It's just as crazy as saying "We don't need math, because every problem can be described using human language".
In other words, that might be true as long as your problem is not complex enough to be able to be understood using human language.
You want to solve a real problem? It's way more complex with so many moving parts you can't just take LLM to solve it, because that takes an actual understanding of a problem.
If you ever write code for a living first thing you notice is that people can't explain what they need by using natural language ( which is what English, Mandarin etc is), even if they don't need to get into details.
Maybe more apt for me would be, “We don’t need to teach math, because we have calculators.” Like…yeah, maybe a lot of people won’t need the vast amount of domain knowledge that exists in programming, but all this stuff originates from human knowledge. If it breaks, what do you do then?
I think someone else in the thread said good programming is about the architecture (maintainable, scalable, robust, secure). Many LLMs are legit black boxes, and it takes humans to understand what’s coming out, why, is it valid.
Even if we have a fancy calculator doing things, there still needs to be people who do math and can check. I’ve worked more with analytics than LLMs, and more times than I can count, the data was bad. You have to validate before everything else, otherwise garbage in, garbage out.
It’s sounds like a poignant quote, but it also feels superficial. Like, something a smart person would say to a crowd to make them say, “Ahh!” but also doesn’t hold water long.
And because they are such black boxes, there's the sector of Explainable AI which attempts to provide transparency.
However, in order to understand data from explainable AI, you still need domain experts that have experience in interpreting what that data means and how to make changes.
It's almost as if any reasonably complex string of operations requires study. And that's what tech marketing forgets. As you said, it all has to come from somewhere.
I don't see how it would be possible to completely replace programmers. The reason we have programming languages instead of using natural language is that the latter has ambiguities. If you start having to describe your software's behaviour in natural language, then one of three things can happen:
either this new natural programming language has to make assumptions about what you intend, and thus will only be capable of outputting a certain class of software (i.e. you can't actually create anything new),
or you need to learn a new way of describing things unambiguously, and now you're back to programming but with a new language,
or you spend forever going back and forth with the generator until it gives you the output you want, and this would take a lot longer to do than just having an experienced programmer write it.
Here’s an honest take. Let me preface this with some credential: I’m an AI Engineer with many years in field. I’m directly working on right now multiple projects that augment and automate code generation, documentation, completion and even system design/understanding. We’re not there yet. But the pace of progress in how fast we are improving our code-AI is astounding. Exponential growth in capability and accuracy and utility.
As an anecdotal example; a few years ago I decided I would try to learn Rust (programming language), because it seemed interesting and we had a practical use case for a performant, memory-efficient compiled language. It didn’t really work out for me, tbh. I just didn’t have the time to get very fluent with it enough to be effective.
Now I’m on a project which also uses Rust. But with ChatGPT and some other models I’ve deployed (Mixtral is really good!) I was basically writing correct, effective Rust code within a week—accepted and merged to main.
I’m actively using AI code models to write code to train, fine-tune, and deploy AI code models. See where this is going? That’s exponential growth.
I honestly don’t know if I’d recommend to my young kids programming as a career now even if it has been very lucrative for me and will take me to my retirement just fine. It excites me and scares me at the same time.
There is more to a program then writing logic. Good engineers are people who understand how to interpret problems and translate the inherent lack of logic in natural language into something that machines are able to understand (or vice versa).
The models out there right now can truly accelerate the speed of that translation - but translation will still be needed.
An anecdote for an anecdote. Part of my job is maintaining a set of EKS clusters where downtime is... undesirable (five nines...). I actively use chatgpt and copilot when adjusting the code that describes the clusters - however these tools are not able to understand and explain impacts of things like upgrading the control plane. For that you need a human who can interpret the needs/hopes/desires/etc of the stakeholders.
Yeah I get it 100%. But that’s what I’m saying. I’m already working on and with models that have entire codebase level fine-tuning and understanding. The company I work at is not the first pioneer in this space. Problem understanding and interpretation— all of what you said is true— there are causal models being developed (I am aware of one team in my company doing exactly that) to address that side of software engineering.
So. I don’t think we are really disagreeing here. Yes, clearly AI models aren’t eliminating humans from software today; but I also really don’t think that day is all that far away. And there will always be need for humans to build systems that serve humans; but the way we do it is going to change so fundamentally that “learn C, learn Rust, learn Python” will all be obsolete sentiments of a bygone era.
Let's be clear - current AI models are being used by poor leadership to remove bad developers (good ones don't tend to stick around). This however does place some pressure on the greater tech job market (but I'd argue no different then any other downturn we have all lived through).
That said, until the issues with being confidently incorrect are resolved (and I bet people a lot smarter then me are tackling the problem) it's nothing better then a suped up IDE. Now if you have a public resources you can point me to that can look at a meta repo full of dozens of tools and help me convert the python scripts that are wrappers of wrappers( and so on) into something sane I'm all ears.
I highly doubt we will ever get to the point where you don't need to understand how an algorithm works - and for that you need to understand core concepts like recursion and loops. As humans brains are designed for pattern recognition - that means writing a program to solve a sodoku puzzle.
It's not really about the coding, it's about the process of solving the problem. And ai is very far away from being able to do that. The language you learn to code in is probably not the one you will use much of you life. It will just get replaced by which ai you will use to code.
I can kind of see his point, but the things he is suggesting instead (biology, chemistry, finance) don't make sense for several reasons.
Besides the obvious "why couldn't AI just replace those people too" (even though it may take an extra few years), there is also a question of how many people can actually have a deep enough expertise to make meaningful contributions there - if we're talking about a massive increase of the amount of people going into those fields.
Anyone who understands how these models works can see plain as day we have reached peak LLM. Its enshitifying on itself and we are seeing its decline in real time with quality of generated content. Dont believe me? Go follow some senior engineers.
There is a reason they didn't offer specific examples. LLM can still scale by size, logical optimization, training optimization, and more importantly integration. The current implementation is reaching it's limits but pace of growth is also happening very quickly. AI reduces workload, but it is likely going to require designers and validators for a long time.
For sure evidence is mounting that model size benefit is not returning the quality expected. Its also had the larger net impact of enshitifying itself with negative feedback loops between training data, humans and back to training. This one being quantified as a large declining trend in quality. It can only get worse as privacy, IP laws and other regulations start coming into place. The growth this hype master is selling is pure fiction.
Secondhand is better value, still new cost right now is nothing short of price fixing. You only need look at the size reduction in memory since A100 was released to know what’s happening to gpu’s.
We need serious competition, hopefully intel is able to but foreign competition would be best.
Fediverse is sadly not as popular as we would like sorry cant help here. That said i follow some researchers blogs and a quick search should land you with some good sources depending on your field of interest
You asked the question already answered. Pick your platform and you will find a lot of public research on the topic. Specifically for programming even more so
Large language models are amazingly useful coding tools. They help developers write code more quickly.
They are nowhere near being able to actually replace developers. They can't know when their code doesn't make sense (which is frequently). They can't know where to integrate new code into an existing application. They can't debug themselves.
Try to replace developers with an MBA using a large language model AI, and once the MBA fails, you'll be hiring developers again - if your business still exists.
Every few years, something comes along that makes bean counters who are desperate to cut costs, and scammers who are desperate for a few bucks, declare that programming is over. Code will self-write! No-code editors will replace developers! LLMs can do it all!
No. No, they can't. They're just another tool in the developer toolbox.
I've been a developer for over 20 years and when I see Autogen generate code, decide to execute that code and then fix errors by making a decision to install dependencies, I can tell you I'm concerned. LLMs are a tool, but a tool that might evolve to replace us. I expect a lot of software roles in ten years to look more like an MBA that has the ability to orchestrate AI agents to complete a task. Coding skills will still matter, but not as much as soft skills will.
Think about a modern application. Think about the file structure, how the individual sources interrelate, how non-code assets are stored, how applications are deployed, and all the other bits and pieces that go into an application. An AI can't know any of that without being trained - by a human - on the specifics of that application's needs.
I use Copilot for my job. It's very nice, and makes my job easier. And if my boss fired me and the rest of the team and tried to do it himself, the application would be down in a day, then irrevocably destroyed in a week. Then he'd be fired, we'd be rehired, and we - unlike my now-former boss - would know things like how to revert the changes he made when he broke everything while trying to make Copilot create a whole new feature for the application.
AI code generation is pretty cool, but without the capacity to know what code actually should be generated, it's useless.
It's just going to create a summary story about the code base and reference that story as it implements features, not that different that a human. It's not necessarily something it can do now but it will come. Developers are not special, and I was never talking about Copilot.
I don't think most people grok just how hard implementing that kind of joined-up thinking and metacognition is.
You're right, developers aren't special, except in those ways all humans are, but we're a very long way indeed from being able to simulate them in AI - especially in large language models. Humans automatically engage in joined-up thinking, second-order logic, and so on, without having to consciously try. Those are all things a large language model literally can't do.
It doesn't know anything. It can't conceptualize a "summary story," or understand parts that it might get wrong in such a story. It's glorified autocomplete.
And that can be extraordinarily useful, but only if we're honest with ourselves about what it is and is not capable of.
Companies that decide to replace their developers with one guy using ChatGPT or Gemini or something will fail, and that's going to be true for the foreseeable future.
Try for a second to think beyond what they're able to do now and think about the future. Also, educate yourself on Autogen and CrewAI, you actually haven't addressed anything I said because you're too busy pontificating.
Try for a second to think beyond what they're able to do now and think about the future.
I am. In the future, they will need to be able to perform tasks using joined-up thinking, second-order logic, and metacognition if they're going to replace people like me with AI. And that is a very hard goal to achieve. Maybe not P = NP hard, but by no means trivial.
Also, educate yourself on Autogen and CrewAI, you actually haven't addressed anything I said because you're too busy pontificating.
I have. My company looked at Autogen. We concluded it wasn't worth it. The solution to AI agents not being able to actually understand what they're doing isn't to amplify the problem by creating teams of them.
Every few years, something new comes along driven by incredible hype, and people declare programming to be dead. They insist a robot will be able to do my job. I have yet to see a technology that will plausibly do that in ten years, let alone now. And all the hype is built on a foundation of ignorance over how complicated a modern, enterprise-ready application is, and how necessary being able to think about its many moving parts is.
You know who doesn't suffer from that ignorance? Microsoft, the creators of Autogen. And they're currently hiring developers, not laying them off and replacing them with Autogen.
Well, I sometimes see a few tools at my job, which are supposed to be kinda usable by people like that. In reality they can't 90% of time.
That'd be because many people think that engineers deal in intermediate technical details, and the general idea is clear for this MBA. In fact it's not.
I'm so sick of the hyper push for specialization, it may be efficient but it's soul crushing. IDK maybe it's ADHD but I'd rather not do any one thing for more than 2 weeks.
Most people got jobs with no experience, or even education less than 40 years ago as long as they showed up and acted confident. Nowadays entry level internships want MScs and years of work xp with something that was invented yesterday
You mistake 'businesses were more willing to take a chance' for 'businesses didn't want experience'. Unless it's a bullshit retail job, anybody with no experience is a net burden their first month.
You remember when everyone was predicting that we are a couple of years away from fully self-driving cars. I think we are now a full decade after those couple of years and I don't see any fully self driving car on the road taking over human drivers.
We are now at the honeymoon of the AI and I can only assume that there would be a huge downward correction of some AI stocks who are overvalued and overhyped, like NVIDIA. They are like crypto stock, now on the moon tomorrow, back to Earth.
If you were able to generate near life-like images and simulacrams of human speech why would you tell anyone?
Money. The answer is money.
Quantum computing wouldn't be developed just to break encryption, the exponential increase in compute power would fuel a technological revolution. The encryption breaking would be the byproduct.
Waymo exists and is now moving passengers around in three major cities. It's not taking over yet, but it's here and growing.The timeframe didn't meet the hype but the technology is there.
Yes, the technology is there but it is not Level 5, it is 3.5-4 at best.
The point with a full self-driving car is that complexity increases exponentially once you reach 98-99% and the last 1-2% are extremely difficult to crack, because there are so many corner cases and cases you can't really predict and you need to make a car that drives safer than humans if you really want to commercialize this service.
Same with generative AI, the leap at first was huge, but comparing GPT 3.5 to 4 or even 3 to 4 wasn't so great. And I can only assume that from now on achieving progress will get exponentially harder and it will require usage of different yet unknown algorithms and models and advances will be a lot more modest.
And I don't know for you but ChatGPT isn't 100% correct especially when asking more niche questions or sending more complex queries and often it hallucinates and sometimes those hallucinations sound extremely plausible.
vulcanpost.com
Hot