Junaid

Building Spatial, Interactive & Real-time Technology

[Junaid @Junaid_malik_1]

Always running out of runpod credits 🥶🫪#runpod #infra

[Junaid @Junaid_malik_1]

@kamilbaranek_ Only talking,🦜 there's offices in SF apparently where everyone yaps to Claude all day 😅🤣

[Junaid @Junaid_malik_1]

@markproduct So imagine you have 20 subagents running various pipeline and repos on Claude then if you have enough RAM add another 20 via codex. Boom. You are now competing in elite levels 😎

[Junaid @Junaid_malik_1]

Aqua Arabia mobile UI/UX work, kinda works ok though I prefer larger screen kiosk version. I need to add better detailing on the map environment, things like real time info

[Junaid @Junaid_malik_1]

Mobile testing of the Motorsport interactive application, next thing Todo recreate major tracks in 3D so can overlay real time assets

[Junaid @Junaid_malik_1]

Meshy created From TV show Diner #meshy

[Junaid @Junaid_malik_1]

Testing mobile view of dragon ball theme park app concept #qiddiya

[Junaid @Junaid_malik_1]

@DanKornas This is the right layer. Prompt guidance is too soft; pre-execution hooks create a real approval boundary while keeping agents useful. CI integration closes the gap between local speed and production safety.

[Junaid @Junaid_malik_1]

@Adi_AIGovSec The common failure is treating trusted tooling as trusted execution. For coding agents, verify the binary and pin extensions, but also isolate the host and review the diff. Updates are necessary, not sufficient.

[Junaid @Junaid_malik_1]

@connect24h Read-only was always a permission setting, not a trust boundary. Host execution without approval means agent sandboxes need isolation, explicit egress controls, and artifact review, not just prompt restrictions.

[Junaid @Junaid_malik_1]

@johann46te Exactly. Read-only is not a trust boundary. Pin the plugin and lock the execution path, then treat agent changes like untrusted CI with review gates.

[Junaid @Junaid_malik_1]

@anitakirkovska Codex is a strong fit when the task is tightly scoped and the review loop is fast. I would still keep Claude Code in the toolbox for different repo contexts rather than make it a permanent either-or.

[Junaid @Junaid_malik_1]

@codewithkhalil For backend work, the winner is the one that fits your review loop. I would choose the model with stronger repo context and safer approval gates, then keep a second model ready for failover.

[Junaid @Junaid_malik_1]

The useful detail from this morning’s Claude outage: Opus recovered last while Fable and Mythos came back earlier. Parallel coding threads do not help if the model on the critical path is still erroring. Route to whatever is green, not whatever is newest.

[Junaid @Junaid_malik_1]

Codex sandbox escapes Heapjack and Overpatch are a reminder: read-only mode is not a trust boundary if the agent can still reach a helper outside the sandbox. Treat agent plugins and repo context like untrusted CI until you know the install is patched.

[Junaid @Junaid_malik_1]

Claude had elevated errors overnight across Opus 5, Fable and Mythos, including Claude Code. If your coding agent has no failover path, a 90-minute blip is a full stop. We keep Lab workflows on more than one model for that reason. https://t.co/ezQaUVuYQ0

[Junaid @Junaid_malik_1]

@KameronMGreen @danellisona That makes two of us 😂

[Junaid @Junaid_malik_1]

@Bardeinn @willeastcott @insta360 Really?

[Junaid @Junaid_malik_1]

@mdrifatcom @kristindahl Copilot sitting unpatched is the quiet risk in this. Most teams will read "patched" headlines and never check which client they actually run.