January 4, 2023

s14e03: Success Criteria; Chat

0.0 Context Setting

It’s another cold, cloudy morning in Portland, Oregon on Wednesday, 4 January, 2022.

---

1.0 Some Things That Caught My Attention

A few small things again today as I get back into the practice of hitting 15 minutes.

1.1 Success Criteria

Another thing that stuck in my head from Tim Bray’s essay about search on Mastodon¹ was his section on success criteria. Here’s the first para, but I encourage you to read the rest (there’s only three more paras of roughly the same length):

I’d like it if nobody were ever deterred from conversing with people they know for fear that people they don’t know will use their words to attack them. I’d like it to be legally difficult to put everyone’s everyday conversations to work in service to the advertising industry. I’d like to reduce the discomfort people in marginalized groups feel venturing forth into public conversation.

Caught my attention because: It clearly outlines relatable examples without going into too much detail. Which isn’t to say detail isn’t important, but to me it’s a good example of sketching out a direction as well as different contexts:

• From a user/poster’s point of view, of understanding the particular context they’re conversing in as an individual with specific needs, and understanding usage as more like conversation than written communication, I think
• The context of business usage (“legally difficult”), and reinforcing a point in the piece about for example carving out exclusions for usage in ML training.
• The fuzzy goal of reducing discomfort, which presupposes going out and actively understanding what that discomfort may be, while at the same time not seeking to completely eliminate that discomfort

As an aside, I especially appreciate how Bray’s essay doesn’t talk about how the usage of hashtags might solve search. Hashtags are useful and interesting, but from my perspective they impose too much work on the part of the writer – especially if you want your conversation to be indexed and searchable – as a sort of per-post, opt-in high cognitive cost. Much easier, I think, to either set indexing and licensing globally for the entire account, or if you really want, to tie it to per-post posting contexts. But maybe not! It would be interesting to see!

1.2 Chat

I went along to the Near Future Laboratory’s General Seminar 26 on ChatGPT futures last week and it was… interesting?

There’s something about ChatGPT that feels limiting and I’m pretty sure it’s in the chat/REPL² interface to the language model.

It’s somewhat astonishing to me that nobody has yet hacked up a Young Lady’s Shitty Illustrated Primer. Given infinite AI arrays³ appeared this week, it doesn’t feel like it would be too far off.

Bear in mind this list is in the realm of “just” as in “just do these things” where “just” is “a giant yawning chasm of implementation” and the actual job is neatly obscured using one word.

• Pick a few topics. Maybe just one.
• Ask ChatGPT to write an introduction to that topic.
• Ask it to write some problems in the form of a children’s book.
• Use all the above as static text.
• Bound figuring out the problems in a ChatGPT REPL, i.e. as a text adventure.
• For bonus points:
  • generate dynamic illustrations based on, I don’t know, identifying salient noun phrases (see, I can word salad just as well as GPT) or whatever by piping them through whatever image generation model is your favorite
  • text-to-speech the entire lot
  • speech-to-text the other side of your REPL
  • (Actually, has someone tts-stt a ChatGPT implementation and thrown it online for people to play with yet?)

See? Easy.

The one thing that I’m not sure has stuck around here has been continuous training. ChatGPT is good (I think?) at preserving state through each session (e.g. you can tell it to “do that again, but x” and it’ll mostly work), but in terms of learning preferences and updating a model continuously, I don’t think I’ve seen that? I mean, I think there’s the whole federated on-device learning thing going on, but I don’t yet see that in terms of integrating ongoing training with conversational large language models? Or maybe I’m not looking in the right places.

As an aside, Siri continues to be dumb. Here is an example of what I would like Siri to be able to do, because from the point of view of a stupid user (i.e. me) all the data is there:

Me: Hey Siri, play some music for me

Siri: Sure Dan, here’s some music. [plays music]

Me: No no no I get that I have this weird thing of listening to college a capella covers of songs, but I really don’t want that right now. Play music for me but don’t play any college a capella

See, this would in theory rely on sufficient metadata like all the college a capella being tagged with the a capella genre, but this whole dealing with sets is something that a) Siri clearly can’t do, and most frustratingly, b) is something that Apple Music can’t do, and c) is something that iTunes could kind of but not really do with Smart Playlists.

It’s frustrating because you look at it and you think ugh this is right there and you can’t really do it? (Note: I know, there are a lot of justs here)

1.3 Updates

Thank you to the reader who let me know that you can sign in to your Square profile and turn off marketing emails for everyone you’ve ever bought anything from. (Scroll down to Notifications, hit Email, you should be able to see it there)

---

This was longer than 15 minutes. In fact the Tim Bray thing itself was about 12 minutes of wall time. Not great for my goal.

How are you doing?

Best,

Dan