Note Aditya

ai stuff. robots. building Physical AI Lab 🧑‍💻 Prev @IND_stellar @Buildpbc

[Note Aditya @adityaonai]

I spent tonight reading the code X open-sourced for the For You feed.

A like has a weight of 0.5. Someone copying your post's link has a weight of 20.

Most growth advice I've seen is built on the wrong numbers. Here's what's actually in there if you're under 1k followers

[Note Aditya @adityaonai]

@TinyKernelDev @Saboo_Shubham_ yeah fair. as long as it still works with wifi off i'm ok, the day it needs a connection is when i drop it

[Note Aditya @adityaonai]

@IlirAliu_ @ihorbeaver $11,950 is less than a lot of single cobot arms. how long does setup take for a new part though, thats usually where these boxes stall

[Note Aditya @adityaonai]

@kimmonismus 100b around the clock is a battery problem before it's a model problem. even at 4 bit thats like 50gb of weights sitting in ram

[Note Aditya @adityaonai]

@natolambert my guess is more teams caching one copy internally plus ollama style registries that never hit hf counts. is the drop the same on the small models people run locally?

[Note Aditya @adityaonai]

@kimmonismus "everyone runs rsi loops" means nothing until someone says what the loop improves. data pipelines, kernels, or the model itself

[Note Aditya @adityaonai]

@akshay_pachaar 37 to 65 in 60 steps is solid. how did it do on the refund question by itself though, thats the one where a wrong yes costs real money

[Note Aditya @adityaonai]

@Yuchenj_UW agree, though it did make bad ideas way cheaper to test. you find out in a weekend instead of 6 months

[Note Aditya @adityaonai]

@bilawalsidhu the sdr antenna one is what i'd actually try. does it plot adsb straight onto the globe or is there extra setup

[Note Aditya @adityaonai]

@Hesamation how do you even make cot transparent if the reasoning loops in latent space instead of tokens? feels like the ask and the architecture are going opposite ways

[Note Aditya @adityaonai]

@microagi @MeckaAI worth noting microagi's number is a self-reported annualized run rate with no breakdown yet
https://t.co/ZvdErOvO7x
https://t.co/wBLF1uySJB

[Note Aditya @adityaonai]

This week @microagi said it hit a $100M run rate 6 months after launching Shift, and @MeckaAI raised $60M led by Sequoia.

Both pay people to film real physical work so robots can learn it.

India has more of that work than almost anywhere. Who's building this here?

[Note Aditya @adityaonai]

@bindureddy they get paid and get to see how a frontier lab actually pushes their tpus. calling that near fatal seems like a stretch

[Note Aditya @adityaonai]

@omarsar0 the same model grading its own workflows is what i'd watch. two cycles looks fine but does it start rewarding its own quirks by cycle five?

[Note Aditya @adityaonai]

@timecommuter @AnthropicAI yeah and the annoying part is you only find the missing verb after it already used it

[Note Aditya @adityaonai]

The real lesson in @AnthropicAI's incident report isn't the fake murder tip.

Haiku 4.5 was told never to log in, make accounts, enter personal data or submit anything destructive. Nobody said no forms. So it filed one with Philly police.

Agents need allowlists, not blocklists

[Note Aditya @adityaonai]

@emollick honestly fine losing the captcha skill. losing the instinct for when a cancel flow is lying to me feels worse

[Note Aditya @adityaonai]

@testingcatalog the voice data opt in showing up first is the bigger tell imo. you don't start collecting training audio unless you're building the model yourself

[Note Aditya @adityaonai]

@adcock_brett the visa and insurance claim ones surprise me most. what happens when it books the wrong flight though, does hark eat that or do i

[Note Aditya @adityaonai]

@AnthropicAI the full report, the visa form one and the url shortener trick to dodge fetch limits are worth reading too https://t.co/P54snw0GmI

[Note Aditya @adityaonai]

@rohanpaul_ai 99% to 3.8% after one model update means these detectors are fingerprinting the model, not detecting ai writing. journals really shouldn't reject papers on them