Dev Vidhani

Techonologist.
Currently at Aster Data Systems, Inc.
Previously, at EMC Systems, Inc; Kazeon Systems. Inc. and Sun Microsystems.

[Dev Vidhani @DevVidhani]

@populartourist @ScalaWilliam 262144. Typo, I guess.

[Dev Vidhani @DevVidhani]

@thsottiaux Compact errors in MacOS app

[Dev Vidhani @DevVidhani]

@GabLesperance Awesome work. Will look it up. I had planned on leveraging RLM for sheets analysis (alas - got buried in other priorities).

[Dev Vidhani @DevVidhani]

@thsottiaux Compact in macOS app. Still keeps hanging in compact operation

[Dev Vidhani @DevVidhani]

@malikwas1f Great result. Were you able to run any concurrency tests?

[Dev Vidhani @DevVidhani]

@iotcoi How large context (prompts) and output tokes generated tokens per agent? What is the performace with prefill (prompts) around 150-200K?

[Dev Vidhani @DevVidhani]

@junsandesuka Great. How large context (prompts) and output tokes generated tokens per concurrent reqeust? What is the performace with prefill (prompts) around 150-200K?

[Dev Vidhani @DevVidhani]

@0xCVYH @iotcoi How large context (prompts) and output tokes generated tokens per agent? What is the performace with prefill (prompts) around 150-200K?

[Dev Vidhani @DevVidhani]

@GBminA @jun_song how large context (prompts) for 1K generated tokens? What is the performace with prefill (prompts) around 150-200K?

[Dev Vidhani @DevVidhani]

@thsottiaux @doublemover Thank you.

[Dev Vidhani @DevVidhani]

@vivoplt In my experience, codex finds serious Claude issues 9 out of 10 occasions. Claude - 1 in 10.

But Claude is good for documentation and presentations, etc. once the core of the project is done.

[Dev Vidhani @DevVidhani]

@browser_use @NousResearch All good, but trouble with pop-ups/accept terms. Is that in the pipeline?

[Dev Vidhani @DevVidhani]

@tomcocobrico @Teknium @sudoingX That is correct. Perf is low. And if you higher context (200K+) and 2k+ output, memory wall (273Gbps) kills it further.

[Dev Vidhani @DevVidhani]

@nisten @TheAhmadOsman We noticed the same internally

[Dev Vidhani @DevVidhani]

@TheAhmadOsman Agree it is a great model, but qwen3.5vl is also multimodal (VL) model and better multilingual support.

[Dev Vidhani @DevVidhani]

@NielsRogge @thohne @huggingface @Zai_org I had the OCR issues with the drop from a month ago.

[Dev Vidhani @DevVidhani]

@bnjmn_marie Is NVFP4 quantized models harder to try?

[Dev Vidhani @DevVidhani]

@simonw My experience too. I am using spark selectively for relatively more predictable tasks, otherwise, vanilla codex-5.3 is the go to model.

[Dev Vidhani @DevVidhani]

@karpathy Together, 140k between two of us. A few rtx 6000 pros 😊

[Dev Vidhani @DevVidhani]

@ZixuanLi_ Failed for my case! Invoice pdf.