I keep watching for the AI bubble to burst.
-
I keep watching for the AI bubble to burst. It's been so expensive for so little returns.
At work, a senior leader asked me to present on AI. As a manager, I'm expected to adopt it, whether I can use it or not.
Our usage is tracked, so I've had to do some stuff with it. I tried setting up a complex process, since the easy things aren't really time-savers. I presented on this testing, even though I doubt I'm actually going to use it.
Every day, the AI tools seem to interpret the exact same prompts differently. So when I validate the output, it has changed.
How can I add an AI process into my workflow when the output is different every day? How much extra time am I wasting in validating the output? And don't tell me that my prompt is bad. It works until suddenly it doesn't. The tool is bad. Tools shouldn't do that.
I can write programming logic to do the same things. I know the code isn't going to give me different output every time I open it.
The senior leaders are desperate to find use cases for AI. The companies have invested a lot of money and want to see returns, but it's not happening. When are they going to cut their losses?
In general, machine learning is good for situations where the value of a correct answer is significantly higher than the cost of an incorrect one. Prefetching is my favourite example (not least because that’s what I used it for in my PhD): if you prefetch the right data, the processor avoids a long idle time. If you prefetch the wrong data, you discard it and you’re not in a worse position than if you hadn’t used it.
Machine translation is a great case study for this: for any given situation, is an incorrect translation worse than no translation? For translating a toot, the worst-case cost is that I am briefly confused and move on. I don’t have time to learn every human language and I don’t have the money to pay someone to translate every toot I see, so machine translation is better than nothing. For translating a user manual or a contract, the worst-case cost may be someone dies or I am exposed to unbounded legal liability (or both), so machine translation is worse than nothing.
The corollary is that machine learning is also useful if the cost of generating an optimal solution is high and it’s possible to mechanically generated solutions. Stochastically generating plausible solution-shapes things then checking if they actually are solutions and picking the one that works best with a cheap-to-run machine-learning system providing the inputs is great. Vulnerability discovery with guided fuzzing works like this, for example.
The big issue with LLMs is that, generally, text is something we use for high-stakes problems. The set of things where the desired output is a pile of text, where the cost of incorrect text is low or the cost of checking the text is low, is very small. This includes programming: if you have a sufficiently detailed spec that you can mechanically check whether an implementation satisfies it, you almost certainly have something that can be used to synthesise a program already.
Some cases, such as searching documentation, are right on the cusp. Is it better to have a machine that answers a question about a system but gives you a right answer 60% of the time and plausible nonsense 40% of the time than nothing? What if it’s 90%? At some point, the Paradox of Automation kicks in: system efficiency does not correlate with component efficiency. You get a significant dip when the system is good enough that people stop checking the answers but it’s still producing a lot of mistakes, only now those make it further along the process.
-
I keep watching for the AI bubble to burst. It's been so expensive for so little returns.
At work, a senior leader asked me to present on AI. As a manager, I'm expected to adopt it, whether I can use it or not.
Our usage is tracked, so I've had to do some stuff with it. I tried setting up a complex process, since the easy things aren't really time-savers. I presented on this testing, even though I doubt I'm actually going to use it.
Every day, the AI tools seem to interpret the exact same prompts differently. So when I validate the output, it has changed.
How can I add an AI process into my workflow when the output is different every day? How much extra time am I wasting in validating the output? And don't tell me that my prompt is bad. It works until suddenly it doesn't. The tool is bad. Tools shouldn't do that.
I can write programming logic to do the same things. I know the code isn't going to give me different output every time I open it.
The senior leaders are desperate to find use cases for AI. The companies have invested a lot of money and want to see returns, but it's not happening. When are they going to cut their losses?
@VeeRat yes 'we' are traced on AI usage supposed to be the clue for everything. Company pay company wants return
Nothing is free in this brave new world
-
I keep watching for the AI bubble to burst. It's been so expensive for so little returns.
At work, a senior leader asked me to present on AI. As a manager, I'm expected to adopt it, whether I can use it or not.
Our usage is tracked, so I've had to do some stuff with it. I tried setting up a complex process, since the easy things aren't really time-savers. I presented on this testing, even though I doubt I'm actually going to use it.
Every day, the AI tools seem to interpret the exact same prompts differently. So when I validate the output, it has changed.
How can I add an AI process into my workflow when the output is different every day? How much extra time am I wasting in validating the output? And don't tell me that my prompt is bad. It works until suddenly it doesn't. The tool is bad. Tools shouldn't do that.
I can write programming logic to do the same things. I know the code isn't going to give me different output every time I open it.
The senior leaders are desperate to find use cases for AI. The companies have invested a lot of money and want to see returns, but it's not happening. When are they going to cut their losses?
@VeeRat I told an LLM about my org and role, and asked it for a reasonable budget, team size, how management should respond to my input, etc.
Suddenly I no longer had to use LLMs.
-
I keep watching for the AI bubble to burst. It's been so expensive for so little returns.
At work, a senior leader asked me to present on AI. As a manager, I'm expected to adopt it, whether I can use it or not.
Our usage is tracked, so I've had to do some stuff with it. I tried setting up a complex process, since the easy things aren't really time-savers. I presented on this testing, even though I doubt I'm actually going to use it.
Every day, the AI tools seem to interpret the exact same prompts differently. So when I validate the output, it has changed.
How can I add an AI process into my workflow when the output is different every day? How much extra time am I wasting in validating the output? And don't tell me that my prompt is bad. It works until suddenly it doesn't. The tool is bad. Tools shouldn't do that.
I can write programming logic to do the same things. I know the code isn't going to give me different output every time I open it.
The senior leaders are desperate to find use cases for AI. The companies have invested a lot of money and want to see returns, but it's not happening. When are they going to cut their losses?
@VeeRat Same here, I received daily pressure for using LLM to do my work or find use cases that works. Now I'm keeping a list of every generated garbage I see in my company and try to demonstrate the problem. I have some co-workers that are becoming blind about the quality, when I confront them, they says "That's look fine to me". I think I'm half way to step back and look the ship sink. The number of incidents QoQ seems raising ~15% but nobody care.
-
I keep watching for the AI bubble to burst. It's been so expensive for so little returns.
At work, a senior leader asked me to present on AI. As a manager, I'm expected to adopt it, whether I can use it or not.
Our usage is tracked, so I've had to do some stuff with it. I tried setting up a complex process, since the easy things aren't really time-savers. I presented on this testing, even though I doubt I'm actually going to use it.
Every day, the AI tools seem to interpret the exact same prompts differently. So when I validate the output, it has changed.
How can I add an AI process into my workflow when the output is different every day? How much extra time am I wasting in validating the output? And don't tell me that my prompt is bad. It works until suddenly it doesn't. The tool is bad. Tools shouldn't do that.
I can write programming logic to do the same things. I know the code isn't going to give me different output every time I open it.
The senior leaders are desperate to find use cases for AI. The companies have invested a lot of money and want to see returns, but it's not happening. When are they going to cut their losses?
@VeeRat LLMs are built to be non-deterministic. Therefore, you can only every reliably use this tool if your desired output is also non-deterministic. That's a fundamental thing, AI evangelists often don't understand.
-
@jmcrookston @diazona @VeeRat
Sadly, the one that comes up again and again: Reading and writing email.
️@tbh @jmcrookston @diazona @VeeRat This begs the question: How does this use case look exactly:
– An autonomous agent who replies to your emails automatically? (What could possibly go wrong?)
– A summary and drafting tool? (has to be supervised, supervision costs time, reduces time saved) -
@VeeRat Same here, I received daily pressure for using LLM to do my work or find use cases that works. Now I'm keeping a list of every generated garbage I see in my company and try to demonstrate the problem. I have some co-workers that are becoming blind about the quality, when I confront them, they says "That's look fine to me". I think I'm half way to step back and look the ship sink. The number of incidents QoQ seems raising ~15% but nobody care.
@hugopoi @VeeRat When I point out errors they haven't spotted themselves but should have (if you give me your AI slop to review, at least have the decency to review it yourself first), they usually just get defensive and say "it wasn't me, it was the AI" like it had nothing to do with this. This loss of accountability prevents change.
-
In general, machine learning is good for situations where the value of a correct answer is significantly higher than the cost of an incorrect one. Prefetching is my favourite example (not least because that’s what I used it for in my PhD): if you prefetch the right data, the processor avoids a long idle time. If you prefetch the wrong data, you discard it and you’re not in a worse position than if you hadn’t used it.
Machine translation is a great case study for this: for any given situation, is an incorrect translation worse than no translation? For translating a toot, the worst-case cost is that I am briefly confused and move on. I don’t have time to learn every human language and I don’t have the money to pay someone to translate every toot I see, so machine translation is better than nothing. For translating a user manual or a contract, the worst-case cost may be someone dies or I am exposed to unbounded legal liability (or both), so machine translation is worse than nothing.
The corollary is that machine learning is also useful if the cost of generating an optimal solution is high and it’s possible to mechanically generated solutions. Stochastically generating plausible solution-shapes things then checking if they actually are solutions and picking the one that works best with a cheap-to-run machine-learning system providing the inputs is great. Vulnerability discovery with guided fuzzing works like this, for example.
The big issue with LLMs is that, generally, text is something we use for high-stakes problems. The set of things where the desired output is a pile of text, where the cost of incorrect text is low or the cost of checking the text is low, is very small. This includes programming: if you have a sufficiently detailed spec that you can mechanically check whether an implementation satisfies it, you almost certainly have something that can be used to synthesise a program already.
Some cases, such as searching documentation, are right on the cusp. Is it better to have a machine that answers a question about a system but gives you a right answer 60% of the time and plausible nonsense 40% of the time than nothing? What if it’s 90%? At some point, the Paradox of Automation kicks in: system efficiency does not correlate with component efficiency. You get a significant dip when the system is good enough that people stop checking the answers but it’s still producing a lot of mistakes, only now those make it further along the process.
@david_chisnall @VeeRat "Machine translation is a great case study for this: for any given situation, is an incorrect translation worse than no translation? For translating a toot, the worst-case cost is that I am briefly confused and move on."
That might be the case on a social media platform, but can decide over life and death in other cases where you are not confused but actually believe the answer.
-
@david_chisnall @VeeRat "Machine translation is a great case study for this: for any given situation, is an incorrect translation worse than no translation? For translating a toot, the worst-case cost is that I am briefly confused and move on."
That might be the case on a social media platform, but can decide over life and death in other cases where you are not confused but actually believe the answer.
@david_chisnall @VeeRat "What if it’s 90%?"
Depends on how you handle it:
– You still have to check everything, because there's still a margin for error, yet you get less reward for finding errors than if it was 60%
– You say, 10% error margin is perfect, you just blindly ship it and let your customers find your errors for your. Great, your reputation goes down the drain but you increased your velocity -
@Lightfighter @Uair @VeeRat Devastating for the environment
-
@kauer @smiteri @n_dimension @VeeRat Nowadays the consensus of LLM evangelists seems to be "write a War and Peace length instruction for the LLM to get it perfectly right". When I have finished my War and Peace small enhancement instruction, I might have already coded the enhancement myself, maybe even twice. Don't forget that by doing it myself, I'm already testing this on a basic level (tests would still be needed) or with TDD beyond that and have an understanding of the code.
-
@VeeRat make the AI once generate code with does the job. Then run the code. The code always does the same thing as good or bad than yesterday. Don't ask the AI to come up with code to do the same task than yesterday. Sell it as "a nice method to use less tokens for the similar result". Get a raise for optimizing the bullshit machine.
@assimilateborg @VeeRat Assuming, you do the same job every day.
-
@VeeRat Lately all the Microsoft IDE stuff has flipped into "CoPilot does everything" and replaced working IntelliSense with garbage.
This last week was when I noticed it because SQL Server Management Studio stopped being able to autocomplete an existing table name.
Just days ago I could trust tab completion and suddenly that's dead too.
-
J jwcph@helvede.net shared this topic
