Just on the back-end of that post.
Looks like companies will be more cost-conscious, but that token demand is going to grow at a rate that will require increasing amounts of memory/chip production.
So hold off on selling memory manufacturers because margins are going to keep increasing for another year due to the production shortages.
Iām not getting the cost here. I have a pro plan with claude. Costs me $200/month. I use it hours a day and never have a problem. Maybe that gets capped if youāre compiling, I dunno. But I was using the unlimited API before and it was only costing me $1000. Given that one person with AI can outcode 5-10 people, I donāt see the concern over $1000 per coder.
I donāt think that is necessarily the speed up in a professional development shop with many different developers working together etc. I saw studies showing a 30% to 50% speed up. Significant to be sure, but not one person out coding 5 to 10 people, which i guess would be a 80% to 90% speed up.
I suspect thatās either on average, or for people that arenāt taking full advantage. Because Iām definitely getting stuff done significantly faster than if I had multiple developers on staff. And my friend said one good Ai developer is outcoding their entire team.
I think it must depend on the kind of work being done, and the environment itās being done in.
As I recall, Linus Torvalds of linux fame at one point remarked that he now uses llms for new projects but thinks it would be terrible for maintenance. More generally, programming low level kernel code of an operating system like linux is going to be very different from quick, almost one-and-done projects for a large enterprise. That is just one example.
So far from what Iāve seen, the user drives the actual efficiency gain. We definitely have people that are significantly better at wielding AI as a tool than others. Partly better prompt engineering, partly having the knowledgebase to know what to ask for in the first place.
theyāll make it up in volume
Currently using AI to create scraper programs. Iāve had these hand built in the past, every time I use them they break and I have to pay even more money and have ays of delays to get them to work. I dread using them, they break every time and cost me a ton to fix.
Aiās going through the programs, finding all the break points and fragile points down to the millisecond, and building a program to do the scrape. Iām ending up with a program that is robust, is 3-10X as fast as my old models, and cost me nothing to build. Yay.
Starting to get a bit complex. I run ubuntu with windows inside a virtual machine. So now Iām running claude on linux and inside the virtual machine, and the two are talking back and forth; the linxu version runs the overall strategy, the windows copy operates the scraper.
