Farzad Mahmoodinobar

nlp@amazon

[Farzad Mahmoodinobar @fmnobar]

@thsottiaux After updates, reasoning levels sometimes change (e.g. from xhigh to medium) on macOS app

[Farzad Mahmoodinobar @fmnobar]

@thsottiaux @AnthonyBerlin received the same error on two threads (Mac app). fixed after restarting codex. here are the threads:
019dbd29-ba48-74d2-b70c-d897453df057
019e753a-31c9-7842-b62c-97d74e790f16

[Farzad Mahmoodinobar @fmnobar]

@ajambrosino

[Farzad Mahmoodinobar @fmnobar]

@JustinBleuel @ChatGPTapp Chats within a project in the Mac app do not have a model selector. Web UI has it

[Farzad Mahmoodinobar @fmnobar]

@Dimillian Well, mostly unchanged after more review. So far remote barely works. Happy to share session info if it helps

[Farzad Mahmoodinobar @fmnobar]

@Dimillian Turning off private relay did in fact improve

[Farzad Mahmoodinobar @fmnobar]

@simonw @bradenjhancock That said, we pick the smallest model that can get us there (cost to serve/latency tradeoff) so I haven’t seen many use cases where the latest models are needed due to diminishing marginal returns.

[Farzad Mahmoodinobar @fmnobar]

@simonw @bradenjhancock I think I had mentioned this in your earlier posts. Domain-specific use cases can benefit a lot from fine tuning. We use this in machine translation in Amazon since we use Amazon- specific terminology that does not translate well with generic translation tools.

[Farzad Mahmoodinobar @fmnobar]

@thsottiaux Computer use in menu bar indicator could show which thread is using it. That’ll help me decide whether to cancel it or not. Sometimes I cannot tell why a computer use is active.

[Farzad Mahmoodinobar @fmnobar]

@thsottiaux My sessions some times reset back to medium thinking level. Would be great to make the thinking level persistent after app restarts/updates

[Farzad Mahmoodinobar @fmnobar]

@Dimillian 4. on the backlog, I'd love for codex to review them overnight, rank/prioritize them and even implement the easier/clear ones and submit the PR that I can review when I wake up - changes traffic pattern to night time and gets my work done faster

[Farzad Mahmoodinobar @fmnobar]

@Dimillian a few requests:
1. menu bar with 5h/7d limits (I built a menu bar app to show me these limits but coming directly from app would be preferable)
2. ability to queue plan and non-plan requests on a thread. rn, if one is plan mode, others won't be able to implement either

[Farzad Mahmoodinobar @fmnobar]

@Dimillian 3. backlog tracking. I've added a backlog thread to each of my projects but would rather have that as a tool

[Farzad Mahmoodinobar @fmnobar]

Install:

git clone https://t.co/8P9Ssbf7d5
cd codexStatusBar
./install.sh

[Farzad Mahmoodinobar @fmnobar]

built a small macOS app to show Codex limits in the menu bar. I'm sure @OpenAI will add this feature at some point but in the meantime, clone, install and run. it shows up like this in your menu bar:

https://t.co/r2Ln3jo9p2

[Farzad Mahmoodinobar @fmnobar]

@thsottiaux One more request @thsottiaux - I usually end up adding a few requests to the same thread and codex works through them over night. If one of them is in Plan mode, all of them switch to plan mode. Would be great to make plan mode request specific

[Farzad Mahmoodinobar @fmnobar]

@thsottiaux I keep a `backlog.md` for each of my projects. would love for Codex to review that at night, rank by priority, scope and maybe even create PRs for the ones that are clear to execute and update me in the morning (like ChatGPT Pulse for Codex). it'll help flatten the traffic curve

[Farzad Mahmoodinobar @fmnobar]

@simonw Domain-specific machine translation (terminology, formality, etc.) - we have a bunch in prod.

[Towards Data Science @TDataScience]

In machine learning, the bias-variance tradeoff is a classic concept to understand model errors. But experienced data scientists are now encountering the double descent phenomenon, which challenges traditional expectations of model performance. Learn more from @fmnobar's latest article.

https://t.co/E5OpT4R058

[Towards Data Science @TDataScience]

🌟 Start the year on the right foot with our hand-picked resources for early-stage data science and AI learners. Dive into articles from industry experts and level up your skills in 2025.

🖊️ by @BM_DataDowntime, @helloiamleonie, @eyalkazin, @fmnobar, @ubersturm, @MayaMurad0 + more

https://t.co/SWN2N7yS30