Vinit Mehta

automating everything except art, poetry and sunsets
growth at @nanonets agents

[Nanonets @nanonets]

Our new product, Atlas is live on Product Hunt today. 🔥

If you're a builder, you and your whole team are likely struggling to maintain one single company context across Claude, GPT, your sales tool, your CRM.

We built Atlas because that's just a stupid problem to still have in the big 2026.

Atlas basically maps out everything that makes your company yours, your tone, how you work, what you sell, how you position it, and stores it as a context graph that any AI tool can pull from.

One source of truth for the whole org. So you and your team can switch tools, systems, workspaces whenever you want and nothing resets.

P.S. First 200 teams get white-glove onboarding at just $99/month.

[Vinit Mehta @winitmehta]

@BhavyamArora6 Waah tujhe kaise pata ye sab

[Vinit Mehta @winitmehta]

@nanonets Really exciting. Everyone building doc heavy workflows should use this!

[Nanonets @nanonets]

Nanonets OCR-3 is live.

This is the most accurate OCR model in the world currently.

87.4 on OLM-OCR (Global #1)
85.9 on IDP Leaderboard (Global #1)
90.5 on OmniDocBench

OCR-3 also ships with two critical features that foundational models and VLMs miss today - confidence scores and bounding boxes.

[Vinit Mehta @winitmehta]

https://t.co/Gz3vAoOaoe

[Vinit Mehta @winitmehta]

BREAKING: Claude can now replace a $200,000/year analyst.

Here's what it does better than your entire junior team:

> Builds DCF models from scratch
> Runs full financial reconciliation
> Writes investor memos and board reports
> Generates variance analysis with commentary

One prompt. Everything under 10 minutes.

I've written a full guide on all 4 workflows - exact prompts, step by step, completely free. Find it below ↓

[Vinit Mehta @winitmehta]

https://t.co/QRyF0SOT6R

[Vinit Mehta @winitmehta]

https://t.co/fcRA1LAZ2H

[Vinit Mehta @winitmehta]

https://t.co/xVAZi3AXoZ

[Vinit Mehta @winitmehta]

GPT-5.4 just launched this week. And if history is anything to go by, they're about to make their older models a lot worse.

Every time @OpenAI (or @Google or @AnthropicAI ) drops a new model, people notice the same thing: their old model starts behaving differently. Slower. Dumber. Less reliable.

The obvious explanation is: you now have something better to compare it to, so the old one just feels worse by contrast.

That's called prompt drift - the idea that YOU changed, not the model. You got more sophisticated, your prompts evolved, your expectations went up.

OpenAI used this explanation for years. Officially. But there's another explanation. Model drift: where the model itself actually changes, silently, underneath you.

So which one is it? Psychological or real?

Well several reputed org. (like Stanford) and individual users have concluded from that the older models do get dumber.

And post the backlash from multiple power users on X/reddit, labs like OpenAI, Google, and Anthropic have all confirmed cases where their older models degraded after an update. Every single time, they called it unintentional.

But practically, a model cannot degrade on its own. It is INTENTIONAL.

The question is: why are labs not being honest about it?

I wrote a full breakdown - prompt drift vs model drift, what the research actually says, what official documents and what our studies reveal. Link in replies

[Vinit Mehta @winitmehta]

https://t.co/wrFZE6WOiL

[Vinit Mehta @winitmehta]

The program that built the atomic bomb was kept secret for 3 years but today, the most crucial arms race is playing out in the press and twitter publicly.

Last week, the US and Israel launched Operation Epic Fury against Iran. 900 strikes in 12 hours. A pace that would've taken days, maybe a week, in any conflict before this decade.

The thing that made it possible was AI fusing drone feeds, satellite imagery, and telecom intercepts faster than any human team ever could. Compressing weeks of targeting work into a single morning.

I think at times when every country has arms and nukes. This is the new unfair advantage.

So I went down a rabbit hole on this and wrote an insane breakdown of the whole thing that includes:

Where does every country participating in wars stand in terms of AI arms, situation of China and can it dethrone the US in this race anytime?

Link in the replies.

[Vinit Mehta @winitmehta]

https://t.co/3wjq5tzQ64

[Vinit Mehta @winitmehta]

unpopular opinion: GPT/gemini/claude are terrible choices for document parsing bcs they're too good at everything else

the training budget is finite. when it goes to reasoning and coding, it doesn't go to maintaining granular OCR accuracy on edge cases. so invoices, lab reports, insurance claims - they become collateral damage in someone else's roadmap.

a fine-tuned model purpose-built for your document type doesn't have this problem. it doesn't know how to write code or reason about philosophy. it just knows your forms, and it knows them reliably.

plus a system processing 10k documents daily costs approximately $5k annually for on-premises deployment versus $50k for frontier inference.

Link for the full breakdown in the replies.

[Vinit Mehta @winitmehta]

https://t.co/d4iPeizdDR

[Vinit Mehta @winitmehta]

I spent days trying to figure out what Claude did during the Venezuela raid that took down Maduro.

The Pentagon won't say. Anthropic won't say. Reuters couldn't verify it.

So I dug through everything public and made my best case. Please give it a read

[Nanonets @nanonets]

https://t.co/VVw260v4Lw

[Vinit Mehta @winitmehta]

@meenaji7428 hello 👋

[Vinit Mehta @winitmehta]

@impeculiar1b @winitmehta

[Vinit Mehta @winitmehta]

@civilianweb3 hello 👋

[Vinit Mehta @winitmehta]

@0xmim9 hello 👋