I got the Codex $100 plan. It’s great at following instructions and it’s fast, but continues to struggle as the code base tends to get larger (where CC excels at) and it consumes tokens a heck of a lot faster. I’m this close 🤏 to upgrading but @OpenAI gotta solve these problems.
This 17 year old from Jharkhand did more journalism sitting at his home than the entire Indian Media combined did in 12+ years.
This is inspiring stuff, Sarthak may have fixed CBSE forever.
Legendary stuff, Must Watch 👏
Not quite as PocketOS level, but encountered this on a rather large codebase I’m working on with Claude Code today and it shook me even more reading this article. @AnthropicAI@ClaudeDevs
Love the @loops platform but was bummed they didn't have an official python library. Threw their entire documentation on Codex, which did an okay-ish job one shot, then Claude Code took it home with tests and some of the weird bugs Codex left.
https://t.co/mPvcF2qIVB
@soldni@staghado Perhaps. But enterprises don’t really care much about public benchmarks, they care more about benchmarks on their own data/documents, and I’ve seen even stellar models with excellent public benchmarks fail terribly on customer use cases way too many times.
@staghado@soldni Have worked with 200+ enterprise customers on Document Processing/Understanding use-cases. None of them use any of the open source models you list, not because they don’t want to use them but because self-hosting these models and scaling for inference is a hassle.
@soldni PaddleOCR is not that great compared to Textract or Google Doc AI. Haven’t heard much about the rest of them you mention vs. almost always heard about the models being used for OCR at enterprises. Mistral’s comparison is strategic, since they want to sell to enterprises.
It’s been a while since I’ve joined this platform but I don’t see much value here anymore. The algorithm seems to be perpetually broken, posts click-baity, overall very boring, a buggy app, and the list goes on.
“How do you build secure, reliable, scalable systems from scratch?”
Answer: You don’t. 😂
The entire digital economy sits on a precarious stack where if Linux boots and the power grid doesn’t fail, we call it a success.
A small audio model launch --
gpt-4o-transcribe-diarize
This is a diarization-focused ASR model, it's big and slow so we recommend running it offline, but it excels at differentiating speakers, and you can provide voice samples for known speakers up front.
I’m not gonna lie but Claude Code is just not cutting it. ESPECIALLY, it’s buggy command line interface when threads get long. Tried Co-Pilot with GPT 5 and surprisingly pleasant experience snappy, accurate code, and solid instructions following. Probably going to switch to Codex
@elonmusk Hey @elonmusk been trying to subscribe FSD for a while now from the app, but it keeps erroring out. Tried 3 payment methods so far. What gives?