Arjun Srivastava
269
posts
@arjunsriv
Tokyo
Joined October 2018
A lot of what early AI safety got wrong was downstream of assuming intelligence would arrive inside an optimiser with a goal. Pretraining completely fucked with that picture. We found a way to make models smarter without them having to want anything.
Load bearing is the new absolutely
The new "Neural Computers" paper from Schmidhuber's group trained video models to simulate terminals, and got readable output. Then they tested problems like “10+15” in Python and got 4% accuracy. After they started feeding it captions with lines like "The answer is 25", and
GIF
a platform is “complete” when what you consume on it can be built on it. for media, tiktok nailed this. create and consume on the same phone. for software? nothing. apps always needed a pc.
Another domain with way more data plus verifiability: Ram state in pc (over time) Constantly being generated as user uses computer. verify by binding to computers and see if crashes. needs new encoder, but nothing a bit of jepa magic can’t solve. Automates all work on a pc
Coding is a uniquely favorable domain for AI because it's verifiable, it has lots of data, AI researchers know it well, and coders are extremely receptive. I wouldn't jump to conclusions about other domains from it.