CCSLM by Sean O'Connor https://padlet.com/seanhaddps/s02446xe3vf4hlm61ukv en-us 2026-10-02 04:09:49 UTC 2026-10-02 09:38:33 UTC hello@padlet.com CCSLM seanhaddps https://padlet.com/seanhaddps/s02446xe3vf4hlm61ukv/wish/4080068849 Neural networks that are:

Linear all the way through on the data channel.

Non-linear via the context channel.

]]> 2026-10-02 04:14:26 UTC https://padlet.com/seanhaddps/s02446xe3vf4hlm61ukv/wish/4080068849 Expert Alignment seanhaddps https://padlet.com/seanhaddps/s02446xe3vf4hlm61ukv/wish/4080079017 The output of one expert and the input of the next are aligned by backpropagation.

SVD allows us to see this.

]]> 2026-10-02 04:25:31 UTC https://padlet.com/seanhaddps/s02446xe3vf4hlm61ukv/wish/4080079017 LSH CCSLM seanhaddps https://padlet.com/seanhaddps/s02446xe3vf4hlm61ukv/wish/4080418319 Using locality sensitive hashing to create experts in concatenated context selected linear mappings.

]]> 2026-10-02 09:30:52 UTC https://padlet.com/seanhaddps/s02446xe3vf4hlm61ukv/wish/4080418319