devbisme
One of my friends was told recently: “You won’t be beaten by AI. You’ll be beaten by someone using AI.”
Well, no. You’ll be beaten by someone selling AI.
And I don’t know why people keep wasting time with this ChatGPT thing, which is not even “AI” and is only language generation, for anything engineering. It’s not made for that, even OpenAI have put very clear disclaimers about its ouput not to be trusted for any particular purpose, and people still try and still expect. That’s almost the definition of craziness.
Although I heard that one of my former colleagues who is not retired uses a licensed version in the company he works in, training it on very specific datasets and they have good results.
But when the public model is trained with the Internet, which we know is full of junk, then you can expect junkout sometimes. Already it seems some junk has been reingested by LLM models, thus amplifying the junkitude.
I have a different opinion. No use debating it. In five years, we’ll probably have a definite answer as to whether it’s valuable or not.
(Note: Back in 94/95, I didn’t think the Internet would amount to all that much. We tend to overestimate how much effect something should have in the short term, and then get completely fooled by how much effect it has in the long term due to compounding.)
sometimes. but most of the times it is the exactly opposite: something gets hyped up way more then it should, can’t live up to any expectations and in the meantime does more harm then good. See Blockchain-tech or the Dot-Com bubble.
That’s why I stick with judging the status quo and on this base public available AI is not suitable to support engineering work. In the special purpose fields with very constrain training it can sometimes work, but I still would argue that the efforts put into training these networks and sorting out their hallucinations amounts to more then having a human do the task from the beginning.
I can imagine this same report being written in the 1980’s: “Compiler-generated code much worse than hand-assembly!” But I doubt anyone today wants to let go of their compilers.
The report makes a big deal about an increase in “code churn” which is indicated by code that has been committed and then modified within two weeks, implying that the committed code was somehow “wrong”. But my experience shows a different aspect of this. During development of RepoRecon, I initially wrote a shell script to scan Github repositories. Eventually, that script became unwieldly, so I said “Copilot, re-write this in Python.” Within 30 seconds, I had a Python version. It didn’t work, but it was 80% correct. I picked through it and fixed the errors. Those corrections would have been classified as “churn”. However, I was able to convert the shell script to working Python in about an hour with Copilot, versus a day if I had been working manually. I’ll take some churn if it saves me seven hours.
The report also talks about how AI-generated code tends to repeat blocks of code rather than create re-usable components. This is true: the AI still doesn’t have a global sense of how an application is put together which would allow it to look for functionality that could be used in multiple places. (Hooray! Humans are still good for something.) But these AI tools have only been in common use for about a year. Expect their capabilities to expand. As Ethan Mollick says: “The AI you’re using today is the worst one you’ll ever use.”
An escape from daily technology. Of course, half the time I am operating using AS (artificial stupidity), and the shelf over my lathe is littered with failed pieces, each of which elicited a naughty word when my tool caught it in the wrong spot. Sorry mods – I’m done with the OT tangent
When a technology becomes commonplace, it stops getting labelled AI. When AI can competently write programs and design schematics, savvy humans will just move up the food chain. I mean, when’s the last time you pulled out a road map and planned a route? You just bring up the map app on your phone or car dashboard.
Thanks! The stripey wenge looks great unfinished with just a green-scratch-pad burnish, as pretty much any finish just turns wenge dark and loses the figure. The others are maple, walnut, and juniper, which I will give an oil-varnish finish and a coat of wax. Still kinda new to it.
