Yeah. I have a set of 5 review prompts attacking different problems, an adversarial review, and then a final synthesis, with the best results gotten by multiple passes using multiple models (the adversarial review stage combining all passes into one review per model and the synthesis picking the best of the two or three adversarial reviews). Expensive but it actually finds real problems that the single pass reviews rarely seem to find.
No one reads the terms and conditions. I went to a resort and read the T&C they made you sign to sign in and was told I was the only person in months who had actually done so.
And even I have mostly given up on the website T&C because most of them are so lengthy, a lot like I've given up on disabling javascript since the modern web frequently won't even render anything if you disable it.
It's not JUST about output, it's also about learning. The output is a way to measure the learning, it's not important of itself except as a reflection of the student's mastery of the curriculum.
You're thinking of the output as the goal rather than as the means to an end.
400k isn’t crazy for the FANG set but it’s still a subset of the developer market and hundreds of thousands of those jobs have been cut in the last few years as they all collectively work to lower SWE pay.
60k a year it needs to be a full irreplaceable part of the infrastructure for I think. There are very few kinds of software that meet that bar right now (certain design tools etc that have no replacement). 12k/year is in the expensive but reasonable for the right tooling category (Matlab etc.).
I don’t know what the future holds. I know the big AI companies are banking on being able to charge for a replacement SWE that works 24/7. Still not convinced these are it yet, as useful as they can be under the right circumstances.
Yeah, it’s so easy to generate code that you can do a whole codebase rewrite in a day.
Is this a good idea? Probably not—in the past we would only do that when the architecture was causing serious problems since it always has tons of behaviors that will accidentally not get carried forward, some of which are load bearing and will cause bugs.
Now we can do it in an afternoon and get the same long term bug behavior.
Yeah, my experience has been exactly that. And the person paying the bills will try as hard as possible to avoid/delay a metal rev as possible because they’re expensive and time consuming so they blow up the schedule if you can’t release the chip with rev. A.
One place I worked at did fast iteration by pushing as much of the risk as they could off the silicon and by using several distinct ASICs instead of a single monolithic one which would have had better performance on its own. Gave them the ability to rev the different parts at the rate they needed it at a cost to software complexity and hardware compatibility and cost.
Yeah, that was my experience. The model was worse in every way than ChatGPT or Claude or even Composer. I tried it out and used it when my other limits were hit, but only as a last resort. And I stopped doing even that because the model was so bad.
If you're wealthy enough $250 is just the price of smoking (especially for someone that can afford a 1st class seat). I wonder why they don't have escalating non-monetary punishments?
There has been a measurable and noticeable drop in attainment starting with smartphones entering the classroom, supercharged by COVID chaos, and finally with AI cheating being just the latest assault on learning.
Ask teachers that have been teaching for 10 years. Ask the professors how today's kids are different than the ones of yesteryear.
The move to de-tech the classroom will eventually help out I expect, but keeping kids (and adults!!!) from using cognitive shortcuts so they can develop their own sense of what's reasonable instead of taking information from a bought-and-paid-for oracle is going to remain a problem.