Paul Kuruvilla

Started actively working on the next codecrafters.io challenge today: "Build your own SQLite". I've got the stages planned out, now onto building the test cases and trying the challenge out myself. Will post updates here as I go.

@jarredsumner @typedfemale If you haven't tried it yet, Nurri has a similar ratio and tastes far better imo (vanilla)

@typedfemale calorie : protein ratio is v good

people are calling fairlife "chud soylent"

@gabriel1 Also you gotta say "make no mistakes", else codex assumes that you want mistakes

@gabriel1 Everything except the last 3 words make sense. Mention early returns and it's going to write the stupidest early returns ever. just can't resist following dem instructions

only bottleneck is consuming code, so make sure to tell codex that you want just that: "write extremely easy to consume code, optimize for how easy the code is to read. make the code skimmable. avoid cleverness. use early returns."

Funny how everyone's optimising for DX now that agents can code - why couldn't y'all have done this when us humans were doing the work???

@a1zhang @ankrgyl Ah. Hard to tell without seeing the full prompt but splitting into n judges, one per criterion, might make it more reliable?

@RohitPaulK @ankrgyl Judge is given a rubric of items the answer must satisfy. It returns a list of which items are satisfied and which aren't, reward is only given to this answer if it satisfies all.

Ran a small eval today on an LM using GPT-5.2 as a judge. Model scores 10%, but paper reports it scoring 34%. I see that the paper uses GPT-5.1 as a judge; for the sake of consistency I change it. Switch to GPT-5.1 as a judge. Model now scores 43.5%... bro

@ankrgyl @a1zhang Curious to know this too

@a1zhang what is the judge's definition?

@schwad_rb This was fun!

Rain is growing rain.schwadlabs.io

@badlogicgames @vokaysh "enough rope to hang itself" 🤣

i can't speak for david. what i see is this: if you let agents build or extend a codebase with only minor or no supervision, you get unmaintainable garbage, because the agent makes terrible decisions that compound, both big and small. those decisions make it hard for both you and the agent to keep modifying the code base, until eventually it's unrecoverable. why does the agent make bad decisions? i can't tell for sure, but my gut tells me that training data can currently not capture the holistic thinking needed to design and evolve complex systems. that's one part of the problem. related to that, and oversimplified: agents output the "mean quality" of the code they saw during training. most of that code is very bad. specifically tests, which humans are terrible at writing at. another part of the problem is that specification via prompt is not precise enough, so the agent has to fill in the blanks, giving it enough rope to hang itself. the more detailed your spec gets, so the agent gets constrained and less likely to produce crap, the closer you are to handwriting the code yourself, as that's the most detailed version of the spec that can exist. so then you gain nothing. back to prompt spec it is, which means the agent fills in blanks, which means we get suboptimal or truely bad results. using agents can still be a net productivity boost (see other posts in my thread), but it is not easy to come up with consistent workflows that produce both production quality maintainable code while retaining the speed advantages agents give you.

recommended reading sure to ruffle some feathers. but it's largely true for now. keeping the complexity off the bay is really hard, espwcially if you go full agent orchestration. even if you don't, and human in the loop a lot, automation bias kicks in and your reviews of agent generated code become mostly performative.

@geoffreylitt Agree, also because the stakes are far lower

Hot take, grocery shopping makes more sense than travel planning as an AI use case. Frequent and semi-repetitive

@comma_ai E2E long is awesome. Would be great if it could handle traffic lights and stop signs (hyundai tucson 22)

What do you want to see in openpilot this year?

@eurydicelives Work feels useful most of school felt useless and enforced for historical reasons

I know I'm doing a lot of these posts but if you hated school and hate work less (or even like it) I am interested in all the reasons for that difference. "They pay me" is not an answer if I couldn't pay you your current salary to go back to school.

@echantech1 Are you using opus 4.6 fast mode? Had a similar increase and that was our root cause (we'll gladly pay though, assuming costs will come down eventually)

Cursor pricing is wild. In the last week, using it full time, I blew through $1000 of tokens. 6 months ago, I would only spend that in a month. Something wierd is going on with their pricing.

@rs545837 I'd love to understand this better too - how much of it is pure LLM calls vs. all the other infra bits in-between? Don't think users will put up with the delay in the long-term

I don't understand why these code review tools are so slow.

@zebriez Surprised that no one mentioned @rtwlz here

who is the person you know that does the coolest side projects?

@Sirupsen @samlambert @PlanetScale Is this real

@samlambert @PlanetScale did I ever tell you I lose my vision for 2-4 hours when I have diet coke?

we are very well stocked with diet coke at @PlanetScale

In the build your own x world this is like the sequel to jurassic park

@SherryYanJiang First rule of taste club: we don't talk about taste club

ok but really can someone tell me what taste even mean anymore

@jeffdfeng openclaw for $1b, moltbook for $3b, things really be crazy out there

so meta has just acquired moltbook, just weeks after openai hired the openclaw creator how much is the social graph of autonomous agents really worth?

@RhysSullivan By that logic does it still make sense at $250?

A $15-$25 PR review bot that catches an incident which would've cost the company $5m in breached SLAs and reputation is a no brainer

@danielmerja What's a way one could use this endpoint in an "attack"?

Queues for post-processing? Totally fair game. But leaving a detailed /health endpoint wide open to the internet with failed job counts, queue depths, etc.? That's not "vibe coding", that's reconnaissance gift-wrapped for attackers 😭