Evan Solomon

Suddenly writing the code became ~free and some of the people thought “finally I can make all the code good!” and other people thought “finally I can make all the code bad!” and neither group really understands why the other guys are making it harder than it needs to be

@markjaquith @opencode Then you get updated docs and engineers can see if they are unintentionally violating some architecture design

@markjaquith @opencode Yea and ai only makes that easier. Imagine some ci job that checks all the coding agent docs and updates any that are made less accurate by the current change. Haven’t tried that but seems doable I think.

If you use CLI coding agents like @opencode or Claude Code, how are you keeping a working document for feature development? Are you just using AGENTS.md etc? Committing that to the branch? How are you merging that with repo-level instructions? How are you remembering to remove the feature working document before merging?

@markjaquith @opencode Honestly if the only thing these docs ever accomplished was recording what was kind of in someone’s brain when the feature was created that would still be good. I think they will do much more but that would be a win on its own imo.

@markjaquith @opencode Yea for sure. I think it’s hard for incorrect docs to be worse than none especially in the age of ai. I think causing the information to get created is most important and we (it) can sort out what seems wrong later eg add “check to see if it still seems accurate, update it, etc.”

@markjaquith @opencode Yea that seems like a good criticism. Seems easy to improve upon. Like put the branch name in the doc and have general instructions to search for the doc with the name. Maybe that’s still bad. My main goal was actually just recording the human’s intent, helping the agent was 2nd.

Yeah. I used to do that, then got annoyed at starting each session with “go look at this specific plan file”. Plus then it’s another tool call to go read it. Whereas if you can layer on feature instructions into the main file, it’s automatically loaded into context. I’m jumping around branches a lot and I like not having to do any warmup. Like literally can change branches and say “what’s next” and we’re off.

@markjaquith @opencode You have to tell it…probably. I guess it could go browsing around and figure it out. We should probably put some kind of general guidance in the main agents file like “coding-plans has ai plans for features check that our think about if any are relevant.”

@evansolomon @opencode And how does the agent know which plan corresponds to what you’re working on?

@markjaquith @opencode I don’t think it conflicts with a general agents file but maybe it does or should?

@markjaquith @opencode Outs wasn’t a repo with that pattern until recently. It seemed insane to me to say that the prompts are the new source code and then throw away the prompts. I made a new directory called coding agent plans or something and committed it. A couple people have followed.

@devahaz @Ryan_Gasoline You are 100% on her side even though you don’t know what happened beforehand?

@Ryan_Gasoline Well we don’t know what happened beforehand, but sounds like the dog was touching or licking her or something. Dog should be in a bag/carrier and not touching anyone else. I wouldn’t mind the dog there (vs a restaurant) but if someone does it’s on the dog owner not her!

I am 1000% on her side. Nobody has to like your cute little dog, follow the rules or don’t bring your dog. Dog owner entitlement is out of control. And yeah, if you’re gonna act all hard at first don’t try to play the victim later, pick a lane!

@leahculver Sorry! It mostly optimizes for photos where your face is kind of near the middle + some randomness. It does not check (yet!) whether your eyes are open 😔

I also love the face detection software on the front door but I feel like they could have used a better photo of me.

Pro tip for @ycombinator Demo Day… the WiFi is better in the women’s restroom on the first floor since there’s a WAP in there. I cannot speak to what is in the men’s restroom.

@markjaquith I think maybe humans don’t “know” anything either? Humans can make the same logical mistakes. Maybe we aren’t doing anything more than predicting the next token.

I think it’s different in terms of fuzziness. Like think how math or logic works, where you establish knowns, which are rock solid, and then you build on those. LLMs don’t ever lock that down. There are no facts, just weights. A useful example is how you can trick LLMs into saying “the doctor is the boy’s mother” even though the riddle given explicitly precludes this. The LLM isn’t doing reasoning. It just sees a familiar riddle pattern and follows the weights to the wrong answer.

It’s hard to keep this in mind because when pattern matching is working, it masquerades as knowledge *extremely well*

@markjaquith @dhh Butterfly anime meme guy holding a NPE “is this luxury?”

"Ruby is a luxury language" This is such good framing by @dhh. youtube.com/watch?v=vagyIc…

Don’t forget to go for a walk today

Forget the tariffs they should just charge $5 to watch a video of whoever wrote this explaining it to Trump and the budget would be very easy to balance imo

@markjaquith I cannot believe I read this whole thing and you did not link to the code. This seems so ridiculous I think it must be sending a message that I don't understand.

Further optimizations have it down to about 100ms for the first solution, and about one second for all solutions (depends on the day). More optimizations still available!

Was nerd-sniped by @helenhousandi when she showed me this 3D printed calendar puzzle. There are ten pentomino pieces and a board containing 12 months, 7 weekdays, 31 month days. Every day, it's a new puzzle, to arrange the pieces such that the current date peeks through. 🧵

@rebryk @ycombinator We wanted to make it a challenge! There actually are some basic heuristics for this that bias toward photos where your face is bigger and closer to the center. But it does not consider how focused (or not) your face is. The photos change so the next one might be easier.

I’ve spent a few seconds to find myself… @ycombinator could you fix that? 😆

@dsiroker @felipekiwi90 We did make it. There is an s24 company now working on a feature like this as part of a larger product.

@felipekiwi90 I think made by YC but definitely a very good product idea. I predict a YC portfolio company looking to pivot will steal this concept and commercialize it!

Just walked into the @ycombinator office to hang with @koomen and this automatically popped up in the lobby based on face recognition. Wow. The future is already here - it's just not evenly distributed.