I’ve only seen a guardrail block legitimate work. I’ve never seen it stop something I would have wanted stopped. Anyone have counter examples?

The moment you realize it’s easier to build your own text editor (code browser) than to configure a vim plugin.

I’ve started to build my own text editor. Is this the new form of vim config tinkering?

An amazing thing about lisp is the code looks like it was designed yesterday despite being very old. It is truly timeless.

If I were to give previous self advice on getting up to speed on AI I would say: get a $100/$200 month plan and learn to max it out every week doing “vibe coding” on new projects. Don’t look at the code (again new low risk projects). Learn the Ralph loop technique and actually implement a few. Not because you will work that way but because it changes how you think about working with the models. Just play and try everything.

Looking back at the last 30 days or so working with AI my highest leverage thing is my QA stateful agent. It’s found and fixed a bug, fixed app inconsistency with docs, or made an improvement almost every waking hour since it started ( it kicks off another run once merged). It is essentially a classic Ralph loop with a journal and some self reflection. Try using agents for QA instead of just code generation.

I’ve created a complete web-based video editing software in 2 days in Common Lisp. Swank/REPL over MCP lets LLMs fly. Use the app as it is built and hot loaded. Error conditions auto trigger to the LLM so it fixes before you notice.

The learning to utilize AI with coding path is so predictable you can literally tell how many hours someone has been coding with AI by how they use it and words they say. Hearing someone talk about it for 5 minutes I know exactly where they are on that path.

These agents are REALLY good at QA testing. Make them keep a journal of findings and tests and plan next run then loop.

I honestly didn’t know JS compiles to machine code while it runs. I figured all scripting languages were interpreted

Having some success with a QA agent. Point it at your docs and a copy of your app and code and let it hammer on your app inside a VM. Mine takes notes on what it’s tried each run and fixes things it finds.

ChatGPT realtime audio thing has completely replaced google for me for quick one-off questions about random things.

I make my agents write all agent code in dependency free Common Lisp as an homage to the original AI.

I think I’m on to something here. Instead of creating loops and agents. Have them create themselves. Start with seed prompt of single goal/role and folder conventions. code, memory, crons, cache, inbox, outbox, for-operator. Run them in a sandbox from supervisor. Inject secrets tmpfs. Everything is a file. Agents create crons for themselves approved by supervisor. Creating new agents is just a single file in a folder with seed prompt. Inbox is operator prompts or itself from cron runs. Supervisor sets loop limits and max crons to keep agents from spinning out of control. I have 3 of these going and seems to be working well so far.

Anyone experiment with AI landing page optimization where AI selects experiments, runs the A/B split test, selects the winner and loops? Seems like an obvious idea.

I mostly use Claude but the real time audio thing ChatGPT released is awesome. I use it all the time now.

Fable has been cut off early