I’m a dev. I’ve been for a while. My boss does a lot technology watch. He brings in a lot of cool ideas and information. He’s down to earth. Cool guy. I like him, but he’s now convinced that AI LLMs are about to swallow the world and the pressure to inject this stuff everywhere in our org is driving me nuts.
I enjoy every part of making software, from discussing with the clients and the future users to coding to deployment. I am NOT excited at the prospect of transitioning from designing an architecture and coding it to ChatGPT prompting. This sort of black box magic irks me to no end. Nobody understands it! I don’t want to read yet another article about how an AI enthusiast is baffled at how good an LLM is at coding. Why are they baffled? They have “AI” twelves times in their bio! If they don’t understand it who does?!
I’ve based twenty years of career on being attentive, inquisitive, creative and thorough. By now, in-depth understanding of my tools and more importantly of my work is basically an urge.
Maybe I’m just feeling threatened, or turning into “old man yells at cloud”. If you ask me I’m mostly worried about my field becoming uninteresting. Anyways, that was the rant. TGIF, tomorrow I touch grass.
This sort of black box magic irks me to no end. Nobody understands it!
And that’s why it’s not going to swallow the world. It’s a toy, not a tool.
Tools behave consistently and predictably. You know what you call a tool that doesn’t behave consistently and predictably? “Broken”, that’s what.
Maybe I’m just feeling threatened, or turning into “old man yells at cloud”.
I was “old man yells at cloud” about cryptocurrency for years, and now it’s dead, exactly as I predicted.
Sometimes, the old man is right.
Having an AI help you code is like having a junior developer who is blazing fast, enthusiastic, and listens well. However, it doesn’t think about what it writes, it does no testing, and it doesn’t understand the big picture at all. For very simple tasks, it gets the job done very fast, but for complex tasks no matter how many times you explain it, it is never going to get it. I don’t think there’s any worry about AI replacing developers any time in the foreseeable future.
Edit: fixed voice to text issues.
in fairness… this is more a limitation of the current technology. your look at gpt4 and going not an expert. but what about gpt5 or 6… or some of the newer ideas like microsoft plan for 1 million token model using attention dialation mechnism. The point being we are still on the ground floor. And these models have emgerent functionality
That’s out of the scope of the foreseeable future though. It’s speculating on the capabilities of things that are completely theoretical at this point. It’ll be interesting to see what happens. I’m not holding my breath.
1 billion*
But you can work with it to write all the tests/acceptance criteria and then have the AI run the code against the tests. We spent a lot of time developing processes for humans writing code, we need to continue integrating the machines into these processes. It might not do 100% of the work you’re currently doing, but it could do maybe 50% reliably. That’s still pretty disruptive!
This is a pretty apt analogy, I think.
We’ve been using copilot at work, and it’s really surprised me with some slick suggestions that “mostly work”. But I don’t think it could have written anything beyond the boilerplate my team has done.
(I also spend way too much time watching Copilot and Intellisense fight, and it pisses me off to no end.)
Lol this is what I was thinking too. The junior dev is also a black box. AI automation seems more like delegating than programming to me.
AI can code assist; it’s quite helpful for that. Predictive text, learning a less familiar language, converting pseudo, etc.
But it couldn’t possibly replace senior developers long-term. It just looks new and exciting, especially to people who don’t truly understand how it works. We still need to have human developers capable of writing their own new code.
-
AI is entirely derivative, it’s just copying the human devs of yester-year. If AI does the majority of coding then it becomes incapable of learning, thus necessitating human coders anyway. It also is only going to generate solutions to broad-strokes problems that it already has in its dataset, or convert pseudocode into functional code (which still requires a dev know enough to write pseudo).
-
It also currently has no way of validating what it writes. It’s trying to replicate what our writing looks like contextually, it doesn’t comprehend it. If it ever starts training on itself as it ages, it will stagnate and require human review, which means needing humans that understand code. And that’s not including the poor practices it will already have because so many devs are inconsistent about things like writing comments, documentation, or unit testing. AI doesn’t have its own bias but it inevitably learns to imitate ours.
-
And what about bug-testing? When the AI writes something that breaks, who do you ask for help? The AI doesn’t comprehend the context of the code its reading if you paste it back, it doesn’t remember writing it. You need people who understand how the code works to be able to recognise why it might be breaking.
AI devs are the fast food of coding. It will never be as good quality as something from an experienced professional. But if you’re an awful cook, it still makes it fast and easy to get a sad, flat cheeseburger.
I’ve worked with devs who are the equivalent of line cooks and are also producing sad, flat cheeseburgers: code of poor quality that still sees production because the client doesn’t know any better. IMO, those are the only devs that need to be concerned, because those are the ones that are easy to replace.
If AI coding causes any problems within the job market for devs, it will be that it replaces graduate/junior developers so well that fewer devs get the mentoring or experience to become seniors, and the demand for seniors will rack up significantly. It seems more likely that developers will split into two separate specialisations, not that our single track will be replaced.
-
https://en.wikipedia.org/wiki/Gartner_hype_cycle
Your boss is on the first peak
I see LLM AIs acting more as an assistant rather than being the primary contributor in software projects.
For example, I’m starting a very ambitious personal project and wanted to practice writing a proper project plan and requirements document.
I had no clue where to start, so I pulled up ChatGPT and after some prompting I now have workable rough-drafts that just need some fine details added in and I can focus on actual programming.
Yeah, I agree with most not so interesting code (which is surprisingly much, if I think about it, especially average frontend/backend apps, client side oriented boilerplaty code (say e.g. React UI…)).
But coding a nice smart architecture, something novel/innovative (I think where the art of software engineering really lies IMHO)… well I’m not even thinking anymore about using AI (for now at least), it just confuses me, writes dumb code, and writing back and forth with it is cumbersome (to get better code), so that I just code it myself (being a fast typer is reallly helpful I think…). (I’m using it often though as some kind of StackOverflow replacement, but letting the AI code…? nah).
I think it’ll likely take a few years still where I really seriously can/have to think about using AI productively in these cases (where it may even teach me a few things about language features I haven’t known yet)…
This is based on someone else’s reply I read once. Developers have been trying to put themselves out of their own jobs since the beginning. Automating/scripting things, creating tools, IDEs, etc.
Development is so much more than generating/writing boilerplate code. Code plays such a small role as opposed to figuring out how to solve a problem or even figuring out what the problem is in the first place.
I spent several days figuring out why an HTTP POST in prod wasn’t working. But an identical one was working locally. Turns out there was an application server change that deceased the max request param size. The Dockerfile was configured so that the patch version (semver) was updated automatically. This was a super interesting challenge (felt like Sherlock Holmes with this one).
Try having ChatGPT/etc. figure that one out.
All of this hubbub might produce some kind of toolset that could augment what we already do (i.e. IDE). But replacing people entirely? I don’t think so.
AI LLMs are the best rubber duck. Like a rubber duck, it’s probably not going to “solve” anything for you directly, but LLMs can be a great tool to unlock your potential.
Non-AI rubber ducks don’t steal your code and sell it to your competitors, though.
If a human can access your public repo and read comments posted on public forums, are they stealing your code? LLMs are just aggregators of a great many resources and they aren’t doing anything more than a biological human can already do. The LLM can do so more efficiently than a biological human, while perhaps being more prone to error as it doesn’t completely understand why something is written the way it is. As such any current AI model is prone to signpost errors, but in my experience it has been very good at organizing the broader solution.
I can give you two examples. I started trying to find out how a .Net API call was made. I was trying to implement a retry logic for a call, and I got the answer I asked. I then realized that the AI could do more for me. I asked it to write the routine for me and it suggested using a library which is well suited for that purpose. I asked that it rewrite it without using an external library and it spit it out. I could have written this completely from scratch, in fact I had already come up with something similar but I was missing the API call I was initially looking for. That said, the result actually had some parts I would have had to go back and add, so it saved me a lot of time doing something I already knew how to do.
In a second case, I asked if to solve a problem which at its heart was a binary search. To validate that the answer was correct it would need to go one extra step, but to answer the question it wasn’t necessary to actually perform that last validation step. I was looking for the answer 10, but I got the AI to give me answers in the range of 9-11. It understand the basic concepts, but it still needs a biological human to validate what it generates.
I’m talking about asking the AI “where’s the bug in this code” and pasting a snippet of code from my non-public repository.
Nope, I fully agree with you.
These “AI” tools have no more understanding of what they crap out than a toddler who has learned to swear (someone else made that comparison, I’m just borrowing it).
Have you ever done a code review with someone, asked about a specific part, and they say “I dunno; I copied it from GPT and it just seems to work” ? I have, and it’s absolutely infuriating. Yeah, they could have copied the same from Stack Overflow, I suppose, but I’d treat it the same. But somehow they expect copy/paste from an “AI” to get a pass?
Even without dipping into the “replacement theory” of it, these kinds of tools just allow people who don’t know what they’re doing to pretend like they do while creating a mountain of technical debt. Even experienced devs who use it are starting down a slippery slope, IMO.
At least code on Stack Overflow was written by a human who presumably has some idea of what it actually does.
The replacement theory brings up this weird conundrum where the LLMs need to consume human work to train themselves, the very work they seek to replace. So once the plan succeeds, how does it progress further?
Yeah that’s one of the major issues I have with it. It gives people a way to take their responsibilities, delegate it to an AI, and wash their hands of the inevitable subpar result. Not even just in programming, I think over time we’re going to see more and more metrics replaced with AI scores and businesses escaping liability by blaming it on those AI decisions.
Back in the realm of programming, I’m seeing more and more often people “saving time” by trying to use GPT to do the first 90% but then just not doing the last 90% at all that GPT couldn’t do.
Oh God, I can see it now. Someone makes an AI for filtering job applications, it’s great, all the employers use it. Before a human ever sees a resume, the AI decides whether to silently discard it. For reasons known to literally no one, the AI doesn’t like your name and always discards your resume, no matter how many times you change it. Everybody uses the same AI, so you never get a job again. You end up on the street, penniless, through no fault of your own.
According to this Ted talk (at about 8:15) Amazon tried a resume filtering AI but discovered it was filtering out women.
Yes, and it’ll eventually be worked out to the point that it’s mostly accurate, but there will always be edge cases like the one I described above; they’ll just be rare enough that nobody cares or even believes that it’s happening.
Now, humans reviewing job applications are also subject to biases and will unfairly reject applicants, but that only shuts you out of one company. AIs, on the other hand, are exact copies of each other, so an AI that’s biased against you will shut you out of all companies.
And, again, no one will care that this system has randomly decided to ruin your life.
I agree with you - I think my main issue is that using LLMs to write code is a crutch, and over time the quality of software will decrease because it will be made by a program perfected to generate the most likely next word, rather than understanding what it’s doing. If anything, having a basic understanding of LLMs makes me trust them less.
deleted by creator
completely unrelated to the conversation, but I love your username!
To cool down your boss, you can always tell him that he’s putting your company at great legal risk, there’s no reason to think that LLMs are not violating copyright laws and software licenses, and moreover the case might settle differently in different countries (in case you export your code).
Thanks for reminding me of that. We’re content creators ourselves, so respecting copyright is a big deal.
I think everyone in the software industry is feeling this right now. It was the same way with NFT’s, Blockchain, AR, Even back in the late 2000’s everyone wanted Social Media (crazies). This too shall pass.
Maybe in 6 months, maybe in 2 years, we will hit a point where the limitations and use case are understood well enough and it will just be another tool in our belts and not the end all.
My advice, just smile to your boss and say “sure let’s try it”. Learn something new and look forward to the next one.
They have “AI” twelves times in their bio! If they don’t understand it who does?!
That’s usually a good sign they don’t understand it at all
I’ve recently started using CodeWhisperer and I’ve found I use the creation aspect much less than I do auto-complete. That may change in the future, but right now it’s pretty nice starting to type something and VSCode completes what I was going to type.
I think the part that annoys me the most is the hype around it, just like blockchain. People who don’t know any better claiming magic.
We’ve had a few sequence specific architectures over the years. GRU, LSTM and now Transformers. They were all better than the last at the task of sequence specific transformations, and at least for the last one the specific task was language translation. We eventually figured out these guys have a bit of clairvoyance too, they could make accurate predictions based on past data, or at least accurate enough to bet on, and you can bet traders of various stripes have already made billions off that fact. I’ve even seen a transformer based weather model. It did OK, but transformers are better at language.
And that’s all it is! ChatGPT is a Transformer in the predictive stance. It looks at a transcript of a conversation and thinks what a human is most likely to say next. It’s a very complex transformation of historical data. If you give it the exact same transcript, it gives the exact same answer. It is in the literally mathematically rigorous sense entirely incapable of an original thought. Any perceived sentience is a shadow of OpenAI’s army of annotators or the corpus it was trained on, and I have a hard time assigning sentience to tomorrow’s forecast, which may well have used similar technology. It’s just an ultra fancy search engine index.
Anyways, that’s my rant done I guess. Call it a cynical engineer’s opinion. To be clear I think it’s a fantastic and useful technology, and it WILL change how we interact with machines. It can do fancy things with the combination of “shell” code driving it’s UI like multi-step “agents” or running code, and I actually hope OpenAI extends it far into the future, but I sincerely think any form of AGI will be something entirely different to LLMs, or at least they’ll only form a small part of it as an encoder/decoder for it’s thoughts.
EDIT: Added some paragraph spacing. Sorry, went into a more broad AI rant rather than staying on topic about coding specifically lol
