"This AI unlocks 10 hours a week per person to finally address the backlog of strategic projects we've delayed for the past two years."

That's the right way to think about it, at least initially. And then there's the technical debt backlog, the documentation, and the aspirational projects and improvements that nobody has had time to work on.

"While creating the tests the AI will discard those tests that work in the old code and fail in the new code"

That certainly hasn't been my experience. Recent AI knows exactly what to do in many cases. Writes tests, checks for bugs in it's work and fixes them. By the time I see it everything is down to a few choices I can make.

Are you using recent models? I'm frequently amazed and what they can do with a good prompt and the appropriate context.

"he has given himself four weeks to prove it, with the final goal of rewriting 98% of Cisco's legacy code within 24 months"

My guess is that him and his team will spend the first 4 weeks getting the AI to examine the codebase and whatever documentation may be available. Documentation quickly goes bad so they'll have to compare it against the current implementation. AI will write all of that up in some detailed markdown files. Then they will have AI use that to give them a set of options for the refactoring/rewrite. Seems doable up to that point.

Then they can determine how feasible it all is, where is the low hanging fruit, what incremental steps to take, etc.

"trying to build it with limited access to tokens"

That's the limiting factor these days I agree. You can subscribe for a weekly or monthly allotment of tokens and burn through them in a few days if you aren't careful. Or you can fund a wallet that gets drawn down as your tasks consume tokens and see exactly how much it costs, and keep feeding the beast.

With multiple agents running in parallel this can get expensive real quick.

If its like all the other legacy code I've worked with over the years, it will now be looking like a big ball of paper mache'. New features and fixes were pasted onto it so thickly as time passed that it is brittle, delicate, and opaque.

I bet this guy spends the first month just trying to figure out what's there and how it could be modularized. He will get AI to do this for him of course, and then it will put an implementation plan together for him that it will work from later. There will be an extensive test suite so he can compare the behavior of his new code against the legacy version. Ideally he would phase in pieces of his reimplementation little by little instead of one huge code drop in 2 years.

The article doesn't give any technical details. From what I can tell, the Cisco IOS is written in C language. The original implementation was written in the 80's. I always disliked any situations where I had to interact with it back in the day, maybe it has improved.

By now the codebase is probably a barnacle-encrusted mess. Wouldn't surprise me if they want to just recreate all the existing functionality from scratch in Rust. Can some kid and a couple of his buddies make significant headway in a month with unlimited AI assistance? Could be.

"he has given himself four weeks to prove it, with the final goal of refactoring 98% of Cisco’s legacy code within 24 months", might be able to do that.

Sounds like it is much more than assumptions seeing as how "American officials had concluded that she was a Chinese intelligence asset".

"Anyone can use an LLM."

Yeah but apparently a lot of people suck at it, including several of the trolls here.

"Where is your evidence to support your causal assertions?"

They said it themselves.

"Probably half of the suggestions AI wants to do for me is stupid or else just plain wrong", I'm still waiting to see your work. Maybe it's just you.

"Everyone should provide evidence to support their personal experiences", you made the claim buddy. Got nothing?

"Probably half of the suggestions AI wants to do for me is stupid or else just plain wrong"", let's see your work.

"psychologically manipulative shills spewing nonsense divorced from reality". I'm not trying to sell you anything, asshole.

'how to be an end user of an LLM constitutes some kind of "skill" is frankly absurd on its face'

Hilarious! Go look for a job telling the interviewer that.

"This may be acceptable when quality is not important, as things can get lost in translation." Yeah that's why you look at the proposal. Make some tweaks as desired.

If you can't figure out how to articulate what you want and choose from some clearly described alternative ways to do it, you sure aren't going to do a good job of skipping over all of that and coding it by hand.

"don't moral panic over blowjobs anymore", I guess it would depend on how his wife feels about it rather than what you think.

It smells like the standard honey pot. Pretty girl positioned to catch his eye while he is overseas on business and she strikes it up with him. A few more meets, some dinner engagements, and then she is travelling around with him. Introducing him to business opportunities and laying on the flattery. If the relationship ever gets into blowjob territory or beyond he is doomed, China will have him on a blackmail hook he can never shake.

"the code is shit, fuck it, I am doing it myself"

If you've learned how to use AI tools properly you don't start out by having it write any code. You tell it what you want to accomplish in as much detail as you are comfortable with, and tell it to create a proposal for you to review. This is the most important step, and anyone who skips it should be rejected. A good model will propose several options and describe the pros and cons.

You make some careful decisions and tell it to generate an implementation plan. Once you are happy with all of that you can give it to an inexpensive model (and even the cheap ones are pretty good these days) to work from and it will get the job done with minimal handholding.

Or you can look at the implementation plan and do it yourself, but why bother?