I've been testing this out and I think my SWE career is dead in the water.
Genuinely wondering what value I bring to my employer right now. What value I will bring in a few months when this gets cheaper.
I think we're screwed. I may only be an SDE 2 at FAANG but I don't think I have promotion opportunities in my future anymore.
Your job is just going to change. You may or may not appreciate/enjoy what it becomes necessarily, but it doesn't mean that you are going to not have a job.
People underestimate how people hate looking at terminals and "weird looking combination of characters" even if they didn't have to write them. If anything, you will likely have more career opportunities in the future, than ever.
And if you get a chance to wet your fingers in cybersecurity - I would take it.
> And if you get a chance to wet your fingers in cybersecurity - I would take it.
Could you explain more? Did some ethical hacking at hackthebox.eu (one insane box, one hard box and a few mediums). But I do not see how I will give additional value to a model.
Just a SWE and data analyst at work, so maybe I am missing something.
I think this really depends on what is interesting to you, personally. Something that you can have fun while doing, without breaking laws.
For example, I'm a privacy nerd, so I like reverse engineering proprietary software to figure out how it works and what data it collects.
I also like getting full access to the hardware I own - like a robot vacuum (bonus point: you'll also learn soldering, probably, which might come in handy if robots take over). Or my Mac studio that imposes some limitations on me on how many active user sessions I can have.
These kinds of things have put me on a path where I've learned how hardware or networking works on deepest levels, what goes through these pipes, how I can place myself in the middle, how I can enter places someone didn't want me to enter.
And once you know how to do these things, you know how to apply this knowledge in defence.
Essentially: always be curious and always try to say "but I want to" when something that doesn't cross the boundary of your physical property says no to you (legally).
Yes, models like Mythos may find vulnerabilities, but your knowledge will make it possible to point it in the right direction, and understand where it's mistaken, or to understand the output when it's right, and what is the right course of action.
Ah, the “I am an expert so I can guide it argument”. Not sure if you are right or wrong. I do know this is the argument that many software engineers claim as well.
Yea, I don’t know if it will hold up. I hope so. It could. I don’t know if it would or wouldn’t.
Get any model, any reasoning level, ask it to tackle a challenge, have it come up with a plan. Then ask it "are you sure? This feels wrong", and it will now think it's wrong. Do that again in a loop and you'll see how unnecessary human judgment actually is.
Or alternatively, have fable write some complex code. Then ask it to do an adversarial review of that code in a clean session. You'll find that it will find issues in the code that it just wrote.
Now imagine you're a layperson who doesn't know which one is true.
Human expertise is never going to become irrelevant.
Yea fair. I have that when I ask an LLM to prove the Riemann hypothesis. I am not mathematically mature. So I can’t see if it approaches it in any way that might yield some insight.
I don't see those career opportunities.
AI is really incredible but in my personal projects it can one-shot things.
I'm trying to figure out how I can get to the point where I have hard problems that AI can't solve, at least not yet.
Because your personal projects likely are not very complex and not high stakes. And you are not responsible to anyone but yourself.
If you're working at a place where this is true about the the organization, then sure, that job will likely be gone. But that was never a good place for your career regardless.
I have 4 concurrent personal projects that are quite complex, but low stakes. I can have SOTA models go wild on them (because low stakes), but they can't one shot anything there. And I can't really work on more than one at a time, even if AI is doing coding - it still requires supervision.
I also frequently nuke these projects and start over because they made a mess there, but I collected necessary knowledge on how to guide them better. You can't do this on a production project, not when there are deadlines and stakeholders.
But just in case some organizations decide to embrace the "trust it blindly" model anyway - cybersecurity specialization will ensure you always have a job.
If you think the job is just writing code then yes you are screwed, just like if you thought your job was just making punch cards. In most roles you have more responsibilities than plainly converting words into text. You're probably not being paid to simply be a human calculator (otherwise you'd be paid a lot less!).
I don't think the job is just writing code. But my career has mostly been about getting a ticket and writing a solution for it. I would really like to solve novel problems, but I don't get many novel problems to solve.
I can architect things but the issue is that Claude can architect things too.
I haven't tried Fable yet but my experience with Claude is it does not engineer things well. Without direction from me, it will either over-engineer things to the point of absurdity, or do the total opposite and have little to no abstraction with repeated code everywhere.
Yeah. I’m not looking forward to years of retraining to earn half the salary either. Us old timers at least got a good 15-20 years out of it. Bananas.
You do realize that this is likely a 10 trillion parameter model that takes something like 20 terabytes of RAM to run inference? Calculate the price for all this VRAM .... It's not getting cheaper in the next few "months".
So this is the one, huh?
If not this one, then definitely 2 model step changes down the line
I agree. Software engineering as we know it is dead. Wonder what it'll evolve into.
I'd say you're cooked if you don't have multi-agent harnesses burning tokens right now. That's going to be a pre-requisite very soon