thanks! we asked the model to generate some synonyms and antonyms (in this case, we have "cold" and "impassive" vs "affectionate" and "sensitive")
Then, we ask the model to behave that way (with a prompt), and store the difference in activations for each pair. Then, a PCA can be used to extract the principal component, giving use the steering vector. We do most of this using the repeng library, and the author goes into a bit more detail on how it's done on her [blog](https://vgel.me/posts/representation-engineering/#How_do_we_...?)
I don't think it's about not using the GPLv3 --- a lot of software is stuck in GPLv2 because the maintainers can't get everyone who contributed (or their estate) to sign off on a license upgrade to the better but incompatible GPLv3. Even the linux kernel is stuck!
But this isn't an "escape hatch" for breaking the GPL; there's a reason the FSF made the "cure clause" explicit in the GPLv3. The explicit language should make life easier and save on court costs not only for corporate lawyers, but even more so for the excellent folks at the Free Software Conservancy, the biggest defenders against GPL violations.
Q: Why didn’t you simply relicense your GPLv2 projects under GPLv3 and let the community benefit from the cure approach that way?
A: In many cases, this would not be possible or practical. For example, we couldn’t relicense our Linux kernel contributions under GPLv3, because the overall license of the Linux kernel is “GPLv2 only”.
I don't think the "browser-only strawman" is condescending elitists dismissing users --- I'm a self-described GNU zealot on most days the only software I use on are apps like firefox, ssh, and vim and a few other bits like gnome/xmonad, liquid haskell, latex, and rustc.
I have no clue how the "average person" uses a computer, but for me it really is all web. Hell, if I had consistent LTE on my laptop, I could very just spend the whole day in a full-screen firefox with a web-ssh connection to my work server for development.
Average home user non-Internet use case # 0: Printing.
If for no other reason, I will never suggest Linux as a desktop option to any of my non-IT-professional friends & family for this one reason.
Use case # -1 would be "owning a laptop." Linux continues to be finicky on laptops - even when all the hardware works, you still end up with rando issues like poor battery life and fans that never turn off due to imperfect laptop power management.
(That said, there is one version of Linux that has cracked both of these nuts for the case of people who really do just want to use the Web: Chrome OS. And I do recommend that to anyone I think it'll work for.)
If you really were all-web, you would just use Chrome OS. As you said, though, you're actually a user of liquid haskell, latex, and rustc. In other words, either a computing hobbyist and/or developer.
There are actually pretty decent webIDEs for all of those things (shareLaTeX's vim mode is pretty nice, too), and I did consider switching to a chromebook. I was actually about to buy one before I decided that the max 8GB of RAM was way too little for my usual browsing habits. That, and I'm dependent on some firefox addons that just don't have good chrome equivalents, so chromeOS is sort of a no-go.
Edit: apparently there's a sizable segment of the market that agrees with both of us!
> In the third quarter of 2016, Chromebooks made up 54% of computer shipments to K-12 classrooms in the United States, says IDC analyst Linn Huang. That market share figure even factors in iPads, which themselves have been successful in education.
This course of study seems to take just as much, if not more, time than a math ugrad+master's at a typical uni. The people who could really benefit from having a path of textbooks like this fall into two categories: those who don't have access a university (due to poverty, rural location, and/or being a shift worker, for example), and high school students.
And I think the latter category is more important than people realize. When I was a high school student, I benefitted greatly from t'Hooft's theorist.html (like this, but for physics, and put together by one of the Greats of the field). It's part of what got me really interested in physics, it was a whole lot of fun, and it actually did a pretty good job of preparing me for graduate-level coursework. Eventually I left physics and math for CS (I'm in PL, so this is even less drastic of career change than one might think), but I still have warm memories of working along t'Hooft's guide and checking off topics as I finished the problems in each textbook.
>Once the type program's been compiled to a Haskell module, it becomes in essence a type-level library. You can include it in your regular Haskell programs and use the type class instances it exports in your own types to enforce whatever invariants you care about.
Cool! Is there an example of this? I poked around the repo and reddit/HN posts on it, but couldn't find anything...
Then, we ask the model to behave that way (with a prompt), and store the difference in activations for each pair. Then, a PCA can be used to extract the principal component, giving use the steering vector. We do most of this using the repeng library, and the author goes into a bit more detail on how it's done on her [blog](https://vgel.me/posts/representation-engineering/#How_do_we_...?)