Anybody who has worked in IT support in the trenches in enterprise IT will tell you there are some Excel power users who basically run the company, are macros wizards and actual ninjas.. about 0.1% of the workforce.
-
RE: https://mastodon.social/@404mediaco/116908074107231828
Anybody who has worked in IT support in the trenches in enterprise IT will tell you there are some Excel power users who basically run the company, are macros wizards and actual ninjas.. about 0.1% of the workforce. About 99% of people can’t align a table in Word.
Giving the 99% of people tools which cost $$$ per user a month and letting them do anything is like giving a child a car, and being surprised when they ram the car into a wall three days later and cost $10k after achieving nothing.
@GossiTheDog
One thing that has occurred to me is how much people use it for search, where search is (still, if used correctly) broadly factual and free at the point of use, and the big GenAI providers tend to provide affirmatory answers while burning through compute. -
@brnrd @GossiTheDog There is a ‘certain amount of waste’ built into the modern way of writing, testing and deploying software.
That amount can be totally insane when changing a minor detail. If the pipeline is not able to isolate what to build, test and deploy it will just do it all. Is it necessary? No. Is it able to prevent mistakes? Maybe.
@jeroenvanbergen @brnrd @GossiTheDog So CI/CD pipelines can't even do what 'make' did back in the day?
-
This isn’t, btw, a Kevin Doesn’t Use AI rant. I use GenAI extensively for various testing things. I had access to Mythos before almost everybody. I pay for Copilot Pro+. I pay for Gemini. I AI generate terrible songs. I vibecode security scanners. My personal spend exceeds £500 this month to date.
Do I think you should make every employee depend on these third party GenAI tools for their job? No. Its ridiculous. You’re also *locked in* to costs you can’t afford which *will rise*. It’s a cliff.
I love this because it’s not a binary AI=good || AI=bad post.
What we have done is reset back to the “early GitHub days” where everybody was whipping up projects, regardless of skill.
AI without knowledge is bad. It’s like a firearm in that sense.
-
RE: https://mastodon.social/@404mediaco/116908074107231828
Anybody who has worked in IT support in the trenches in enterprise IT will tell you there are some Excel power users who basically run the company, are macros wizards and actual ninjas.. about 0.1% of the workforce. About 99% of people can’t align a table in Word.
Giving the 99% of people tools which cost $$$ per user a month and letting them do anything is like giving a child a car, and being surprised when they ram the car into a wall three days later and cost $10k after achieving nothing.
@GossiTheDog Early in my IT career, I had to do a Help Desk call to the business department (at a school district). They were in a panic because they had a shared Excel spreadsheet that was mission critical, but the person who made it had quit. They didn't know how it worked and things were no longer adding up. I shouldn't have, but I looked at the sheet and figured out where the error was. Over the next several weeks came in help desk tickets from this department needing help modifying the sheet and expanding it. I refused to do any of them and had to have a call with their department manager explaining that I was not there to teach them how to use Excel and he agreed they shouldn't be using IT people to do the business calculations they claimed they knew how to do.
I can't imagine what it must be like if those people are using AI.
-
In an era where companies need to become more efficient and diverse they’ve basically picked the least efficient way to do it, with the biggest risks and highest costs - because everybody else is doing it.
I know somebody at one of the big 4 who has written something in Claude that prompts Claude each twenty minutes for a question, then feeds Claude’s question back into Claude to use their tokens - because token usage is factored into employee evaluations. What are we even doing.
@GossiTheDog CEOs and board members were sold the scam that LLMs would magically turn into AI someday and that it would, in turn, successfully replace all human employees (except them of course!) They love the idea of eliminating humans — or at least eliminating paying them.
I think, though, the most impressive thing is most of the scammers actually are sniffing their own fumes and believe the same ridiculousness. They actually don't understand the nature of their own product... Which... I guess isn't so unusual for people like these, but darn it's a heck of a crazy thing to watch happen. People who know nothing selling to people who know nothing. And all either side truly needs to do is just frigging ask someone who understands the tech. It can be useful, but not like this.
-
@bontchev @GossiTheDog Why would your hypothetical employee have become annoyed with you?
@xan @GossiTheDog Because I'm an annoying person. Pedantic, insisting on excellence, working long and odd hours, with a sarcastic sense of humor, etc.
-
@bontchev Claude with effectively unlimited usage isn't going to stay $180/year for long. When they do usage based billing like everyone else has been forced to do it's really easy to have a 2 sentence input that costs $100
(It still might be cheaper than hiring a developer but it is still very expensive)
@malwareminigun You might be right but I'm talking about the right now, not about the future. And even before the boss bought me the Pro subscription, I managed to do quite a lot of work with the free version, which offers like 9 times less tokens per 5-hour session.
-
@malwareminigun You might be right but I'm talking about the right now, not about the future. And even before the boss bought me the Pro subscription, I managed to do quite a lot of work with the free version, which offers like 9 times less tokens per 5-hour session.
@bontchev the future in question is months not years away. It's already here for pretty much every other provider. The ~$1 Trillion that was spent on hardware to power this stuff isn't getting paid back at $200/user/year
-
@bontchev the future in question is months not years away. It's already here for pretty much every other provider. The ~$1 Trillion that was spent on hardware to power this stuff isn't getting paid back at $200/user/year
@malwareminigun That might be so but I was talking about the *now*. And now, Claude is perfectly affordable for what it can do. When it becomes too expensive, I'll stop using it.
-
@malwareminigun That might be so but I was talking about the *now*. And now, Claude is perfectly affordable for what it can do. When it becomes too expensive, I'll stop using it.
@bontchev you're missing the point. This thread started with people talking about how AI was crazy expensive, and you saying, in my experience, it's cheap. The reason it's cheap for you right now is you're using a setup that subsidizes the crap out of it. Anyone using a path that doesn't have such subsidy is currently paying for it and it's not cheap. Those folks are paying those costs right now.
I'm not saying it's valueless, or don't use it, or you should change anything about what you are doing right now, i'm explaining why people are saying it's expensive. The employee at the beginning of this thread is easily costing a $1000 a day, just in AI usage.
-
@GossiTheDog PS: we did raise these concerns when management announced everyone was getting Copilot. Took months (and months) to get them to agree to a test. “No one will be able to see anything they don’t already have access to,” they said.
️Yes. Exactly.
@Spartan_1986 @GossiTheDog A lot easier to *find* the stuff that you have access to but shouldn’t.
-
@GossiTheDog And then there is the non-monetary costs. We recently completed our first penetration test against Copilot in my corp and to say we found a lot of secret and confidential stuff out there just for the prompting is an understatement. The company totally believed Microsoft when they said everything would be safe guarded, yet I personally found a document with every marketing service account name and password. Vender contracts, company secrets, legal documents: we found it all. Copilot in a corporate environment is the single largest amplifier of poor IAM configurations. It is the largest insider threat I’ve ever seen.
Parameters of the test?
And as it in the compartmentalised corpo environment or public?
-
In an era where companies need to become more efficient and diverse they’ve basically picked the least efficient way to do it, with the biggest risks and highest costs - because everybody else is doing it.
I know somebody at one of the big 4 who has written something in Claude that prompts Claude each twenty minutes for a question, then feeds Claude’s question back into Claude to use their tokens - because token usage is factored into employee evaluations. What are we even doing.
@GossiTheDog That sounds like reinforcement training to me. /s
-
@GossiTheDog And then there is the non-monetary costs. We recently completed our first penetration test against Copilot in my corp and to say we found a lot of secret and confidential stuff out there just for the prompting is an understatement. The company totally believed Microsoft when they said everything would be safe guarded, yet I personally found a document with every marketing service account name and password. Vender contracts, company secrets, legal documents: we found it all. Copilot in a corporate environment is the single largest amplifier of poor IAM configurations. It is the largest insider threat I’ve ever seen.
@Spartan_1986 @GossiTheDog The obscurity is gone
-
In an era where companies need to become more efficient and diverse they’ve basically picked the least efficient way to do it, with the biggest risks and highest costs - because everybody else is doing it.
I know somebody at one of the big 4 who has written something in Claude that prompts Claude each twenty minutes for a question, then feeds Claude’s question back into Claude to use their tokens - because token usage is factored into employee evaluations. What are we even doing.
@GossiTheDog The security risk surrounding all these users with access to LLM's with access to all that corporate data is chilling.
-
@malwareminigun @GossiTheDog No attribution necessary, it's AI generated.
@hans5524 @malwareminigun @GossiTheDog Yep it is, this exact piece made the rounds a few weeks ago. There was a very thorough thread where every little AI tell was poured over in detail.
-
In an era where companies need to become more efficient and diverse they’ve basically picked the least efficient way to do it, with the biggest risks and highest costs - because everybody else is doing it.
I know somebody at one of the big 4 who has written something in Claude that prompts Claude each twenty minutes for a question, then feeds Claude’s question back into Claude to use their tokens - because token usage is factored into employee evaluations. What are we even doing.
@GossiTheDog there’s an old tale about optimizing for the wrong things. Something about getting rid of rats only made a market for breeding them.
Guess that means we’ll burn the planet and use up all the drinkable water to autocomplete our homework and live life without having to deal with people.
-
In an era where companies need to become more efficient and diverse they’ve basically picked the least efficient way to do it, with the biggest risks and highest costs - because everybody else is doing it.
I know somebody at one of the big 4 who has written something in Claude that prompts Claude each twenty minutes for a question, then feeds Claude’s question back into Claude to use their tokens - because token usage is factored into employee evaluations. What are we even doing.
@GossiTheDog @paninid That's actually the smartest use of AI (given the constraints imposed) that I've heard about so far
️ -
@bontchev you're missing the point. This thread started with people talking about how AI was crazy expensive, and you saying, in my experience, it's cheap. The reason it's cheap for you right now is you're using a setup that subsidizes the crap out of it. Anyone using a path that doesn't have such subsidy is currently paying for it and it's not cheap. Those folks are paying those costs right now.
I'm not saying it's valueless, or don't use it, or you should change anything about what you are doing right now, i'm explaining why people are saying it's expensive. The employee at the beginning of this thread is easily costing a $1000 a day, just in AI usage.
@malwareminigun I don't dispute that it's cheap right now because it is subsidized - but that's besides the point. The point is that it *is* (or at least can be, as demonstrated in my case) very cheap right now compared to the value it brings.
Yes, that's likely to change. When it changes, I'll revaluate. I am not arguing that AI won't become too expensive to use. I'm arguing that AI can be quite cheap to use (while bringing tremendous value) *right now* - for whatever reasons.
-
RE: https://mastodon.social/@404mediaco/116908074107231828
Anybody who has worked in IT support in the trenches in enterprise IT will tell you there are some Excel power users who basically run the company, are macros wizards and actual ninjas.. about 0.1% of the workforce. About 99% of people can’t align a table in Word.
Giving the 99% of people tools which cost $$$ per user a month and letting them do anything is like giving a child a car, and being surprised when they ram the car into a wall three days later and cost $10k after achieving nothing.
@GossiTheDog
And the tech firms didn't even only give cars to their children but explicitly told them to drive as much and as fast as they possibly could 🤯
