Discussion
trouble with both the Microsoft and FAA fuckups is that both are fully explicable by normal idiot management decision making around IT
so before we can state in an article/episode that either was caused more or less directly by the AI, and not standard pre-AI incompetence, we need more direct evidence
it's most frustrating
Did you see the presser yesterday where they blamed the old software that still worked for not being able to handle the data that the new software was hurling at it, and blamed the Congress (100% controlled by their coterie) for not giving them $50 billion to do it right. That's like blaming vandalism when you learn the dumb way that a reflecting pool is not a parade route. They scoffed at the idea that one could expect old systems to work smoothly with new systems. That's exactly the plan that they tried to get away with. Fortune favors the bold, but so do the Darwin Awards, the National Enquirer and the Police Gazette.
@davidgerard I don't know about Microsoft, but the FAA failure sounds like bog standard systems rollout by people who have no business doing a systems rollout
Even if you take their words at face value and don't read the subtext- they rolled out a system without doing proper testing of any kind and apparently didn't do a phased rollout. For a life or death system? Don't let these people near a computer ever again
@Jer yep, is it a duck or is it a duck with prion disease?
@davidgerard I also will bet that the new systems make no attempts to be efficient and they put systems coded as modern web based systems on networks built to handle 1980s level network code.
When they say the old systems were "overwhelmed" by the new systems, bloated modern network code immediately came to mind
And one thing ai agents love to produce is network code that uses far more bandwidth than it should need. Because that's the shit it has been trained on
@davidgerard @MisuseCase they are one and the same.
The choice to use untested and generally unrestricted AI is reflective of human decision making.
Any direct evidence will be laundered under a human name, anyway.
But I think for Microsoft we can confidently say it's at least indirectly responsible, if not for the fuckup happening in the first place, then for the time-to-diagnose, which was uncharacteristically long for Microsoft.
RE: https://bsky.app/profile/did:plc:5aeagkpy2ch5eencyi2elahk/post/3n2bfkuytgk2k
@boxobark.ing god tnose incident updates, slop generated stories of spinning the gacha for a solution
I opened a support ticket for Azure (unrelated to M365 outage) about a preview feature not enabling. After the second AI slop repeating back everything I wrote, and hallucinating how you even enable preview features, I told the bot to get me a person who knows the system.
Two days later I finally find the relevant Learn article on my own, and fix the issue (wrong feature selected due to crappy documentation). The bot never responded.
I think the real story here is not that it happened, but how their incident response flailed around as if they don't even know their own infrastructure anymore. That's a cultural shift that can be pinned on dogmatic "AI" adoption even if you can't point directly to a PR from Claude as the trigger.
@boxobark.ing @davidgerard
OK, so… MIcrosoft have always had a soft spot in their own understanding of their cloud infrastructure. I was a CTO of a company that was using Azure to provide data to a fleet of devices, and one time we rolled out an update that could;t read the result of a query from Azure, so it would send the request again. We only rolled out the update to about 1% of the fleet before the load on the SSH authentication servers overwhelmed the entire Azure west European infrastructure because crypto is expensive.
Anyway the important thing was that you couldn’t see the issue on any of the customer facing dashboards - if we looked on the client side we could see more than 90% of requests failing but on the Azure side all we could see was lower than expected traffic. When we reported the problem to MS (initially as one endpoint having an issue) it took them many many hours to realise what the actual issue was taking down their entizre network. And they only figured it out because we did some investigating on our side to see exactly where in the connection things were failing. They couldn’t even see their entire service was down (for all customers, not just us) - though they finally twigged when I explained it wasn’t just the one endpoint, but all of them.
Anyway, this situation looks very similar in its response shape, but mine happened 8-9 years ago, so I wouldn’t be too quick to point the finger at AI here.
@davidgerard much as it would tickle my schadenfreude for it to be the fault of AI, I've been one of these f#$!wit managers this thread is giving a good kicking to and sometimes, shit just breaks.
Health checks between devices fail, so they don't fall over, someone types a '1' instead of a '2' in the change, the validation test didn't cover the edge case, etc, etc
@davidgerard you are definitely correct. However, the "we are jamming generated code into everything" people and the "buffoonish fuckwits who fire 1000s to juice the stock price" people ARE THE SAME PEOPLE.
They are responsible, and we can remind everyone of that. The mechanism of failure is downstream, and I am arguing, irrelevant.
@davidgerard Journalists should’ve asked the FAA what caused the “data overload” and what guarantees there are that this won’t happen again? Blaming old equipment isn’t enough of an explanation.
@davidgerard “never attribute to artificial intelligence that which can be adequately explained by natural stupidity”
@davidgerard capitalism
It's capitalism
AI is fine when used properly
Everything this week is capitalism problems
@jpaskaruk that third line turns out to be a near empty set however
@davidgerard Hanlons Razor for me. I can absolutely see someone doing an update on a Friday that took it down. Whether they couldn’t be arsed to check properly or were pushed to do it is just extra detail.
@davidgerard AI enables more idiot managers to fuck up way more things much faster, what's not to like 🤡
@davidgerard I think MS is more about the cloud compute cartel that the hyperscalers run. They don’t need to be good so they arent
I don't feel like it's my responsibility to tell anyone "perhaps this wasn't a direct failure of AI, it's more complex than that"
When those who have foisted AI into everything have done less than nothing to educate people about what this technology really is, what it can and can't do.
If the general public thinks "they put AI in it and then it broke" well then. How is that different from "they put AI in it and now it's amazing" ?
@futurebird @davidgerard yes. Don't ascribe to malice that which can be explained by incompetence. It works both ways round. And also if you substitute AI for incompetence.
@futurebird yeah, but it sort of is my responsibility to do that! i need to be precise and i need more facts for that
@davidgerard @futurebird 💯 I agree that people in general have no obligation to give these clowns any benefit of the doubt, I still want good journalists to preserve their credibility and dig for the facts. Your poll the other day about covering Humble Bundle vs MS was a good example of that.
@glecharles @futurebird also i'd never hear the end of it from the nitpickers lol
@futurebird It's an explanation I want to be true, therefore it is true. /s 🤪 @davidgerard
@thejessiekirk @futurebird @davidgerard@circumstances.run
Yes, why saving them from getting hoist by their own petard? They kept "putting AI into" everything, now they're blaming AI. It actually fits my agenda nicely, so who am I to dig them out of the hole they dug for themselves?
@thejessiekirk @futurebird so many times i see the *best* headline for a Pivot story and I look and it isn't true/solid
There is also a big incentive to hide any evidence that this is the case if true.
So, in a way we can only look at the wider picture. Have there been more software failures as of late? Have they happened in ways that we might not normally expect?
And even if there was evidence that it was vibe-coded software that failed some would still say that it was just "not prompted correctly" or "not reviewed enough" etc. so in the end it's always human error.
But that's the point.
@davidgerard @futurebird Yes, in the case of the FAA, admitting it was AI would make them look dumb. 800 million dollars dumb
@davidgerard I think we have a deep national shift in risk management philosophy. Historically, we view infrastructure risk like NASA does. Put a ton of research in up front to understand all variables, and manage redundancies in hopes to avoid catastrophic failure.
We're now in the SpaceX(Musk) Iterative Development era with little or no redundancies. Failure is expected. Manage fallout. The "You have to break a few yokes" philosophy.
Its deeply wreckless and callus.