Adrita Mahata

A temporary pattern of consciousness pretending to be ‘Adrita.’ | Building AI that automates the jobs we created to avoid asking what we’re actually for.

[Adrita Mahata @adrixy0]

https://t.co/YLjmHpu42I

[Adrita Mahata @adrixy0]

@AlfieJCarter Agents

[Adrita Mahata @adrixy0]

@genspark_ai Useful Idea

[Chetaslua @chetaslua]

🚨 Google New Image Model

> Instant-ramen (successor of nano-banana)

Ramen is cooked time to serve soon , we will share results as soon as we get hands on it 😉

[Neeraj @neerajjj6785]

Claude:
You've used 90% of session limit

Me instantly:

[ayesha @ayesha_fatiima]

Stripe is a $65 BILLION company??????
SIXTY FIVE BILLION??????
For simply moving money
between bank accounts??????

Notion is a $10 BILLION company??????
TEN BILLION??????
For turning documents
into an all-in-one workspace??????

Calendly is a $3 BILLION company??????
Three. Billion. Dollars?????
For basically letting people
book a time slot????????

Slack is a $27 BILLION company??????
For essentially being
a team chat app????????

And you think your SaaS idea
is too simple?

Think again.

[Higgsfield AI 🧩 @higgsfield]

Unlimited Seedance for 30 days, starting today.

Higgsfield is the only place with real unlimited Seedance, powered by an official @BytePlusGlobal partnership.

Purchase now and get unlimited Seedance for 30 days.

[Nico @ncsgy]

Ethereum can already start preparing accounts for a post quantum world, without waiting for a hard fork.

Today, it would be just 0.07$ .

Further audits incoming. Though I squeezed in a review with Fable before Uncle Sam crashed my party. Verity formal proof included for my lean enjoyers

https://t.co/hfOx08X17Q

[Adrita Mahata @adrixy0]

@theflowgrammer @n8n_io @MrBeast Congrats, gonna be following your next journey 🌊

[Adrita Mahata @adrixy0]

@BratDotAI Yeah,
Maybe 1 out of 10 businesses survives,
Not 1 out of 10 business man

[Adrita Mahata @adrixy0]

@0xIlyy Gemini ofc

[Adrita Mahata @adrixy0]

@ShubhamTotu Research about AEOS 👁️👁️

[Adrita Mahata @adrixy0]

1,000 GB / 80 GB = 12.5 GPUs. That math works only if your batch size is 1, context window is zero, and you want 5 tokens/sec for exactly one user.

Even the core ML team lead at Cerebras (who explicitly benchmarks against Nvidia clusters) notes that it takes a minimum of ~240 to 280 parallelized GPUs just to host and properly run a 1T instance(GPT 4o) in production.

Don’t take my word watch this breakdown on how cluster stitching vs. wafer-scale handles 1T+ models: https://t.co/Kz9VrG5nix

[Adrita Mahata @adrixy0]

You’re conflating basic model storage with the mass parallelization needed for real-time production at OpenAI scale.

A ~1T param GPT-5 equivalent in FP16/BF16 needs ~1-2TB VRAM just for the weights. Add KV cache, activations, and headroom for concurrentt users?

If you just want to run it locally for a few tokens/sec, sure - quantize heavily or 3-4 512GB Mac studios or go wafer-scale like Cerebras.
Their single-chip engines handle trillions of params with insane bandwidth…

Again context matters 👁️👁️

[Adrita Mahata @adrixy0]

One of the most interesting AI agent benchmarks I’ve come across recently:

https://t.co/KmmjU0u75W

[Adrita Mahata @adrixy0]

@Kunalchavan_ Fs

[Adrita Mahata @adrixy0]

i recorded one talking head. that was my entire job here.

research, fact-check, script, captions, the edit.. all of it ran as agents on @papercliping, handing work down the line like a tiny media company.

then the part i can't stop thinking about.. my research agent corrected me. i briefed the science wrong, it pushed back and shipped its version, not mine.

even the edit was code in @HyperFrames_ (open-sourced by @HeyGen). no editing app touched it.

i didn't make a video. i approved a company that did.

[Adrita Mahata @adrixy0]

@aronprins @papercliping Love the paperclip desktop app

[Adrita Mahata @adrixy0]

@sri9s I like Claude models but codex software

[Adrita Mahata @adrixy0]

The next breakthrough might be in “architecture optimisation” apart from scaling laws

[orcus108 @orcus108]

WHAT THE HELL is happening in AI?

A 3B parameter model just put up coding benchmark scores in the same league as Claude Opus 4.5.

3 BILLION.

The weights are on Hugging Face, anyone can test it.

I genuinely don't know if this is a breakthrough or if the benchmarks are broken.