It may later be succeeded by other technologies, but it should be a thing for a period of at least 6 months.
10% of people regularly doing it will serve as an anchor for the definition of "common", but I don't expect to have this clean data when resolving, and will be vaguely estimating based on this.
It may be wrapped in a no-code user interface, as long as there is still a very general coding AI behind it.
As we move toward a future where AI becomes more integrated into daily life, it’s fascinating to think about how even non-programmers might use AI to create small scripts for everyday tasks. Similarly, sports fans are also embracing tech solutions to enhance their experience. For football enthusiasts, platforms like Stream East offer live streams of matches, making it easy to catch your favorite games from anywhere. While AI might shape our work in 2033, platforms like Stream East are already transforming how we enjoy sports today.
The AI can also do this.
A lot of scripts are just dumb glue code which is not hard for a programmer but which requires a lot of background the average user lacks.
@osmarks The dumb glue code is the syntax part to which I was referring. LLMs are bad at the actual logic part, because they are just text predictors.
@CraigDemel Have you actually used Claude 3.5 Sonnet? I also think you're underestimating the difficulty of programming anything at all for normal people.
@osmarks I'm betting NO because I think the actual work of telling Sonnet their requirements in enough detail to get what they want is more effort than most people will expand.
Also, I recently asked Gemini to tell me what the output of a NAND gate would be given two false inputs, and it failed. Some LLMs have patched things like this, but then fail at only slightly more complex tasks, because the entire architecture is designed to optimize token prediction.
@CraigDemel I consider stochastic-parrot-esque arguments unserious and am ignoring them. I think you underestimate LLMs' ability to guess roughly what you want. I have them do plotting code and quick Python scripts and such all the time from underspecified vague descriptions. I can also ask for tweaks if I miss something.
@osmarks Please continue to ignore the limitations of LLMs so I can buy more NO shares!
I also use LLMs to generate boilerplate code. But once you get off the reservation they start to fail hard.
@CraigDemel The average thing a user wants is not a weird trick but something doing fairly mundane data munging, and 2033 models will obviously be much better.
@osmarks If it's mundane then a programmer will wire a button to it and no one will have to script it.
What part of life would be so underspecified that hundreds of millions of people would have to write unique scripts?
@CraigDemel Programmers don't always bother with this stuff. I imagine it'll replace things like spreadsheets or some manual workflows.