I feel like the AI industry did this to themselves by absolutely shoving it down our throats at every possible chance they had.
AI in general isn’t a bad technology, it just has very limited use cases where it’s actually good at things. Most things it’s used for are things it’s bad at. Kind of like using a steam locomotive to clean the bottom of your pool.
Exactly. And the few things GenAI does well aren’t that compute intensive and can run on local models.
LLMs are terrible at writing new text that matters, i.e. anything at all technical or important. But they’re good at reformatting content into properly structured English, editing text for grammar, and explaining everyday concepts. But they’re so good at those things, that a tiny model running on consumer hardware can do it.
So, I don’t understand how anyone can possibly expect LLMs to be a profitable business. Where are the moats?
In the early days of the steam engine’s success, what helped it find an economic footing was the need to pump water from coal mines. At the time, logistics of transporting things over large distances were difficult, so having the coal supply on-site was an obvious benefit. So while using a steam engine to pump water from a pool would be quite archaic and inefficient, but at least for emptying the pool, it could technically be used.
Of course, a locomotive moving the engine around would be entirely impractical for pumping a somewhat steady stream. Chlorinated water would also make for poor steam, I imagine.
I guess driving a scrubber back and forth would be possible, but I could do that more efficiently by hand. I drink less water, eat less coal, don’t require rails and skilled operators and I’m easier to transport to the site. I’m not sure how the carbon footprint for a human compares to that of extracting and processing iron, but it won’t be favourable if at least one human is involved.
The scrubber couldn’t be as large and heavy though. If you need a heavy scrubber, you’d also need a powerful motivator. That’s where a steam engine might have an edge over me: sheer power output.
To match that, you’d have to employ and pay multiple people. So if the steam engine’s rent seems cheaper than that…
“See the naming conventions at the beginning of the script. Apply all the rules to all the non-conforming variable names and the new custom function we made earlier.“
There were like 100 changes here and there, and I just didn’t feel like doing all that work manually.
In my experience, the trick to this is to get a list of changes to make yourself, rather than trusting the machine to actually get it right. (Unless you’re in a dev/test environment and don’t care much if it fucks it up.)
This code isn’t anywhere near production yet. I’m still going to make a hundred little changes, but an LLM can do all the boring stuff so that I can focus on the interesting bits. There’s also going to be plenty of testing and iterative tweaks, so any major mistakes will be exposed sooner rather than later. Honestly, about half of the errors will be caused by me.
I can count on one hand the number of jobs or functions I’ve seen people try to get AI to do that isn’t already done better by a hard-coded program or an Excel spreadsheet.
I’ve used it as a sort of find and replace in the past, but by feeding it the output from Ripgrep to skip the “find” part. The “replace” was just unwrapping unnecessarily confusing “try” blocks and allowing errors to propagate to the appropriate handlers. This would have required a syntax-aware replacement tool (and some exist but I didn’t really feel like learning one for a one-off).
As a simple “find this text and replace with this other text” it makes no sense.
Kind of like using a steam locomotive to clean the bottom of your pool.
I’m surprised that wasn’t tried on the reflecting pool in D.C. “We love the old-timey trains, don’t we, folks? And now were going to use one in the most amazing way the world has ever seen!”
Powered by “clean coal” of course, right? It’d have been more effective than dumping a couple bottles of hydrogen peroxide into the pool at least since the emissions would prevent future algae growth more effectively.
Oh, they did dump some H202 along the edges, giving a dark blue hue surrounding the algae bloom. Apparently, a comprehensive distribution either didn’t come to mind or was nixed. Alternatively, clearing it up immediately would negate Trump’s claim that the green hue was the result of vandalism and thus, well …
There are lists of technology which will work for AI such as programmer. which require analytics of AI and work with that frame of mind.
We were shown the list in my school 15 years ago and told “some day soon AI will take these jobs, so choose wisely” with a rough percentage for each job.
Okay, let AI take my job then, that would happen anyway. Jobs have been gouged ever since society has mutated apart from its basic building blocks. I at least want peace of mind!
I feel like the AI industry did this to themselves by absolutely shoving it down our throats at every possible chance they had.
AI in general isn’t a bad technology, it just has very limited use cases where it’s actually good at things. Most things it’s used for are things it’s bad at. Kind of like using a steam locomotive to clean the bottom of your pool.
Exactly. And the few things GenAI does well aren’t that compute intensive and can run on local models.
LLMs are terrible at writing new text that matters, i.e. anything at all technical or important. But they’re good at reformatting content into properly structured English, editing text for grammar, and explaining everyday concepts. But they’re so good at those things, that a tiny model running on consumer hardware can do it.
So, I don’t understand how anyone can possibly expect LLMs to be a profitable business. Where are the moats?
In the early days of the steam engine’s success, what helped it find an economic footing was the need to pump water from coal mines. At the time, logistics of transporting things over large distances were difficult, so having the coal supply on-site was an obvious benefit. So while using a steam engine to pump water from a pool would be quite archaic and inefficient, but at least for emptying the pool, it could technically be used.
Of course, a locomotive moving the engine around would be entirely impractical for pumping a somewhat steady stream. Chlorinated water would also make for poor steam, I imagine.
I guess driving a scrubber back and forth would be possible, but I could do that more efficiently by hand. I drink less water, eat less coal, don’t require rails and skilled operators and I’m easier to transport to the site. I’m not sure how the carbon footprint for a human compares to that of extracting and processing iron, but it won’t be favourable if at least one human is involved.
The scrubber couldn’t be as large and heavy though. If you need a heavy scrubber, you’d also need a powerful motivator. That’s where a steam engine might have an edge over me: sheer power output.
To match that, you’d have to employ and pay multiple people. So if the steam engine’s rent seems cheaper than that…
I’ve seen people literally use it as find and replace 🤦♂️
I did that today. It went something like this:
“See the naming conventions at the beginning of the script. Apply all the rules to all the non-conforming variable names and the new custom function we made earlier.“
There were like 100 changes here and there, and I just didn’t feel like doing all that work manually.
In my experience, the trick to this is to get a list of changes to make yourself, rather than trusting the machine to actually get it right. (Unless you’re in a dev/test environment and don’t care much if it fucks it up.)
This code isn’t anywhere near production yet. I’m still going to make a hundred little changes, but an LLM can do all the boring stuff so that I can focus on the interesting bits. There’s also going to be plenty of testing and iterative tweaks, so any major mistakes will be exposed sooner rather than later. Honestly, about half of the errors will be caused by me.
I can count on one hand the number of jobs or functions I’ve seen people try to get AI to do that isn’t already done better by a hard-coded program or an Excel spreadsheet.
I’ve used it as a sort of find and replace in the past, but by feeding it the output from Ripgrep to skip the “find” part. The “replace” was just unwrapping unnecessarily confusing “try” blocks and allowing errors to propagate to the appropriate handlers. This would have required a syntax-aware replacement tool (and some exist but I didn’t really feel like learning one for a one-off).
As a simple “find this text and replace with this other text” it makes no sense.
I’m surprised that wasn’t tried on the reflecting pool in D.C. “We love the old-timey trains, don’t we, folks? And now were going to use one in the most amazing way the world has ever seen!”
Powered by “clean coal” of course, right? It’d have been more effective than dumping a couple bottles of hydrogen peroxide into the pool at least since the emissions would prevent future algae growth more effectively.
Oh, they did dump some H202 along the edges, giving a dark blue hue surrounding the algae bloom. Apparently, a comprehensive distribution either didn’t come to mind or was nixed. Alternatively, clearing it up immediately would negate Trump’s claim that the green hue was the result of vandalism and thus, well …
There are lists of technology which will work for AI such as programmer. which require analytics of AI and work with that frame of mind.
We were shown the list in my school 15 years ago and told “some day soon AI will take these jobs, so choose wisely” with a rough percentage for each job.
Okay, let AI take my job then, that would happen anyway. Jobs have been gouged ever since society has mutated apart from its basic building blocks. I at least want peace of mind!