Instead of cave man feedback cycle of edit text save and restart server I have agent plugged into the app via MCP that can see app state as I use it. If an exception occurs it pauses the app and signals agent to fix then resumes. It can hot load the code as the app runs. You can add feedback in the app itself that has recorded state and page you were on when you made it. Feedback is pumped right into agent to be worked on. You develop the app from within the app itself.
@cybersec I think it comes down to how well are you doing the basics. Detecting employees sharing passwords in a spread sheet, patching everything, detecting secrets leaking in logs, git repos, etc. then doing that every single day consistently. It’s the consistency that’s difficult.
I made a pretty big roadmap of 8-9 milestones with 5-6 phases in each then just kept telling it to do the next most important thing. For about 3-4 weeks non stop. Ansi-test helped and having sbcl to compare against for performance. It made me appreciate SBCL. My first ansi compatible iteration was like 22 times slower than SBCL.
The stdlib was made in like half a day. I had it analyze top imported Go libraries and modeled after that.
Cost per task vs human is already cheap, like near minimum wage cheap. AI cost per task is getting cheaper every release. Current model capability is already better than most humans. There’s no way humans will go back to writing code. Software will be infinitely more complex and costly to maintain for sure but we aren’t going back to writing code.
"Where things stand: I first ran the suite against SBCL. mycl is now 0.93× SBCL overall on this machine (geometric mean), down from 1.09× before the first round."
I vibed a lisp that passes 22k test ansi-test and loads top 100 quicklisp libraries.
@levelsio@ubuntu I’ve never understood why Ubuntu was more popular over Debian. They’ve done crappy things like putting ads for their cloud services in the motd among other things. What do you get over Debian?
@david_nix Models are good at moving toward 3 but you have to ask for it. If you ask a human to keep adding features and never do refactor or cleanup you’ll get 1 too
Making my own Common Lisp implementation has given me an appreciation for SBCL. That thing is highly optimized. I’m close to same speed but AI has been optimizing nearly non stop for a week and a half.
@taylorotwell I’ve been doing this off and on since February except I still do it all locally I just use VMs. Every job is a cheap git clone into fresh VM with dev tools pre installed and a thing that copies the change back out to host
Stateless no long running process, simplicity of copying text files to a server for instant deploy with no rolling processes for zero downtime deploy. There is actually a lot that is great other things don’t have. The language itself kinda sucks but there is some good. Also, these recent vulns are due to C and memory management not the PHP language itself.
@dcooper8 Im actually making one right now. I started two days ago. The verification plan and roadmap is quite something. I’m only on like phase 3 milestone 2 out of 8 phases and I have 22,000 tests.
@noahkagan I made my own: https://t.co/7AUN7Y0Rzp as a challenge to myself that I could replace QB in 24 hours for my own needs. It doesn’t do everything QB does but it works for me. After I built it I realized you don’t even need software. AI can do the whole thing from data.