Just to make this clear up front: I was always very skeptical of LLMs (I'll never call them AI) and still see them as fundamentally limited and unreliable. See my post history if you don't believe me.
Having said that: I work in business software and the models have progressed to the point where typing code and writing classes and functions by hand has become uneconomical. We review the output closely to avoid the idiocy of "vibe-coding", but the quality has improved to the point where we rarely have to outright reject solutions. The discussions have moved to high-level architectural points, where we just request better implementations than the first-shot. We get them, review again and move on. Maybe there are deeper, latent problems beeing embedded in the code-base through this process, but we can't spot any although we are looking intently.
And just like that, my profession has died. Nobody in my team writes code anymore. Things that would have taken days are now done in minutes. We are all just considering design and architecture. We are changing our processes since they are not fit for the new way of doing things. Scrum? What's even the point anymore? We can work through a sprint-backlog that would've taken two weeks in two days now. We do reviews 90% of the time. When we request changes they come back in minutes.
I have a feeling like the ground has been taken from below me. My son wanted to become a developer, that isn't a real job anymore. Everyone around me feels the same way, including our management. We are a small team and a small company, everyone is effective and our jobs are not in danger as long as our product is succesful. But my profession is essentially dead. It's a craft I honed for many, many years and the craft itself is now automated. I wanted to improve my typing skills to code faster, this is now a ridiculous endeavour. We don't know what comes next, but I now only have my analytical skills and knowledge of good architecture to fall back on.
And it all changed so fast. I don't think this will transfer to other professions so easily, but for programming LLMs ended it all in the span of the last four months. I thought I knew what my future holds as a senior dev, now I can barely see past christmas.This is a nightmare.
embedded systems, signal analysis, weird edgecasey lisp stuff...
I work in embedded systems and until recently I would have agreed with you... But it's getting to the point where our agents have access to debuggers/tracers, system I/O logging and basically every tool we would ordinarily use and it's vastly improving the quality of their work.
I always attributed their better performance in web and general software development to more availability of open source examples that they are trained on, but I think the ability to 'self check' was a huge limiting factor. Unit and other pure software testing abstracted away too much of the real system. With real feedback and access to the data sheets it's doing a pretty good job.
Only four months back I told anyone who asked that LLMs would never replace developers, because back then they really couldn't. The modern models in conjunction with their harnesses are a very different thing.
Similar time frame for me. I just spent a few hours documenting how a tool we already have should work (one that's a freaking overcomplicated mess and impossible to manage) and feed those requirements to Argon. It made it 90% accurate in one shot. 5 % difference was an issue with my specifications and 3% a latent bug on the existing framework nobody had noticed.
Freaking scary
That's scary. My most recent experience was a year ago when a coworker vibecoded a bunch of tests that somehow only tested the emulated hardware and forgot to test production code.
Try it again. In the past few months they are good enough for general enterprise dev that what OP said is the reality. We just watch our robots shit out a working solution to a ticket.
doesn't really help you when any proprietary code crossing the network boundary is a security breach.
This is only a short term problem. More and more models are released that are „good enough“ that can be run on local hardware. Prices for components at the moment are a problem, but that’s a money problem not with the technology itself.
What does "good enough" mean here though? At the moment I highly doubt that we'll get anywhere near the performance of frontier models on hardware accessible to the average user, due to the limitations of physics itself. Then on the model side of things I think there's probably a limit on how small we can go before accuracy becomes a major issue, and I don't think that limit would be something that consumer hardware would be able to run (and by consumer hardware I mean your average sub $1000 gaming PC, not bleeding edge M6 Macs)
I dunno man claude one-shotted a firmware build for an ESP32-based project I was working on last weekend. I had to massage it properly and I wouldn't have been able to supervise it properly without my EE degree but it did produce valid and correct firmware.
Can you explain what you mean by "I had to massage it properly" when you one-shotted it? Did you preemptively account for things or something?
I mean, yeah. Asking "do xyz make no mistakes" won't get you what you want. Come up with a plan, decide on what constraints matter, pick libraries you want to use. Decide what doesn't matter. Generally around now I feed it my plan and ask it "do you think I've missed anything?" and read what it suggests and see if I agree with any of it.
From there, plan out what incremental progress looks like. Maybe start with the UART/I2C/SPI library and just reading firmware version off of each child component. Then ask it to go through the plan, increment by increment. Review everything as it does it, watch the console and redirect it if it's making poor choices or just things you know will box yourself in in the future.
On my best day in university my project this weekend would have been weeks of work. It took me a day, not including soldering and packaging after the electronics were functional. It's just undeniably faster at this point. Even if the only thing I used it for was to open the serial port, watch the logs, and fix any bugs I introduced, it would be faster. But you can feed it whole datasheets and it'll construct a client library for your device essentially perfectly on the first try.
I have serious ethical qualms about how LLMs were created, how they're going to be used, the inequality they create, the finances of the industry behind them, the environmental impact... but from a purely technical perspective they can be very powerful tools. But you need to already have domain knowledge to be able to massage it properly - "everyone is engineer" still isn't reality. Someone with experience and good intuition though...
I work in driver developer and once I gave them access to the software verification platforms we use and the lab machines containing the test silicon, the quality improved drastically.
Most of my job is now interacting with people because those things can't meaningfully turn business requirements into architecture yet.
it's interesting how years of security hardening and multi-tier firewall infrastructure to prevent code leaks is all going out the window, isn't it?
I'm not sure exactly what you mean by code leaks here. Like if the AI agent decided to take the proprietary code and post it on social media for some reason?
The same guard rails the humans had are in place for the AI agents too (not that they were super effective for humans to begin with).
anything the model ingests leaves the premises. that's a leak.
I work for a very large company that owns their AI, so that's not a concern in my case.
But even if I wasn't, I would assume the company would have negotiated a contract with whoever provides the AI that meets their requirements. That's above my pay grade.
I would never use a personal AI for company code. Even if public AI offerings aren't blocked on corporate devices, that sounds like a great way to get in trouble for mishandling company IP.
you can't really justify the hosting costs if you have less than 5000 employees
Yes, but someone has to host the data.
Your customers trust your company to host that data because they know you meet certain standards when it comes to security, privacy, and whatever other data handling requirements matter to them.
You can make the same agreements with your suppliers that your customers make with you to ensure their handling of data meets your (and your customer's) requirements. In this case, the "supplier" is the company that hosts the AI models.
don't think i've ever worked for a company that hosts customer data. we develop and build products, customer purchases product, transaction over.
You think they're using in-house models to do their coding? 😂
I mean. You can claim that. And maybe its not via LLM, but if you are using LLM to use other models, it absolutely can.
i have no idea what you are trying to say.
I think "No AI" gets very myopically focused on "LLM"s but there are many other kinds of machine learning, and something agents are phenomenal at, is implementing basic code.
So sure, an LLM might suck at denovo prediction of sequence data. But an LLM can instead pull in a model that is quite good at sequence data.
Being against something doesnt mean we can be in denial of it.
i'm not opposed to machine learning in general, i don't think anyone here is. i am myopically opposed to large language models. no other form of machine learning is causing the damage they currently are.
i've used machine learning techniques in my code for many years. llms are awful at seeing and applying those techniques.