Kaj Bostrom
@bostromk.net
I hate slop and yet I work on generative models
PhD from UT Austin, applied scientist @ AWS
He/him • https://bostromk.net
PhD from UT Austin, applied scientist @ AWS
He/him • https://bostromk.net
Reposted by Kaj Bostrom
As our main result, we find that when a token is in a model’s vocabulary—i.e., when its characters are tokenised as a single symbol—the model may assign it up to 17x more probability than if it had been split into two tokens instead
June 5, 2025 at 10:43 AM
As our main result, we find that when a token is in a model’s vocabulary—i.e., when its characters are tokenised as a single symbol—the model may assign it up to 17x more probability than if it had been split into two tokens instead
Reposted by Kaj Bostrom
Finally found it!
November 22, 2024 at 6:46 PM
Finally found it!
power's back, now I can resume flailing ("running ICLR rebuttal experiments")
November 21, 2024 at 5:10 PM
power's back, now I can resume flailing ("running ICLR rebuttal experiments")
Reposted by Kaj Bostrom
another starter pack, this time for folks (past & current) from Ai2 (@ai2.bsky.social) 😍
go.bsky.app/Qjyc97J
go.bsky.app/Qjyc97J
November 21, 2024 at 4:10 PM
another starter pack, this time for folks (past & current) from Ai2 (@ai2.bsky.social) 😍
go.bsky.app/Qjyc97J
go.bsky.app/Qjyc97J
EMNLP was so fun miami is interesting
November 17, 2024 at 5:06 PM
EMNLP was so fun miami is interesting