I’ve found this to be the case a lot, too. I also spoof my OS because a lot of government sites will refuse to work unless it says Windows. It’s stupid, but here we are.

I got a laptop back in 2018, and it shipped really fast. It’s not my daily driver, but it works well when I’m on the road, and the battery life is pretty good. Granted, I replaced the OS with a distro I prefer and customized the hell out of it, so that might contribute to my experience. Tbh, I was pretty impressed with it (still am), and I was going to buy a Librem 5 when they came out. I wanted to wait and not just throw money at them because I didn’t want to get burned. After all the horror stories and crap reviews, I passed on that and won’t touch the company with a 10 foot pole, and I thank past me for not throwing money at them.

I think that the company started with noble intentions and made a decent product at first, but they got in way over their heads and now they’re floundering.

The original paper itself, for those who are interested.

Overall, this is really interesting research and a really good “first step.” I will be interested to see if this can be replicated on other models. One thing that really stood out, though, was that certain details are obfuscated because of Sonnet being proprietary. Hopefully follow-on work is done on one of the open source models to confirm the method.

One of the notable limitations is quantifying activation’s correlation to text meaning, which will make any sort of controls difficult. Sure, you can just massively increase or decrease a weight, and for some things that will be fine, but for real manual fine tuning, that will prove to be a difficulty.

I suspect this method is likely generalizable (maybe with some tweaks?), and I’d really be interested to see how this type of analysis could be done on other neural networks.