Thanks, that does bring my estimate down by about two orders of magnitude (to 1 error / 10,000 lines being the current standard for “very good” code). I think I mis-remembered the NASA one as 1 per million instead of 1 per 500,000 and thought that it was more typical than it was.
Still, 1 error per 10,000 seems to be definitely achievable today, and can probably be made higher if we need it. Hopefully by the time we’re dealing with AGI we will have better metholodogies.
Here’s some evidence for you to update on.
Thanks, that does bring my estimate down by about two orders of magnitude (to 1 error / 10,000 lines being the current standard for “very good” code). I think I mis-remembered the NASA one as 1 per million instead of 1 per 500,000 and thought that it was more typical than it was.
Still, 1 error per 10,000 seems to be definitely achievable today, and can probably be made higher if we need it. Hopefully by the time we’re dealing with AGI we will have better metholodogies.