It's been a while since I wrote anything so I wrote an article on how to discover the entire x86-64 instruction set in seconds including any hidden instructions and learn their basic properties while on it. There were some pretty interesting results! blog.can.ac/2021/03/22/specu…

jev captcha solver would go so hard

First look at the Tern iOS TestFlight! You will be able to omp on the go.

ladies and gentleman, I am happy to announce I can now draw braces

I was trying to find some code I saw a while back, instead found this gem in the sourcemaps 😆

I keep seeing people rt the claude-gzip tweet saying "oh this was so obvious, look they are discovering, this is what I do since 1942" meanwhile completely missing: - large conversations have thinking signatures + images in b64 which are high in entropy = compression is not that efficient anyway - rest would be hit or miss considering gzip adds latency on both sides => if you have decent bandwidth it might just hurt instead of helping - rep. HTTP requests where you send entire transcripts is a horrible method of exchange, needlessly O(N^2) - more importantly, since they have `store: true` implemented for API, just not OAuth, they clearly know how to optimize it just fine > look at this [horrible] thing they just implemented! i would have implemented [horrible+1] on day 1, smh they are learning the fundamentals like shut up man 🤣

God forbid one's system prompt makes the agent follow user instructions, oh the horrors

gave astra a try again, very beautiful

oh no

he cant keep getting away with it. @_can1357 is on a generational run

Opus, cancel my google AI ultra™️ subs

It seems oh-my-pi (OMP) is one of the best coding agent harnesses to pair with OrcaCyber Zero 1.5. Based on our internal red teaming experience, its architecture is particularly well-suited for vulnerability discovery: LSP integration — Trace symbols, references, and cross-file relationships. Native debugger — Inspect actual runtime behavior instead of guessing. Persistent Python/Bun — Build custom analysis and exploit validation workflows. Parallel subagents — Investigate multiple attack surfaces simultaneously. Context compaction + memory — Sustain deep investigations across massive codebases. Advisor model — Challenge findings and reduce false positives. These capabilities align remarkably well with OrcaCyber Zero 1.5’s specialization in vulnerability research and zero-day discovery. A great cyber model needs more than a prompt. It needs the tools to investigate, reason, and prove. OrcaCyber Zero 1.5 + OMP is looking like a powerful combo. 🐳

having an idea and finding a 2026 preprint on arXiv that's almost identical is so great, I used to go "dang" when this happened but: 1. it's free independent validation 2. clearly they didn't manage to actually use it need some compute but I might just save +30% on frontier tokens in return (both knowledge work and coding) 💪