to solve your problems, because it can only spit out what it’s already
stolen fromseen somewhere elseIt can also throw things against the wall with no concern for fitness-to=purpose. See “None pizza, left beef”.
I’ve worked with a few PMs over my 12 year career that think devs are really only there to code like trained monkeys.
I’m at the point where what I work on requires such a depth of knowledge that I just manage my own projects. Doesn’t help that my work’s PM team consistently brings in new hires only to toss them on the difficult projects no one else is willing to take. They see a project is doomed to fail so they put their least skilled and newest person on it so the seniors don’t suffer any failures.
Simplifying things to a level that is understandable for the PMs just leads to overlooked footguns. Trying to explain a small subset of the footguns just leads to them wildly misinterpreting what is going on, causing more work for me to sort out what terrible misconceptions they’ve blasted out to everyone else.
If you can’t actually be a reliable force multiplier, or even someone I can rely on to get accurate information from other teams, just get out of my way please.
Everybody talks about AI killing programming jobs, but any developer who has had to use it knows it can’t do anything complex in programming. What it’s really going to replace is program managers, customer reps, makes most of HR obsolete, finance analysts, legal teams, and middle management. This people have very structured, rule based day to days. Getting an AI to write a very customized queuing system in Rust to suit your very specific business needs is nearly impossible. Getting AI to summarize Jira boards, analyze candidates experience, highlight key points of meetings (and obsolete most of them altogether), and gather data on outstanding patents is more in its wheelhouse.
I am starting to see a major uptick in recruiters reaching out to me because companies are starting to realize it was a mistake to stop hiring Software Engineers in the hopes that AI would replace them, but now my skills are going to come at a premium just like everyone else in Software Engineering with skills beyond “put a react app together”
Trouble is, you’re basing all that on now, not a year from now, or 6 months from now. It’s too easy to look at it’s weaknesses today and extrapolate. I think people need to get real about coding and AI. Coding is language and rules. Machines can learn that enormously faster and more accurately than humans. The ones who survive will be those who can wield it as a tool for creativity. But if you think it won’t be capable of all the things it’s currently weak at you’re just kidding yourself unfortunately. It’ll be like anything else - a tool for an operator. Middlemen will be wiped out of the process, of course, but those with money remain those without time or expertise, and there will always be a place for people willing to step in at that point. But they won’t be coding. They’ll be designing and solving problems.
It’s tons easier to repkace CEOs, HR, managers and so on than coders. Coders needs to be creative, an HR or manager not so much. Are they leaving three months from now you think?
I’ll start worrying when they are all gone.
I don’t understand how you could understand how LLMs work, and then write this.
Machines can learn that…
Ah, nevermind.
If you’ll excuse me saying, I feel that you are the one who is looking at something and extrapolating.
We are 18 months into AI replacing me in 6 months. I mean… the CEO of OpenAI as well as many researchers have already said LLMs have mostly reached their limit. They are “generalizers” and if you ask them to do anything new they hallucinate quite frequently. Trying to get AI to replace developers when it hasn’t even replaced other menial office jobs is like saying “we taught AI to drive, it will replace all F1 drivers in 6 months”.
McDonald’s tried to get AI to take over order taking. And gave up.
Yeah, it’s not going to be coming for programmer jobs anytime soon. Well, except maybe a certain class of folks that are mostly warming seats that at most get asked to prep a file for compatibility with a new Java version, mostly there to feed management ego about ‘number of developers’ and serve as a bragging point to clients.
The real work of software engineering isn’t the coding. That is like saying that being a doctor is all about reading health charts. Planning, designing, testing and maintaining software is the hard part, and it is often much more political than it is a technical challenge. I’m not worried about getting replaced by AI. In fact, LLMs ability to generate high volumes of code only makes the skills to understand it to be more in demand.
An inherent flaw in transformer architecture (what all LLMs use under the hood) is the quadratic memory cost to context. The model needs 4 times as much memory to remember its last 1000 output tokens as it needed to remember the last 500. When coding anything complex, the amount of code one has to consider quickly grows beyond these limits. At least, if you want it to work.
This is a fundamental flaw with transformer - based LLMs, an inherent limit on the complexity of task they can ‘understand’. It isn’t feasible to just keep throwing memory at the problem, a fundamental change in the underlying model structure is required. This is a subject of intense research, but nothing has emerged yet.
Transformers themselves were old hat and well studied long before these models broke into the mainstream with DallE and ChatGPT.
It’s based on the last few years of messaging. They’ve consistently said AI will do X, Y, and Z, and it ends up doing each of those so poorly that you need pretty much the same staff to babysit the AI. I think it’s actually a net-negative in terms of productivity for technical work because you end up having to go over the output extremely carefully to make sure its correct, whereas you’d have some level of trust with a human employee.
AI certainly has a place in a technical workflow, but it’s nowhere close to replacing human workers, at least not right now. It’ll keep eating at the fringes for the next 5 years minimum, if not indefinitely, and I think the net result will be making human workers more productive, not replacing human workers. And the more productive we are per person, the more valuable that person is, and the more work gets generated.
Copilot can’t even suggest a single Ansible or Terraform task without suggesting invalid/unsupported options. I can’t imagine how bad it is at doing anything actually complex with an actual programming language.
It also doesn’t know what’s going on a couple line before it, so say I am in a language that has options for functional styling using maps and I want to keep that flow going, it will start throwing for loops at you, so you end up having to rewrite it all anyway. I have find I end up spending more time writing the prompts then validating it did what I want correctly (normally not) than just looking at the docs and doing it myself, the bonus being I don’t have to reprompt it again later because now I know how to do it
Let’s assume this is true, just for discussion’s sake. Who’s going to be writing the prompts to get the code then? Surely someone who can understand the requirements, make sure the code functions, and then test it afterwards. That’s a developer.
I think that’s the point? They’re saying that those coders will turn into prompt engineers. They didn’t say they wouldn’t have a job, just that they wouldn’t be “coding”.
Which I don’t believe for a minute. I could see it eventually, but it’s not “2 years” away by any stretch of the imagination.
Possibly. But… Here’s the thing. I’ve dealt with “business rules” engines before at a job. I used a few different ones. The idea is always to make coding simpler so non technical people can do it. Unless you couldn’t tell from context, I’m a software engineer lol. I was the one writing and troubleshooting those tools. And it was harder than if it was just in a “normal” language like Java or whatever.
I have a soft spot for this area and there’s a non zero chance this comment makes me obsess over them again for a bit lol. But the point I’m making is that “normal” coding was always better and more useful.
It’s not a perfect comparison because LLMs output “real” code and not code that is “Scratch-like”, but I just don’t see it happening.
I could see using LLMs exclusively over search engines (as a first place to look that is) in 2 years. But we’ll see.
Definitely be coding less I think. Coding or programming is basically the “grunt work”. The real skill is understanding requirements and translating that into some product.
No, going by them, they just talk to an AI voice and it will pop out a finished product.
I don’t believe for a single instance that what he says is going to happen, this is just a play for funding… But if it were to happen I’m pretty sure most companies would hire anything that moves for those jobs. You have many examples of companies offloading essential parts of their products externally.
I’ve also seen companies hiring tourism graduates (et al non engineering related) giving them a 3/4 week programming course, slapping a “software engineer” sticker on them and off they are to work on products they have no experience to work on. Then it’s up to senior engineers to handle all that crap.
This explains so much about 1 in 4 IT people I meet.
The only people who would say this are people that don’t know programming.
LLMs are not going to replace software devs.
The one thing that LLMs have done for me is to make summarizing and correlating data in documents really easy. Take 20 docs of notes about a project and have it summarize where they are at so I can get up to speed quickly. Works surprisingly well. I haven’t had luck with code requests.
I can see the statement in the same way word processing displaced secretaries.
There used to be two tiers in business. Those who wrote ideas/solutions and those who typed out those ideas into documents to be photocopied and faxed. Now the people who work on problems type their own words and email/slack/teams the information.
In the same way there are programmers who design and solve the problems, and then the coders who take those outlines and make it actually compile.
LLM will disrupt the programmers leaving the problem solvers.
There are still secretaries today. But there aren’t vast secretary pools in every business like 50 years ago.
I wrote a comment about this several months ago on my old kbin.social account. That site is gone and I can’t seem to get a link to it, so I’m just going to repost it here since I feel it’s relevant. My kbin client doesn’t let me copy text posts directly, so I’ve had to use the Select feature of the android app switcher. Unfortunately, the comment didn’t emerge unscathed, and I lack the mental energy to fix it due to covid brain fog (EDIT: it appears that many uses of
Iwere not preserved). The context of the old post was about layoffs, and it can be found here: https://kbin.earth/m/[email protected]/t/12147I want to offer my perspective on the Al thing from the point of view of a senior individual contributor at a larger company. Management loves the idea, but there will be a lot of developers fixing auto-generated code full of bad practices and mysterious bugs at any company that tries to lean on it instead of good devs. A large language model has no concept of good or bad, and it has no logic. happily generate string- templated SQL queries that are ripe for SQL injection. I’ve had to fix this myself. Things get even worse when you have to deal with a shit language like Bash that is absolutely full of God awful footguns. Sometimes you have to use that wretched piece of trash language, and the scripts generated are horrific. Remember that time when Steam on Linux was effectively running rm -rf /* on people’s systems? I’ve had to fix that same type of issue multiple times at my workplace.
I think LLMs will genuinely transform parts of the software industry, but I absolutely do not think they’re going to stand in for competent developers in the near future. Maybe they can help junior developers who don’t have a good grasp on syntax and patterns and such. I’ve personally felt no need to use them, since spend about 95% of my time on architecture, testing, and documentation.
Now, do the higher-ups think the way that do? Absolutely not. I’ve had senior management ask me about how I’m using Al tooling, and they always seem so disappointed when I explain why I personally don’t feel the need for it and what feel its weaknesses are. Bossman sees it as a way to magically multiply IC efficiency for nothing, so absolutely agree that it’s likely playing a part in at least some of these layoffs.
Basically, I think LLMs can be helpful for some folks, but my experience is that the use of LLMs by junior developers absolutely increases the workload of senior developers. Senior developers using LLMs can experience a productivity bump, but only if they’re very critical of the output generated by the model. I am personally much faster just relying on traditional IDE auto complete, since I don’t have to change from “I’m writing code” mode to “I’m reviewing code mode.”
Will there even be a path for junior level developers?
The same one they have now, perhaps with a steeper learning curve. The market for software developers is already saturated with disillusioned junior devs who attended a boot camp with promises of 6 figure salaries. Some of them did really well, but many others ran headlong into the fact that it takes a lot more passion than a boot camp to stand out as a junior dev.
From what I understand, it’s rough out there for junior devs in certain sectors.
The one colleague using AI at my company produced (CUDA) code with lots of memory leaks that required two expert developers to fix. LLMs produce code based on vibes instead of following language syntax and proper coding practices. Maybe that would be ok in a more forgiving high level language, but I don’t trust them at all for low level languages.
I was trying to use it to write a program in python for this macropad I bought and I have yet to get anything usable out of it. It got me closer than I would have been by myself and I don’t have a ton of coding experience so it’s problems are probably partially on me but everything it’s given me has required me to correct it to work.
The problem with this take is the assertion that LLMs are going to take the place of secretaries in your analogy. The reality is that replacing junior devs with LLMs is like replacing secretaries with a network of typewriter monkeys who throw sheets of paper at a drunk MBA who decides what gets faxed.
I’m saying that devs will use LLM’s in the same way they currently use word processing to send emails instead of handing hand written notes to a secretary to format, grammar/spell check, and type.
Good take
There is no reason to believe that LLM will disrupt anyone any time soon. As it stands now the level of workmanship is absolutely terrible and there are more things to be done than anyone has enough labor to do. Making it so skilled professionals can do more literally just makes it so more companies can produce quality of work that is not complete garbage.
Juniors produce progressively more directly usable work with reason and autonomy and are the only way you develop seniors. As it stands LLM do nothing with autonomy and do much of the work they do wrong. Even with improvements they will in near term actually be a coworker. They remain something you a skilled person actually use like a wrench. In the hands of someone who knows nothing they are worth nothing. Thinking this will replace a segment of workers of any stripe is just wrong.
No
It’ll have to improve a magnitude for that effect. Right now it’s basically an improved stack overflow.
…and only sometimes improved. And it’ll stop improving if people stop using Stack Overflow, since that’s one of the main places it’s mined for data.
Nah, it’s built into the editors and repos these days.
?
If no one uses Stack Overflow anymore, then no one posts new answers. So AI has no new info to mine.
They are mining the IDE and GitHub.
You seem to be missing what I’m saying, and missing my point. But I’m not going to try to rephrase it again.
I thought by this point everyone would know how computers work.
That, uh, did not happen.
I don’t know if you noticed but most of the people making decisions in the industry aren’t programmers, they’re MBAs.
Irrelevant, anyone who tries to replace their devs with LLMs will crash and burn. The lessons will be learned. But yes, many executives will make stupid ass decisions around this tech.
It’s really sad how even techheads ignore how rapidly LLM coding has come in the last 3 years and what that means in the long run.
Just look how rapidly voice recognition developed once Google started exploiting all of its users’ voice to text data. There was a point that industry experts stated ‘There will never be a general voice recognition system that is 90%+ across all languages and dialects.’ And google made one within 4 years.
The natural bounty of a no-salary programmer in a box is too great for this to ever stop being developed, and the people with the money only want more money, and not paying devs is something they’ve wanted since the coding industry literally started.
Yes its terrible now, but it is also in its infancy, like voice recognition in the late 90s it is a novelty with many hiccoughs. That won’t be the case for long and anyone who confidently thinks it can’t ever happen will be left without recourse when it does.
But that’s not even the worst part about all of this but I’m not going into black box code because all of you just argue stupid points when I do but just so you know, human programming will be a thing of the past outside of hobbyists and ultra secure systems within 20 years.
Maybe sooner
Maybe in 20 years. Maybe. But this article is quoting CEOs saying 2 years, which is bullshit.
I think it’s just as likely that in 20 years they’ll be crying because they scared enough people away from the career that there aren’t enough developers, when the magic GenAI that can write all code still doesn’t exist.
yeah 2 years is bullshit but with innovation, 10 years is still reasonable and fucking terrifying.
Wrong, this is also exactly what people selling LLMs to people who can’t code would say.
It’s this. When boards and non-tech savvy managers start making decisions based on a slick slide deck and a few visuals, enough will bite that people will be laid off. It’s already happening.
There may be a reckoning after, but wall street likes it when you cut too deep and then bounce back to the “right” (lower) headcount. Even if you’ve broken the company and they just don’t see the glide path.
It’s gonna happen. I hope it’s rare. I’d argue it’s already happening, but I doubt enough people see it underpinning recent lay offs (yet).
That’s not what was said. He specifically said coding.
AI as a general concept probably will at some point. But LLMs have all but reached the end of the line and they’re not nearly smart enough.
LLMs have already reached the end of the line 🤔
I don’t believe that. At least from an implementation perspective we’re extremely early on, and I don’t see why the tech itself can’t be improved either.
Maybe it’s current iteration has hit a wall, but I don’t think anyone can really say what the future holds for it.
I’m not trained in formal computer science, so I’m unable to evaluate the quality of this paper’s argument, but there’s a preprint out that claims to prove that current computing architectures will never be able to advance to AGI, and that rather than accelerating, improvements are only going to slow down due to the exponential increase in resources necessary for any incremental advancements (because it’s an NP-hard problem). That doesn’t prove LLMs are end of the line, but it does suggest that additional improvements are likely to be marginal.
LLMs have been around since roughly
20162017 (comment below corrected me that Attention paper was 2017). While scaling the up has improved their performance/capabilities, there are fundamental limitations on the actual approach. Behind the scenes, LLMs (even multimodal ones like gpt4) are trying to predict what is most expected, while that can be powerful it means they can never innovate or be truth systems.For years we used things like tf-idf to vectorize words, then embeddings, now transformers (supped up embeddings). Each approach has it limits, LLMs are no different. The results we see now are surprisingly good, but don’t overcome the baseline limitations in the underlying model.
The “Attention Is All You Need” paper that birthed modern AI came out in 2017. Before Transformers, “LLMs” were pretty much just Markov chains and statistical language models.
You’re right, I thought that paper came out in 2016.
we’re extremely early on
Oh really! The analysis has been established since the 80’s. Its so far from early on that statement is comical
Transformers, the foundation of modern “AI”, was proposed in 2017. Whatever we called “AI” and “Machine Learning” before that was mostly convolutional networks inspired by the 80’s “Neocognitron”, which is nowhere near as impressive.
The most advanced thing a Convolutional network ever accomplished was DeepDream, and visual Generative AI has skyrocketed in the 10 years since then. Anyone looking at this situation who believes that we have hit bedrock is delusional.
From DeepDream to Midjourney in 10 years is incredible. The next 10 years are going to be very weird.
“at some point” being like 400 years in the future? Sure.
Ok that’s probably a little bit of an exaggeration. 250 years.
