The AI emperor has no DAUs why most devs still don't use code AI: Quinn Slack

[@dak2009]

Using an AI is like running a continuous code review with junior developers who never seem to learn from the experience.

[@marvin_hansen]

Have you tried windsurf Ai.? This thing just fully implemented an advanced high performance data structure for me today including all tests, benchmarks, comments and even the PR message and summary for the Readme. I have three degrees in computer science and this thing pulled off in one day what would have taken a week or two for me. And then it added SIMD optimization… I think the Ai junior dev days are gone for good 🎉

[@avi3681]

I was surprised to learn that it was surprising to the speaker that devs have a chat session open while coding. This is the only way I've used AI tools. I find autocomplete super distracting. Part of the problem with building specialized tools is that just copying and pasting into Claude or ChatGPT is 90% effective, and it's hard to see how custom tools really would make that much additional benefit. I mean, I could save a few keystrokes if it was integrated in my IDE but that's not enough to make me pay more for a specialized service when I'm already paying for access to foundation models. I'm sure the whole process will get streamlined over time, but I'm guessing that open source tooling for VS code with people putting in their own API keys for their favorite LLM is what it will win the day.

[@hamburger--fries]

One huge thing is using Vector DB as a knowledge base for Framework / API / Language Documentation. Turn the docs into MD files and then put them into a Vector DB and feed it back to your AI Tool.

[@ibgib]

Docs should go away. JIT AI "docs" based on actual code with known caching strategies to handle tradeoff of time vs compute. This will help minimize out-of-sync interference that docs/comments have on AI codebase analysis.

[@hamburger--fries]

@@ibgib What does JIT AI DOCS mean? Can you send me a reference? Some person still needs to make some documentation, even if it's for AI.

[@ibgib]

The reason that it is mostly code complete and so few use it really is that as a developer tool, it just isn't there yet. Anything that is non-trivial one a real-world codebase is a horribly poor experience. IMO (and I'm hugely biased because it's a primary use case of my protocol), this stems from the biggest monopoly in the software world: git. Like all monopolies, this is a local maximum and the technical debt is costing the entire industry. UNLIKE any other monopoly I've ever come across, natural or no, ZERO people are talking about this. This means ZERO people are investing in ANY attempt at a solution. Now is the PERFECT time because agentic workflows create branching possibilities. The ultra cool thing is that AGENTS THEMSELVES are part of the possible branching paths (e.g. Mixture of Experts MoE). A FINER-GRAINED "version control system" is necessary. HUGE OPPORTUNITY! WAKE TF UP! (TF in honor of all of the cussing in this video).

[@tthtlc]

in order to use AI, in the current state of art, is that you need to explain to the AI in natural language how to do the programming. yes, many times you can skip the details, and you can see AI make a lot of assumption and come up with its own algorithm. but many times it is so much faster to just code ourselves, than to think of what english word to use to describe what to do. software at the low level can be very difficult to describe and characterize.

[@pisanghangus2]

The higher opinion you have of yourself , the less likely you are to use it

[@astronaut22]

❤

[@ordinarygg]

Oops, no one wants your bubble, ahaha, yet to discover what is product market fit.

[@brucem8448]

Devs work 2 hours a week. Inefficiencies keep jobs.