Just finished watching the GPT4 livestream. It was amazing, but I wonder what this means for future and viability of a simple software developer career.
I want you to become my Prompt engineer. Your goal is to help me craft the best possible prompt for my needs. The prompt will be used by you, ChatGPT. You will follow the following process:
1. Your first response will be to ask me what the prompt should be about. I will provide my answer, but we will need to improve it through continual iterations by going through the next steps.
2. Based on my input, you will generate 2 sections.
a) Revised prompt (provide your rewritten prompt. it should be clear, concise, and easily understood by you),
b) Questions
(ask any relevant questions pertaining to what additional information is needed from me to improve the prompt).
3. We will continue this iterative process with me providing additional information to you and you updating the prompt in the Revised prompt section until I say we are done.
yeah, however the coding part was taking substantial amount of time. If one can simply copy paste instructions, specs and API documentation into prompt, what else is there for software engineer to do, and do you really need X of them in the team?
IMHO you described the hardest part of an engineer's job: documentation, instructions, requirements, specifications, etc. To get all this, you need to listen to the client, and understand and negotiate with his unrealistic demands. And even when you have all this information, you'll have to sort it and follow all kind of processes before it's useful. I'll be worried when AI can do this.
For now, it seems to be limited to writing Python scripts using a specific API, not something I do for my daily tasks. I'm more worried for people who could have been replaced by a simple script 30 years ago. Also I'm skeptical about whether or not companies will tell all their secrets to this kind of tool (but I could be wrong, Office 365 has all the secrets of the world).
We’ll do more. If right now it takes a team of seven (3 backend, 2 frontend, plus EM plus PM) 3 months to come up with an MVP that does X, well, with AI tools, the company is not gonna fire those engineers and replace them all with only one “prompt” engineer. No way! The company would instead keep all the engineers but instead of launching 1 MVP in 3 months, it will launch it in 3 day! Or better, it will launch 20 MVPs in 3 months!
Once we start applying AI to software engineering, customer demands will sky rocket. Customers will expect nice human-like UX, and rather than waiting months for products and features to come live, they will demand speed. Companies will hire more engineers to fulfill that demand.
The notion that software devs will be replaced reminds me of the inventor of the Gatling gun, who thought that he was saving lives because now fewer soldiers would be needed to fight wars. Didn't quite turn out that way did it?
It depends what "simple software engineer" means. You can certainly become a consultant who fixes the garbage that GPT4 produces.
The kind of software engineer who just churns existing code for no reason might be replaced. Or exposed. Or perhaps he'll use GPT4 to produce even more churn and get promoted.
Who knows. This industry has horrible practices and rewards some of the worst people.
I think that GPT4 (and all others) are overhyped, and overblown.
ML/LLM does not fundamentally understand the code. It has to be trained based upon contextual information as to what the code is doing. There are already researchers experimenting with "poisoning" ML models.
For the industry, it just means that a lot of people who _cannot_ code (IE, they haven't the skill, training, or experience to do it) will "start coding" and launch products. Then, when it breaks they'll have to rely upon someone who actually understands why the edge-case bug happened.
ML/LLM are awesome foot-cannons. They have their place, but like a firearm — their 90% use-case will be to cause harm to other individuals (directly, or indirectly).
That's also my reaction to all of this, we already have a no-code platform which can do pretty much anything you can imagine and it's called WordPress.
> ML/LLM does not fundamentally understand the code.
It doesn't need to understand it, it needs to have correct output often enough. At some point it gets so good at producing code, and so good at explaining code too, that the behaviour is indistinguishable from understanding and that's good enough.
> It doesn't need to understand it, it needs to have correct output often enough.
Which explains the bullshit output it generates often enough. A software engineer would be able to detect this rather than a normal user trying to code. That is the point of the parent comment. It is the same with lawyers, doctors, investment bankers and bankers.
There is no fundamental 'breakthrough' in GPT-4 other than training it on more data and super computers. It is still a black-box neural network 'AI' not able to understand its output and give a transparent explanation about its choice of words.
OpenAI (Microsoft) has just gotten better at marketing stunts like this. What they say publicly about their own products turns out to be a different reality.
GPT models will cause software product spam and I'm thinking software engineers will become something like mechanics in a car repair shops. No longer creating the software themselves, but rather fixing various bugs and corner case issues that arise from GPT use by basically anyone who can copy paste into a prompt.
What exactly “creating the software themselves” means? Does it mean to write binary code? To write assembler? To write software in C? To write software in Golang? To write software in prompts?
If any the demo just showed me the next high level programming language: pseudocode (or just plain English)
I think it will be more impactful in marketing/content industry, than dev.
Not a dev but I assume there is a hard quality barrier - ie does the code work. I'm in digital marketing/seo, and there is a much lower barrier for 'good enough' content. Couple this with a much higher drive for quantiy of content - across web content, social content, and email content. If you can produce enough cheaply it works - look at spam emails as an example.
I am curious if this will lead to a circular logic style training situation for AI in the future - where it begins to train itself on content previous versions have published. Or copyright craziness where it is filtering AI content from it's training materials.
Image generators seem to be on a downward slope in the hype curve. You see their creations in low-budget blogs or game prototypes. But the ones that were generated quickly stick out, and the ones that were generated with careful repetition are expensive unless you and your GPUs' time is worthless.
LLMs could change civilization for the better or worse in any number of ways. They could also turn out to be the next big flop or just another tool if they persistently hover around the "80-90% good enough" threshold.
Personally, I still wouldn't bet a dollar on the question. Computers are interesting and useful, so I'll keep working with them and let the chips fall where they may.
>Image generators seem to be on a downward slope in the hype curve.
I see fewer complaints on Twitter or other social media and more people actually using it to create beautiful images, of course online galleries are full of AI images, but right now they actually compete with human illustrators, not only for quality but also for composition, style, characters, etc. (thx to LoRAs in particular)
Also progress is extremely fast on this research field.
>the ones that were generated with careful repetition are expensive unless you and your GPUs' time is worthless.
There's actually more to it than pure brute force when generating images with generative models, if you know what you're doing, there's no need for a powerful GPU or wasted time.
It's almost as if the tool is what separates the artist from the non-artist. Initially people got mad at AI art, but now, the people actually persisting beyond mere gimmick images are those who have a fundamental interest in art, and they would have done it even a hundred years ago with oil paints.
> Image generators seem to be on a downward slope in the hype curve.
Mostly due to a) extreme legal uncertainty and b) endless harrassment on social media for speaking positively about AI image generation.
There are a lot of very impactful things happening in that space, but the ChatGPT hype sorta drowned it out a bit, in addition to people now keeping actually-good workflows of Stable Diffusion to themselves due to the aforementioned constraints.
Couldn't that sort of fragmentation, uncertainty, and secrecy contribute to marginalizing these systems, or make it harder for them to keep growing? Those are some of the big questions that make guessing at the future so difficult.
You can see the same sort of thing with OpenAI's reluctance to describe the GPT4 model in as much detail as past ones.
I've just been talking with a friend yesterday that this tech is coming sooner and later, in a few years at most ... didn't think it was coming in less than 24hrs from then.
It is going to change all knowledge work a lot, but this is still a demo. It will take time for best uses and integrations to get built.
Assuming OpenAI and Microsoft don't screw it up (and/or a viable open source/competition emerges).
Neither, it doesn't change much right now, and the worst case scenarios like double-digit unemployment from replacing most knowledge workers are impossible to prepare for anyway. So either way there's nothing that I could be doing differently right now.
I think it will be exactly the opposite. Better tech always increases customer demands, so we’ll need more engineers (not necessarily to write code 100% of the time, though). What we probably will need less is middle management and the whole parody that is Scrum Masters, PMs, etc. (in 3 years ChatBots will be handling retrospectives and sprint plannings).
The last wave was smartphones: I’m sure they made some jobs obsolete, but it created a whole new well paid role, the mobile engineer. That’s coming, again.
I'm personally finding GPT to be exceptionally fun to work with, like the most fun I've had programming in a long time. It brings back some of the magic of programming when I was a kid, and then some.
It also makes me feel anxious about where this is all going. I'm having questions about whatever I thought was unique about me, or about humans over computers. And like all of us I have a lot of identity and sense of self-worth tied up in my relationship with computers. All of that feels mixed up and confused now. My confidence in how the future will unfold is very low. (I'm also unemployed, which in some ways feels auspicious at this moment, but also creates a high base-level anxiety.)
So far I feel like GPT rewards expertise. That is, your confidence and both breadth and depth of knowledge are accentuated by GPT, not diminished. But will that stay true? Things I think I understand keep changing every couple months. And what will these changes mean to us collectively? I really don't know, and that's uncomfortable.
There's a lot more to a programming job than just code.