For me, it's relaxation. I sit in a comfy chair beside big picture windows overlooking a canyon and the fjord, I give it an instruction, and then just enjoy the view until it beeps at me.
Nafnlaus 🇮🇸 🇺🇦 🇬🇪
@nafnlaus.bsky.social
6030 Followers
712 Following
Mastodon: @nafnlaus@fosstodon.org Twitter: @enn_nafnlaus URL: https://softmaxdroptableartists.bandcamp.com/
#Energy #EVs #Ukraine #AI #Horticulture #Research
-
-
-
Does Not Compute.
-
I love the concept of turning a quarterly financial call into a "Are you considering supervillainy?" test ;)
-
How relevant it is... 🤷♀️ I guess it depends on what you're trying to achieve. In general, finetuning (at least in my experience) doesn't affect deeper layers much; it's rather superficial.
-
Before the advent of abliteration, finetuning of finetunes which had no publicly-released base models was the primary way to uncensor them. 🤷♀️
That said, I certainly can see the advantage to adding in some Fineweb training before retraining. Many epochs on small finetune datasets is of course bad.
-
<ThisIsFine.gif>
-
But you can also just retune a text model; you don't have to work from a base. The data is, as they note, still in the weights.
-
So they're trying to turn a useful chat model into a non-finetuned text completion model?
I'm not getting the application here...
-
X also uses deliberately misleading metrics. Take their video metrics, for example:
236 million "views" on said video is just "how many timelines they put it on". Only 14,8 million actual views, and even the latter is only "played at least 2sec, at least half visible"
No, Tucker Carlson's Trump interview doesn't have 230 million video views on X
Here's how many times it has actually been viewed.
