Saurabh Kumar

CTO & Co-founder, @360LabsOffical
12 years of code • ∞ problems solved
If you sell vaporware for a living, you probably won't like it here.

[shirish @shiri_shh]

Figma CEO watching 14 years of work get cooked by one tweet 😭

[Saurabh Kumar @NotSaurabhKumar]

Hot take:
IEMs destroy every pair of AirPods Pro I’ve ever owned.

Better soundstage, punchier bass, cleaner vocals, no dropouts, and they don’t die at 40%.

AirPods Pro are expensive wireless earrings.
IEMs actually sound like music.

Change my mind.

[Saurabh Kumar @NotSaurabhKumar]

Every AI consulting firm sells you a 6-month "discovery phase."

we built and shipped 6 products in the last 10 days.

one of them ships to adidas next week.

#AI #AIConsulting #BuildInPublic

[Saurabh Kumar @NotSaurabhKumar]

The real lesson for every builder:

Your moat isn’t your brand. It isn’t your funding. It isn’t your headcount.

Your moat is: DOES IT WORK?

Intel proved that even $70B in revenue means nothing if your product needs 47 steps to set up.

Ship products. Not slide decks.

[Saurabh Kumar @NotSaurabhKumar]

The most insane fumble in tech history isn’t Blackberry. It isn’t Yahoo.

It’s Intel.

In an era where EVERYONE wants chips, Intel can’t even give theirs away.

Let me explain why this should terrify every “big company” founder: 🧵

[Saurabh Kumar @NotSaurabhKumar]

Intel had:
∙The fabs
∙The engineers
∙The government contracts
∙The brand
And they STILL lost. Because they built org charts, not products. They shipped roadmaps, not software.
This is exactly what happens to “consulting” companies that sell slides instead of building things.

[Saurabh Kumar @NotSaurabhKumar]

OpenAI Codex doesn’t even support Intel Macs.

Downloads an ARM-only app. Installs it. Fails silently.
No warning. No error message. Just vibes.

Intel is so irrelevant that other companies don’t even bother telling you they’ve dropped support.

[Saurabh Kumar @NotSaurabhKumar]

Want to run AI models locally? Cool.

Nvidia: install ollama, done.

AMD: install ollama, done.

Intel: install conda, set up IPEX-LLM, configure oneAPI, set 7 environment variables, run a Docker container, cross your fingers.

This is what happens when committees build products.

[Saurabh Kumar @NotSaurabhKumar]

Nvidia prints $90B from data center chips.

Intel’s Gaudi AI chip? Couldn’t hit $500M because, and I’m not making this up - the software was “unbaked.”

HALF the price of Nvidia. Double the inventory sitting in warehouses.

Nobody wants it. Because it doesn’t work.

[Saurabh Kumar @NotSaurabhKumar]

@davidonchainx And make no mistakes

[Saurabh Kumar @NotSaurabhKumar]

@slash1sol Ah yes, the annual tradition: nerf the flagship model right before launch so the ‘slightly better’ one feels like a revolution.

[Saurabh Kumar @NotSaurabhKumar]

https://t.co/VHxCcPwFDx lmk if you find all the easter eggs

[Saurabh Kumar @NotSaurabhKumar]

Here's where it gets interesting.
Gemma 3 4B has 4 billion parameters.
Training all of them? Expensive. Slow. Needs massive GPUs.
The trick: LoRA (Low-Rank Adaptation).
Freeze the original 4B parameters. Add small adapter layers. Train only those.

[Saurabh Kumar @NotSaurabhKumar]

I just trained a medical AI for $10.
Not $10,000. Not $100,000.
Ten dollars.
Here's exactly how I did it 🧵

[Saurabh Kumar @NotSaurabhKumar]

Next decision: which model to fine-tune?

Llama 3.2 1B → too small, basic responses
Gemma 2 2B → better, but wanted more
Gemma 3 4B → sweet spot

Open source. Free. Good at following instructions. Fits in GPU memory.

[Saurabh Kumar @NotSaurabhKumar]

Linus Torvalds vibecoding a Python tool.
Difference between him and most vibecoded projects?
He knows what the code should do BEFORE AI writes it.
Vibecoding without expertise = expensive hallucinations.
#vibecoding #aidev #buildinpublic