Skip to main content


Talked to a software engineer at Microsoft working on Copilot Studio today at a social event and he said he was ashamed that he hadn’t written a single line of code in over three months. “I used to take pride in my work.” (They simply create plans in natural language and feed it to the LLM which generates the code. They can’t even do human code reviews anymore as there’s too much code being generated.)

He said a lot of them were waiting for a catastrophic event (something that would take down critical infrastructure) to get top management to reverse course. He seemed to think such a failure was very likely.

Given what we’ve been seeing recently, I tend to agree with him. Although I feel they will just double down. There’s too much money in the pot for them to fold.

#AI #microsoft #LLMs

This entry was edited (1 day ago)

reshared this

in reply to Aral Balkan

I heard a talk from someone, who said something similar, some months back. I'm worried

infosec.exchange/@webhat/11557…

Aral Balkan reshared this.

in reply to Aral Balkan

Could you ask, why in their opinion no catastrophic event has happened yet? Did their overall workload increase?
in reply to dkl

@dkl May's Patch Tuesday addressed 120 separate vulnerabilities, including 17 classified as critical. GitHub's uptime is now zero nines, and they just had 3,800 internal repositories hacked. For a lot of businesses, those would be catastrophic events, but long term Microsoft customers are used to poor security and unreliability.

mrshu.github.io/github-statuse…

@dkl
in reply to mathew

@mathew @dkl This.

By “catastrophic” he meant something that causes people to die, etc. (Medical systems, etc.)

in reply to Aral Balkan

@mathew @dkl No, something that costs Microsoft money. Something like 365 or Azure being unavailable for a couple of days or weeks.
in reply to Aral Balkan

Something TPTB should keep in mind, but I think they are closely examining subterranean sand instead.
in reply to Aral Balkan

"They can’t even do human code reviews anymore as there’s too much code being generated" this is the ticking time bomb. I've been saying for a long time that writing code is easy; it's owning code that's hard. Designing an application to change and grow as future requirements arrive takes deliberate design thought. LLMs are not capable of that.
in reply to Sam Stephens

I think what we're going to see is that LLM codebases will end up at a point where the LLM can no longer successfully modify them, and are complex and convoluted enough to defy human understanding (with the exception of people who are being careful about the code structure, but you cannot do that, and simultaneously generate too much code to review). A similar danger to that of humans writing code without architectural vision, but amplified by the volume of code an LLM generates.
in reply to Aral Balkan

i&i love this one:
"They can’t even do human code reviews anymore as there’s too much code being generated."
in reply to Aral Balkan

in my opinion, the subliminal steering stuff (check arxiv) is ready to happen. This gist is that a user discusses, to recapitulate the plot of the manchurian candidate committing an assassination when shown a trigger with the slopbot. Then the sloperator asks the bot for some code. Even though the code has no semantic connection to political assassinations, when another bot in the same family sees the code, it picks up the instruction (e.g. the political assassination codeword).
This entry was edited (23 hours ago)
in reply to Aral Balkan

Copilot's going to end up on par with bing if they're not more careful.
MS still have pool tables...? Seems like a good LLM-proof career.
in reply to Aral Balkan

I'm forced to use M$ at work. This is just anecdotal but it's getting slower and buggier, lots of people have been complaining. It's certainly not getting amazingly great.
in reply to Aral Balkan

Well either that, or it becoming more expensive than to hire a human programmer.

However one needs to take into account that many people live in a bubble of "OK-ish software". Outside of it there are companies like Atlassian who have products, created by humans, which could be much improved by getting them re-written by AI. There's just so much terrible software out there already.

in reply to Aral Balkan

Betting on disaster to stop them is an illusion; the capital and systems that have tasted the machine's efficiency in erasure and profit will not back down, but will treat victims and software errors as an "acceptable cost" of dominance. When human skill and responsibility fall, humanity falls first💔😔🇵🇸🇵🇸✌️
in reply to Aral Balkan

@glynmoody Yes well cue management that thinks it knows better what to do followed by knowing it better how to do it. Tic tic tic tic tic...
in reply to Aral Balkan

And then they will demand government bail-outs to compensate them for their catastrophic losses.

And will get them.

in reply to Aral Balkan

In a minor aside, I was forced to use Copi-lot the other day to change a date field in an online Word document. No other way
in reply to Aral Balkan

OK, so here's another little bit that fits the pattern.

We're due a catastrophe. The mad king is supposed to utterly ruin our ability to respond to emergency. This seems relatively accomplished. Next step is to cause a massive crisis that topples the last of the old republic.

Synchs with the "data centers" that are really just large sections of land secured for corporate.

They're investing all our 401ks in it through SpaceX. So...yeah. You are correct.

in reply to Aral Balkan

I’m sure they could find a fungible human who’s down in their token-burning quota to be a high-tech whipping boy to fire.
in reply to Aral Balkan

Problem is: the very people who need to say that all these billions of dollars were wasted are the very people whose necks are on the board of directors' chopping block. The likes of Nadella, the CEOs and such, are the one's who'd get axed for wasting so much money but they're also the ones who get to decide whether or not cut losses or double down.
in reply to Aral Balkan

It's always a bad sign when people on the ground are confirming my suspicions about the state of anything. This is no exception.