iiinova — where your digital self keeps evolving

Reposted by Cameron Holmes

Julian Minder @jkminder.bsky.social · 20/10/2025

New paper: Finetuning on narrow domains leaves traces behind. By looking at the difference in activations before and after finetuning, we can interpret what it was finetuned for. And so can our interpretability agent! 🧵

151

Cameron Holmes @cameronholmes.bsky.social · 01/01/2025

I think it's pretty hard to disentangle them really, I was initially skeptical of the (very convenient) argument from the labs about them not being orthogonal, but I'm increasingly buying it.

180

Cameron Holmes @cameronholmes.bsky.social · 01/12/2024

It would have been an all-too-convenient refrain for the "I don't believe in that sci-fi nonsense" AI Safety scepticism line

120

Cameron Holmes @cameronholmes.bsky.social · 24/11/2024

She shall know your ways as if born to them

040

Cameron Holmes @cameronholmes.bsky.social · 23/11/2024

Truly an excellent milestone. Although concessions do follow in (incorrect) episode preferences. Sleepytime falling out of favour was crushing.

000

Cameron Holmes @cameronholmes.bsky.social · 23/11/2024

Yeah I try to follow a similar approach

020

Cameron Holmes @cameronholmes.bsky.social · 23/11/2024

The fora draw a clear distinction between upvotes / agreement votes so I think the culture of upvoting contributions stems from that maybe?

000

Cameron Holmes @cameronholmes.bsky.social · 23/11/2024

Bizarre that was included in the screenshot, doesn't seem like it belongs to the same class as the others at all.

010

Cameron Holmes @cameronholmes.bsky.social · 21/11/2024

Now grappling with whether I'd be in that group or not.

120

Cameron Holmes @cameronholmes.bsky.social · 21/11/2024

Not everyone, then we'd have to read them. If only those inclined to make one did then the mere existence of the doc would probably clarify 90% of scenarios. "Oh they've got one of those docs, we are probably cool"

120

Cameron Holmes @cameronholmes.bsky.social · 21/11/2024

I've been really noticing autumn this year

140

Cameron Holmes @cameronholmes.bsky.social · 17/11/2024

Hoping this generalizes into alignment research

140

Cameron Holmes @cameronholmes.bsky.social · 13/11/2024

The HS2 bat tunnel is even worse value for money once you factor in the updated direct cash transfer effectiveness estimates.

030

Cameron Holmes @cameronholmes.bsky.social · 13/11/2024

Oh yes, an entirely intentional error on my part in the spirit of #DAW 😬

030

Cameron Holmes @cameronholmes.bsky.social · 13/11/2024

deleting dating apps because i want to meet someone the old fashioned way (we caught a wild pig together while not sharing a common language, then met 12 years later under the tree we planted)

020

Cameron Holmes @cameronholmes.bsky.social · 13/11/2024

This is a Draft Amnesty Week #DAW draft. It may not be polished, up to my usual standards, fully thought through, or fully fact-checked.

000

Cameron Holmes @cameronholmes.bsky.social · 13/11/2024

FYI it's Draft Amnesty Week on TPOB, where users can publish scrappy, draft-y, or incomplete posts with impunity. #DAW

430

Cameron Holmes @cameronholmes.bsky.social · 13/11/2024

Circles, but it's just a different app for each emigrating TPOT

040

Cameron Holmes @cameronholmes.bsky.social · 09/11/2024

On new beginnings: This week I handed in my notice, ending 10 years in Product Management, capital markets to start as an Alignment Research Manager in January! 🎉

170