Calvin Alkan

@calvinalkan

Randomly hitting buttons and hoping for the best. Zero bugs so far... 🤞 Building fortress.snicco.io & commandui.com

Joined June 2023

170

Following

523

Followers

Replies Reposts Media

Calvin Alkan
@calvinalkan
Sep 10
For some reason, today, Codex in PI started messing up some tool calls with wrong arguments and instead put (what I assume to be) the reasoning traces in tool call arguments. Is anyone else seeing this? Pretty funny stuff: "Need perhaps AGENTS already loaded. Let's inspect.
1
Calvin Alkan
@calvinalkan
Sep 4
Apparently, adding a simple ESM `export` can make hot path code ~10% slower in JS/V8. Below, the `export const` variant has: ~ +10% cycles ~ +14% instructions ~ +16% branches. You can get incredible performance out of V8, but getting there feels like a science experiment...
1
Calvin Alkan
@calvinalkan
Aug 18
JS Benchmarking is super hard. You need to account for JIT, run each batch in a new process, NEVER run more than one scenario (i.e lib a vs lib b) in the same process or you will totally confuse Turbofan and it's all fake. You need to properly prepare input, you can never use
Mitchell Hashimoto
@mitchellh
Aug 14
A couple more notes on this: First, the xterm.js benchmarks were very noisy. A lot of the results relied on whether they got lucky with the V8 JIT/inliner. e.g. the frame times for ASCII ranged from 20us to 100us (!!!). For the purpose of optimism, I only used the fastest times
1
Calvin Alkan
@calvinalkan
Apr 23
You mount the feedback loop as ro into the session.
Nicolas Hillegeer
@alazyleopard
Apr 23
Replying to @neogoose_btw
How does one prevent the autoresearch agent from just emptying the tests (or modifying the tests in other ways), if one asks it to optimize those? The examples make it look so easy, but I feel like more structure would be helpful.
Calvin Alkan
@calvinalkan
Apr 4
Now do WP-admin
GIF
Dipak Gajjar
@dipakcgajjar
Apr 4
A properly cached WordPress site on FastCGI serves 1,097 requests per second with 5ms TTFB. That's not speculation. That's a benchmark on real hardware. EmDash on Cloudflare Workers? Also fast. Also global. But here's the difference: WordPress gives you that performance on
3