I’ve never adopted a voice assistant. I’ve had one available for a while on my smartphone but after playing with it for a few minutes and then accidentally triggering it a few times, I disabled the voice assistant’s hardware button completely. There’s just not enough value there for me.
My issue with voice assistants is that I don’t know what commands are supported or how to properly use them and there’s no reasonable onboarding path. With menu based software, there’s an exploration path that also serves as training. I can learn the locations of all of the commands I regularly use by simply moving through the menus and I can not only optimize how quickly I reach the desired menu item over time, but I can also learn keyboard shortcuts as well, which good menu systems display.
No such onboarding path and increased adoption and usage flow exists for voice assistant commands and so I’ve never picked them up and probably won’t until they’re far better than they are today.
Some will say that LLMs will fix this. I’m skeptical, but could be convinced. If LLMs can translate more natural language to the appropriate set of commands, maybe things could be improved. Then again, maybe voice assistants face more issues than a failure to alias broadly varied human speech patterns to a much narrower set of supported commands.
The one place I think I would like to use voice commands is in the car. When I’m listening to music, for example, it’d be nice to be able to press a button on my steering wheel and say “How well did this album sell” or “Did Chrissy Amphlett ever perform with Jimmy Barnes?” or something like that, and get a true answer.
An LLM like ChatGPT could handle much of this today, modulo the “true answer” part. That might be OK for my mostly inconsequential music questions, but what if I’m in the car listening to the news and have a question where the accuracy of the answer has more impact.
I do remember completing the user onboarding for the first Alexa tower speaker. It had a companion app that introduced the lingo to trigger responses. Subsequent email marketing introduced the sockets idea of importing “skills” from other partners. Keeping up with the expansion of vocabulary was a bit cumbersome. Over time I settled into just using a few prompts that were reliable. I believe you’d like to see the chat assistants train users natively through dialog rather than resorting to a secondary app for onboarding. Is that the case? Would you ever be into conversing casually with an LLM just for jollies? Or do you see them as task-specific tools only?
Yeah, my thinking is use an LLM for “translation” which they’re generally pretty good at (I got LLM translations going for Firefox last year.) The LLM can know the available commands and also enough real language to accurately translate between the two. Then I don’t have to learn a new, mostly invisible, and potentially changing, syntax and diction. The assistants, IMO, could have aliased a whole lot more and made the tools better but now that we have LLM, why not do the translation/aliasing that way. I do play with LLM chatbots like Bard, Claude, and ChatGPT, but I’m not getting a lot of value out of them yet as it often takes querying several to triangulate on an accurate reply. I also haven’t seen the kinds of improvements over the last year that excite me, outside of the variety and that a couple are more up to date than a year ago. I’ll keep poking at them, but I don’t think LLMs are going to go a lot further than where they are today. Maybe we can shrink them lower their costs, but for any real AGI, we’re gonna need a better approach. LLMs might be a part of AGI, but, IMO, not the critical part.
Very interesting. Considering what you wrote previously (elsewhere) about your thoughts on AGI and LLMs as the wrong approach to it (for jpeg reasons) and your thoughts about interfaces and lack of onboarding as a hurdle to the utility of chat interfaces based on LLMs as a foundation, I interpret that you have an engagement model in mind presently and you are really hoping that LLM would be a more useful way to get there. They might be more useful to you if they weren’t just a regurgitation of the last 10 years of web or some other historical repository. Translation intermediary between languages, aiding interpretation across people, is a great function for LLMs certainly. But beyond that rather dull function, I’m curious what utility or tool you’re dreaming that would require an AGI machine. If you had one, what good would it be for you beyond replacing web queries?