Colin @csgordon@types.pl

I love my job, but I do find it frustrating how much of my job consists of repeatedly arguing obvious stuff like "phrenology is bad even if a computer does it" and "using probabilistic lossily-compressed text databases to guess next words is not sentience" and "understanding Merkle trees does not guarantee you'll make tons of money in speculative trading of unregulated currencies" to students and peers and occasionally university administrators.

Realizing I never did an #introduction here, even though I joined a few months ago.

I'm Colin, a #pl professor at #Drexel University. Historically I've mostly worked in (capability-based) type systems, effect systems, and program verification. Lately I'm also involved in the "linguistics of formal specifications" (particularly writing formal specs in English) though I'm moving my grumbling about linguistics papers that misuse definitions from computability and logic over to my linguistics-focused alt @csgordon@lingo.lol

I will however continue to talk here about #coffee, running #freebsd on laptops (bhyve & encrypted #ZFS home directories ftw), random OS (operating system) kernel stuff, and resisting the ever-present temptation to rewrite everything in #rust (mastodon, FreeBSD wifi drivers, research code that already works, code for classes I teach in Java, i3, etc).

So apparently Stanford HAI has put up a PR blurb about a preprint where they concluded that "AI" (i.e., ChatGPT) "may have already caught up to the persuasive capacity of everyday people, a critical benchmark of human-like performance."

No, AI has done no such thing. If you read their methodology section, they compared ChatGPT's output -- a statistical interpolation of every extended argument on these issues that hundreds (or more) of dedicated humans (many probably experts spending hours of time each) have written over the many years of training data ChatGPT was trained on -- to the argument put together in a short time by some rando they recruited from a mechanical turk knockoff.

They've shown that the statistical average of every written argument for these issues is on par with an internet rando.
https://saturation.social/@Noupside/110052790316882007
https://doi.org/10.31219/osf.io/stakv

Sadly I can't find this forward in English, as I cant find any English binding of the initial trilogy at all, only the German version I'm reading https://www.kobo.com/us/en/ebook/die-neuromancer-trilogie-1
All the English collections I can find include the fourth book

If you want a cool example of language change over time, or just want to feel old (it's a 2-for-1 deal): I'm reading a German translation of the first 3 books of the Sprawl series (apparently published before the 4th book came out), and in the foreword Neil Gaiman pointed out how the meaning of the first sentence of Neuromancer has changed drastically. It refers to the color of a TV screen tuned to a dead channel, and to most people around my age or older this is obviously fluctuating gray and white static, to a generation younger it's blue, and if my son reads this when he's older it'll be black.

Take that back: almost perfect. There are small groups of LLM researchers who are realistic about limitations and are also interested in using them for purposes they're fit for (or specifically trying to make them fit for specific well-scoped purposes) where one doesn't need to assume they "understand" meaning in order to drive benefits.) Those people are usually pretty willing to talk about ethics and have coherent well-reasoned thoughts about the matter.

The other 95% though, are well-captured by the linked post.

Truly the perfect analogy
https://fediscience.org/@MarcSchulder/110023486566798654

Literally every zoom meeting I'm in with > 6 people, at least one of us has to reboot to get sound working. This has been the standard for 3 years now. I'm absolutely astonished that all alternatives are still strictly worse after all this time.