TheCodeEyeAI

Engineer // Building & Learning // Views are personal // DM for collabs ✉️

[TheCodeEyeAI @TheCodeEyeAI]

Gemini 4 Argon wins 13 of the 19 benchmarks in Google's own chart.

And loses Terminal-Bench 4.0 by 9 points.

57.4 vs Opus 5.5's 66.4.

That is the coding benchmark everyone has been quoting all year.

Also: 1M output tokens, $2/$10 intro pricing.

https://t.co/Tp4eFYO86q

[Striver | Building takeUforward @striver_79]

team outing 🫶

- engineering
- sales
- user experience
- product
- design
- marketing
- content

[TheCodeEyeAI @TheCodeEyeAI]

Should a founder-CEO step aside during an investigation like this, or stay at the helm?

Reply with your take.

[TheCodeEyeAI @TheCodeEyeAI]

Nvidia's board just opened an investigation into its own CEO.

The question: did Jensen Huang's affair with a chip executive, alleged to be a Chinese intelligence asset, put anything at risk?

NVDA is down ~5% premarket.

What's confirmed vs. alleged ↓

[TheCodeEyeAI @TheCodeEyeAI]

The real lesson for founders:

Your biggest single point of failure isn't always a chip, a supplier or a model.

Sometimes it's the person at the top.

For 30+ years, Nvidia and Jensen have been the same story. That's the edge. It's also the risk.

[TheCodeEyeAI @TheCodeEyeAI]

Hello everyone. Happy Diwali!

[Postey AI @posteyai]

I connected Claude to my social accounts and it now runs my entire social media

I typed one sentence.

Claude wrote the posts, adapted each one for every platform, and published them.

X → done

LinkedIn → done

Instagram → done

Threads → done

no copy-pasting

no switching tabs

no scheduling tools open

it took me less than 2 minutes

watch it happen in real time 👇

content isn't hard anymore.

distribution was the bottleneck.

so I fixed it.

now Claude doesn't just write my content,

it ships it everywhere.

comment "CLAUDE" and I'll send you the setup.

[Shruti @heyshrutimishra]

City hall's newest employee runs on an NVIDIA GPU.

Malaysia's ITMAX is rolling out AI vision agents built to cover 17,000+ traffic cameras by 2027.

It tests signal timing in an open-source traffic simulator, and NVIDIA says peak-hour corridor travel times are down by up to 26%.

In Spain, 1MillionBot's agents handle municipal procedures, appointment requests and resident questions across 150+ deployments, including Madrid's regional government.

In Taiwan, Linker Vision and the Dutch architects MVRDV simulated 15 years of urban forest growth in 10 minutes. Each design iteration went from two weeks to under an hour.

In France, the Côte-d'Or department put agents across 500+ public buildings. Document search went from weeks to minutes, and operating costs fell 10%, with 30% the target.

Every number here is reported by NVIDIA and its partners, not independently audited.

Still, the direction is clear. Cities already own thousands of cameras. Agents are what turn that footage into decisions.

[TheCodeEyeAI @TheCodeEyeAI]

🔥🔥🔥🔥🔥

[Sabiee @whipofwrist]

Probably the only instance where you’ll get to hear Kohli’s original voice singing our National Anthem 🥹

[TheCodeEyeAI @TheCodeEyeAI]

@nirudhuuu Agar mere ko bhi Claude subscription mil jata 😋😋😋

[TheCodeEyeAI @TheCodeEyeAI]

@piyushgarg_dev one question @piyushgarg_dev sir, who is going to pay for this order 😁??

[TheCodeEyeAI @TheCodeEyeAI]

@piyushgarg_dev @letsblinkit Hi, I have created a cart with 1 item worth ₹179. Please review and make the payment to place the order on Blinkit.
https://t.co/kDLxpqoGb4

[TheCodeEyeAI @TheCodeEyeAI]

@piyushgarg_dev Hi sir, I have created a cart with 1 item worth ₹179. Please review and make the payment to place the order on Blinkit.
https://t.co/kDLxpqoGb4

[Shruti @heyshrutimishra]

Everyone’s hunting for the next side hustle.

But most people are bleeding hundreds a month on charges and don’t even know it.

Before you make more, keep more.

Here’s how to find the leaks in 2 minutes

[Shruti @heyshrutimishra]

Dyna went with a video about craft and the people behind Taku, not a spec sheet.

tells you a lot about who they want building with them

[Shruti @heyshrutimishra]

Benchmark results and accuracy vs document length, from Baidu's report.

[Shruti @heyshrutimishra]

1,200+ pages of the original Chinese Three-Body Problem trilogy, turned into text in a single pass.

That's what Baidu's newly released Unlimited OCR 2 just accomplished.

Page count isn't the hard part. Any OCR tool can go page by page. The hard part is reading a long document in one pass while keeping track of relevant information from earlier pages, so a table that runs onto the next page or numbering set hundreds of pages back stays consistent. That's where AI document parsers usually degrade.

In Baidu's report, MiniMax-M3 fails past about 223 pages, and Gemini 3.1 Pro and Qwen3.8-Flash-Next lose accuracy as documents get longer. Unlimited OCR 2 stays relatively stable. On MarathonDocBench, built from documents hundreds of pages long, it scored 77.1. The next best model in the comparison scored 33.7.

The trick is Sparse Reference Sliding Window Attention (SR-SWA). It works like a person transcribing a book: eyes on the current page, looking back at the most relevant earlier pages. So the cost of each decoding step barely grows with length. At 100 pages it was 7× faster per step than full attention.

The first version already has 10M+ downloads on Hugging Face and 28K stars on GitHub.

What I still want to see: independent runs on messy scans, handwriting and mixed-language PDFs. These numbers come from Baidu's report, and a printed novel is a clean source.

If it holds up, contracts, filings and data rooms could be processed in a single run instead of having to be split into separate documents.

[Shruti @heyshrutimishra]

AI agents don't have a capability problem. People just don't know what to ask them.

Now you can steal a prompt that already works.

@wajo_ai lauched Fompt Feed, where people share the prompts they use daily and Fo runs them for you: job alerts, missed emails, money you forgot to claim.