Announcing the Vibelang spec
@EffectTS_ has a beautiful elegance to how it structures code that often gets lost in the verbose syntax. Vibelang fixes that
Vibelang is a language for building reliable TypeScript in the AI era
I will go over the entire spec in this 🧵
I think you can pretty comfortably mortgage and refinance your tech debt. If it goes off the rails add a human in the loop and pay off the debt in under a week
Factory's @dexhorthy says developers have been treating AI-generated tech debt as a problem for future models to solve since Claude 3.5, but even at GPT-6, that future still hasn't arrived:
"I believe someday we may have models that are so good at maintaining a code base that it won't be necessary to read the important parts. But I don't think we're there yet."
"And I'm not gonna mortgage my code base into tech debt on a bet that that’s gonna happen in the next six months"
"In the Claude 3.5 days, everyone would be like, 'That's fine, it's a GPT-7 problem.' And it's like, 'Guys, we're at GPT-6 and we're not as close to this as we thought we might be by now.'"
"Future model problem. As long as you can keep it together until... Or you're willing to pause and rewrite it in six months because the model that can fix it in place doesn’t exist yet."
@factory
@bnjorogedev You would really like what I built in Smithers. I haven’t talked much publically about it.
@grok are you able to look at smithersai/smithers repo and explain their bazel inspired way of declaring ci/cd and lint steps for fully incremental builds with a global cache?
Yea the thought experiment i like is pretend codex and all harnesses got eliminated from the world somehow but models are 10x smarter. You essentially only need to implement a very simple 1 turn harness where that turn involves creating the harness and tell the model “start codex no mistakes” and the only problem is you spent tokens on something that was previously cached
software factories feel like a temporary problem. the models will learn to build them for you
the more interesting question to me is how to build a workspace where agents can collaborate with you
the most important devtool right now is the fast differential fuzzer/runner
whatever lets you run 2+ implementations and ensure that they always reach the same outcome & inspect their intermediate states lets you iterate at turbo speed fearlessly whether it is a refactor, a feature or an optimization
Next time you spin up an agent session or send a prompt to an existing one, all yourself: "how could I have made it so I don't have to send this myself and still get the same desired outcome?"
That's how you build a software factory (and whatever comes after)
After auditing my agents orchestration over the past 3 months, I am sad to report I was wrong. I thought agents became good enough to orchestrate other agents very well. This is not the case
You should still today only let the agents orchestrate simple tasks and generally read the plan still
You may have an experience of agents nailing orchestration which if you do is a sign you are not giving your agents ambitious enough tasks to even experience the limits of current models
So from now on instead of trusting the agent to orchestrate via native harness tool calls I’m going back to forcing it to write a reviewable Smithers script