Jai
@jaiwithani.bsky.social
This is Jai. Almost no one is evil almost everyone is broken.
Trying to figure out what I want to be when I grow up for 38 years.
Trying to figure out what I want to be when I grow up for 38 years.
I didn't follow the reasoning here. Why wouldn't a physicalist account of consciousness be connected to these?
December 5, 2025 at 2:38 PM
I didn't follow the reasoning here. Why wouldn't a physicalist account of consciousness be connected to these?
Okay yes, but also classic sci-fi: AI will speak in incomprehensible verse briefly punctuated by actual commands and meaningful output
O3 CoT:
O3 CoT:
September 30, 2025 at 10:06 AM
Okay yes, but also classic sci-fi: AI will speak in incomprehensible verse briefly punctuated by actual commands and meaningful output
O3 CoT:
O3 CoT:
Here are some trees I saw one time
March 11, 2025 at 6:39 PM
Here are some trees I saw one time
Claude has escaped containment (Mt Moon)
March 2, 2025 at 8:18 AM
Claude has escaped containment (Mt Moon)
Today I started in earnest on solving the hardest problem in AI Safety, untangling the massive tangle of cables at the London Initiative for Safe AI
January 14, 2025 at 9:47 PM
Today I started in earnest on solving the hardest problem in AI Safety, untangling the massive tangle of cables at the London Initiative for Safe AI
My house has a lot of dietary restrictions. But I still managed a pretty good Beef Wellington without any eggs, dairy, sugar, or beef. Homemade puff pastry and all.
December 26, 2024 at 5:08 AM
My house has a lot of dietary restrictions. But I still managed a pretty good Beef Wellington without any eggs, dairy, sugar, or beef. Homemade puff pastry and all.
I guess this is basically what the frequency-based-n-gram-embeddings approach they show in appendix D is doing - which even they find slightly-but-consistently outperforms the hashes for a given vocab:
Hash embeddings are probably just adding noise from token collisions, right? Drop 'em.
Hash embeddings are probably just adding noise from token collisions, right? Drop 'em.
December 19, 2024 at 4:09 AM
I guess this is basically what the frequency-based-n-gram-embeddings approach they show in appendix D is doing - which even they find slightly-but-consistently outperforms the hashes for a given vocab:
Hash embeddings are probably just adding noise from token collisions, right? Drop 'em.
Hash embeddings are probably just adding noise from token collisions, right? Drop 'em.
New alignment research agenda just dropped
December 18, 2024 at 7:15 PM
New alignment research agenda just dropped
Sweden, welcome to NATO
March 9, 2024 at 4:47 AM
Sweden, welcome to NATO