I believe this is a clear example of the Moloch effect [1].
Companies race towards having the shiniest chatbot or _virtual assistant_, without fully understanding all the implications and customer needs. Everyone wants to have their brand name associated with being "the first" or "the best" in their segment for providing such technology, no matter how trivial it may seem -- e.g. as the article points out: just wrapping already available info in a dialogue.
Given these incentives and high stakes, it's also hard to "exit" the bandwagon, due to various factors like reputation, costs, and the uncertainty around your competitor not doing the same.
The other day I was frustrated with several Linux quirks my laptop was experiencing and decided to give Windows 11 + WSL2 a try.
The sheer amount of bloat, sneaky privacy settings, ads, clunky UI etc. literally make Windows unusable. I was willing to put up with the switch (leveraging WSL2), but the entire operating system feels like a browser with toolbars from the 2000s.
While I agree with your sentiment, I do not believe the author tried to "promote rote-learning" here -- they just presented a nice little analogy between Japanese and programming constructs.
> It should be generally better across a wide range of topics and has improved factuality.
Maybe I am being naive, but this is not telling too much as there is no metric attached to it, so the average user cannot (at least immediately) gauge how impactful this improvement is.
I think it would be beneficial to have benchmarks for assessing the factuality of large language models. Now, several open questions are:
- What is the minimal (necessary) set of questions it must contain?
According to [1], the g4dn.xlarge (a single nVidia GPU instance) appears to have the lowest on-demand price/hr: $0.526. The Tensorbook starts at $3,499 [2] -- that's approx. 6,652 continuous compute hours (~277 days or ~39 weeks or ~76% of a year). Let's round this up to a full year of productive research. I think local hardware may actually be cheaper in the long run for _moderate_ problems, but please correct me if I'm wrong.
There's GeoWizard [1] who is absolutely crushing these challenges -- NMPZ (no moving, panning or zooming) is especially interesting to watch, e.g. [2].
This reminds of IGCC (interactive GCC) [1] which I forked from [2] some time ago and since then I try to maintain it.
I don't use it very often, but it's handy to try out syntax or certain snippets. At its core, it continuously updates a temporary file that gets sent to gcc for compilation. Flags & libs can be specified in a config.