Sarthak Karandikar

e/acc | Into tech + science + philosophy

[Sarthak Karandikar @_itsskofficial_]

I am into AI cause I want to find out whether consciousness is emergent or always existing

[Sarthak Karandikar @_itsskofficial_]

@EvaFox The weekend engine detail is the best bit of this

[Sarthak Karandikar @_itsskofficial_]

@adxtyahq Do the labs run in the browser or do I need my own GPU for the inference ones?

[Sarthak Karandikar @_itsskofficial_]

@tgod34748 Imagine paying that much to watch a man snap it in half

[alphaXiv @askalphaxiv]

"A Fix for Softmax and Cross-Entropy"

Turns out two of the most fundamental functions in deep learning have numerical bugs that can actually destabilize training.

This paper proposes simple fixes to softmax and cross-entropy that prevent tiny gradients from incorrectly rounding to zero, or remaining nonzero when they should vanish, as these errors can cause sudden loss spikes and interfere with knowledge distillation.

On small-scale grokking benchmarks, the fixes eliminate Slingshot loss spikes and improve test accuracy from 94.3% to 100% in one setting. They also prevent student models from drifting away from their teachers during distillation, without requiring higher numerical precision.

https://t.co/0nXRLrQVbe

[Sarthak Karandikar @_itsskofficial_]

@kaavyya This is so real, and the small work never feels like it counts while you are doing it

[Sarthak Karandikar @_itsskofficial_]

@astrobase_space Seeing a propulsion team hiring out of India still surprises me every time, in a good way

[Sarthak Karandikar @_itsskofficial_]

@amorriscode Claude Code on desktop runs most of my week now

[Sarthak Karandikar @_itsskofficial_]

I did all of this in the wrong order
Started my own things first, joined someone else's early stage company much later, and spent the first month realising how much I had been confidently guessing about for years

The advice everyone repeats is to just start your own thing, because nobody learns to swim from the side of the pool
I believed that completely. I now think it is right about the destination and wrong about the route

When you only ever run your own company, you learn one company deeply
That sounds like a lot until you notice it is a sample size of one. One market, one set of customers, one way of making decisions, and no way to tell which lessons generalise
You also see your own failures from the inside, where they are hardest to read. Something goes wrong, you can give four plausible explanations for it, and you have no idea which one is correct
No control group, no second run with one variable changed

The things that surprised me inside someone else's company were not strategic at all, they were procedural
How a decision gets made when two senior people disagree and both are being reasonable
What a customer sounds like when they are not being polite, which is nothing like the calls you have when you are the founder and everyone in the room knows it
I watched people decide things in ten minutes that I would have agonised over for a week, not because they were smarter, but because they had seen it before and knew which bits did not matter

Then the cost argument, which is the one that actually got me
Make a mistake in your own company and you pay with your runway and two years of your life
Make the same mistake inside someone else's and you learn it while they absorb the bill
Most people who turn that down are turning it down on status, not economics. Founder just sounds better than employee

Two or three years is plenty though
Past that you are doing year one again with nicer tools

[Sarthak Karandikar @_itsskofficial_]

@atulit_gaur Random vectors going perpendicular is the one I still cannot picture, is there any intuition for it that is not just algebra?

[Sarthak Karandikar @_itsskofficial_]

@willdepue Been there, and the logs were mostly apologies

[Sarthak Karandikar @_itsskofficial_]

@amritwt Sleep is the one I keep negotiating with

[Sarthak Karandikar @_itsskofficial_]

@AndrewCurran_ Is the skepticism about the models or about wet labs being slow

[Sarthak Karandikar @_itsskofficial_]

@KlausCodes The old phone photo hitting harder than the move itself

[Sarthak Karandikar @_itsskofficial_]

@YufeiWang25 Did being in robot learning narrow where you could even apply, or did it open doors

[Indian Tech & Infra @IndianTechGuide]

🚨 Indian physicist Ashoke Sen has become the first Indian to win the 2026 Oskar Klein Medal for his pioneering contributions to string theory, black-hole physics and quantum gravity.

[Sarthak Karandikar @_itsskofficial_]

@jain_harshit Start with the smaller add on, if you never open it you saved the rest

[Sarthak Karandikar @_itsskofficial_]

@bepituLaz The mocking step is the only one I have fully mastered

[Sarthak Karandikar @_itsskofficial_]

Half the AI tools I try are just a prompt in a trench coat

I do not mind it, the prompt is usually better than mine

What bugs me is the pricing page acting like it invented reasoning

[Sarthak Karandikar @_itsskofficial_]

@mannupaaji CORS never gives you a useful error

[Sarthak Karandikar @_itsskofficial_]

@Pirat_Nation Explains why my 4060 suddenly feels like a flex