full time software engineer and part time security researcher
Everywhere
Joined June 2010
- Tweets857
- Following336
- Followers221
- Likes59.8K
glm-5.3-flash vs glm-5.3-flashx
- 2 request each (bottom to top)
- first one is cold (non-cached), second is warm (cached)
Sayem retweeted
If you think this is fake, go read Hacker News comments on any AI-related post
There’s a shockingly large number of engineers refusing to use AI because “the code it produces isn’t good enough”
X is a bubble, outside of it, AI isn’t as prevalant as it may seem
Met an engineer last night that isn't into AI.
He got laid off in Feb.
He had an interview recently and the CTO ended the interview midway when he said he hadn't used AI much.
Feel bad for the guy, but it's true basically no company would want him.
He told me AI has yet to prove its worth and it's all just hype. I didn't try convince him otherwise. But feel bad for him. Might never get a job again. He's convinced AI is a hype cycle that will die down.
Sayem retweeted
thats because technical screening/recruiting is done wrong
everything should be done with work trials mapped to the actual job
look at the persons ceiling not their floor
resume focused recruiting does not allow you to find hidden gems
pro sports teams never fill their entire roster with "proven veterans" they take chances on high ceiling potential draft picks who are hungry
git bisect is so underrated. You almost never use it on solo projects, but when working in teams and pulling massive change sets, running automated bisect tests saves hours of debugging.
Quick guide + example script:
gist.github.com/sayem314/870…
Sayem retweeted
That would be @37signals - how's it going folks? Seems pretty decent to me.
real world example: github.com/omacom/omarchy/pu…
check what it can do autonomously
github.com/sayem314/test/iss…
check what it can do autonomously
github.com/sayem314/test/iss…
Using @zeddotdev for a while now, got used to it. Glad made the switch from vscode/forks.
tested with different harness, same provider and same model with same effort level and oracle always performs better and also uses less tokens.
Fair scoring: if you want org deployment, IDE integration and MCP ecosystems, codex and opencode are excellent.
If you're one operator living in a terminal who wants a fast personal agent that knows your repo and asks before doing anything dumb, nothing I tested fits better.
Two design bets validated by the field:
1. Line-oriented shell over full-screen TUI. fx, built by a platform team, reached the same call I did.
2. Agents guess silently by default (AMBIG-SWE, ICLR 2026). Making asking a first-class tool call fixes real accuracy.