Clive Chan configures Telegram Codex on home server
Clive Chan configured Telegram Codex on a home server and reached 99 percent completion on remote access while setting PasswordAuthentication to no. The Codex interface displayed a Code Bot conversation with post-fold fine-tuning recommendations at n8000 and a results table showing metrics of 1.250 for BN folded and 1.098 for post-fold FT at lr 4e-4. Researcher Lucas Beyer replied cautioning against reviving batchnorm. The exchange addressed the bot's handling of model training checkpoints and metrics.
@itsclivetime Don't you even dare reviving batchnorm
Clive Chan@itsclivetimegot telegram codex set up on my home server codex remote is 99% of the way there but i just want `PasswordAuthentication no` :(
@itsclivetime Ok just a word of caution if you ever want to scale up: removing BN was key in BiT (not visible enough in the paper, but we had some slides) and also in NFNet.
Clive Chan@itsclivetime@giffmana i’m sorry i don’t want to implement a floating point divider in rtl 😭
#559Clive Chan@ITSCLIVETIMEgot telegram codex set up on my home server
codex remote is 99% of the way there but i just want `PasswordAuthentication no` :(
@giffmana i’m sorry i don’t want to implement a floating point divider in rtl 😭
Lucas Beyer (bl16)@giffmana@itsclivetime Don't you even dare reviving batchnorm