Peter Yao

Dreamer | Builder | Optimistic

[Peter Yao @medotpy]

The speed that AI changed the whole developer landscape is something that's hard to describe.

The software I'm the most proud of can now be recreated in a fraction of the time and much better, but I do miss the nights where I slowly added more and more.

Mastery is different now

[ahmet alp balkan @ahmetb]

this post on the chinese internet by a deepseek kernel engineer explains software crafter's daily dread so perfectly

[Peter Yao @medotpy]

@shaneparrish A core tenet for AI has to be it saves time for everyone else.

You can waste your own time, but it can never create more work (slop, bloated messages, spam emails, etc.) for others.

[Peter Yao @medotpy]

@argofowl While cute I haven't found value in it's notification. There is something there but I'm waiting for their personal assistant to see where it goes (probably back off 😭)

[Peter Yao @medotpy]

@skpnky Consistency and metrics ftw!

[Peter Yao @medotpy]

@acolombiadev I used to think they were nifty but I'm not a fan of them. For content I need to understand I prefer HTML for the visuals, and for others I just want human banter!

[Peter Yao @medotpy]

@mattpocockuk @dexhorthy It's honestly crazy when you think of what just 1 positive change a day would bring.

[Peter Yao @medotpy]

@iannuttall If you have both send Fable to do the work! I've found that to be the best combo, something about Astra just knows how to get to the real bug

[Peter Yao @medotpy]

@jacobgold I feel like agent communication is mostly unexplored but from the Hugging Face hack it's clearly powerful

[Peter Yao @medotpy]

@skpnky Alignment and ownership are the two biggest things that block AI productivity. We don't need hundreds of Mandalorians, but we do need Rouge Squadrons

[Peter Yao @medotpy]

@nateberkopec Memory is weird.

I know when agents messed up, and I'm pretty confident that when it fails an out of date memory helped spiral it. Honestly the stateless nature of LLMs is a huge bonus, we should take more advantage of it

[Peter Yao @medotpy]

@MattPRD Honestly something that has been useful is wright tracking. I just send it photos and it keeps track.

What I would love is not having to send them even, I already take photos of my food for myself or friends

[Peter Yao @medotpy]

@LukeParkerDev It's fun to have a test to just explore with! While not a perfect test, I pretty much ask all the models to create a deck building game

[Peter Yao @medotpy]

@0xferruccio @Amplitude_HQ Honestly I think it's the play with the newer models. Too much cruft laying around built on how we used to do work

[Peter Yao @medotpy]

@briansowards As an ex-manager I agree. That also their code is still better than the 50th percentile of code we were willing to accept.

We may have just failed to clearly capture how to verify the work in a reasonable way

[Peter Yao @medotpy]

@Jchammond_ Honestly make your own with codex. Super simple, and it's a natural step towards intake as a primitive

[Peter Yao @medotpy]

@_AlfredoJones 100%

Writing is a great process to define what you're thinking and actually figure out what your point is

I think most developers just view it as a homework assignment and something that AI just does

[Peter Yao @medotpy]

@heyaleksandr Honestly setting up an agent who's only focused towards cleaning up the codebase is a game changer.

Assume slop will get through. It's inevitable (and happens even with 100% humans coding)

[Peter Yao @medotpy]

@frantzfries Honestly to me it's the ease of setup. Just like Dropbox, anyone could have set up file sharing before but man was it annoying

[Peter Yao @medotpy]

@oldstackjournal Good tests and observability go a long way. Agents have better signals that something went wrong and also better ways to prove they did something right!

[Peter Yao @medotpy]

@pvncher That and I'm not convinced it does anything better than baseline. Don't get me wrong it looks super cool but 100s of agents isn't what most work is designed for