Yeah, beware “no true Scotsman” here. But I think it’s a reasonable hypothesis that GPT-4 is “good enough” or close to it, and when the current assistant services get wired up to LLMs we’ll see a step function in their utility. Apple and Google are both definitely working on this. OpenAI’s voice mode is closer capability-wise, but doesn’t have the integrations; that would be an obvious product for them to do this year.

If this doesn’t seem obvious in 1yr I’d say the hypothesis is likely wrong.

Sounds to me like a silver bullet idea. "LLMs will make it good".

I don't think voice is a good interface. "It chats like a human" is the lowest possible hanging fruit in terms of product design, and bets everything on the smarts of the tech that's behind it.

We are so used to tooling faster than voice. Keyboards and taps are very, very fast. I want digital assistants as smart AND as fast as that, not something smart but incredibly slow to interact with because it needs to dumb itself down to human speech I/O.

To me, this is also not about modality or making it more generic. I just don't want an anthropomorphized smart-ass assistant. I want smart tools that actually assist me directly, no chat.

Agreed. I find it utterly exhausting to talk to most humans, at least when my goal is purely transactional or functional. Human conversation takes work. Let's save it for what matters.

People hate call centers. I don't want to have a conversation with some human or human-parity AI at my airline, I want to change my flight in 3 clicks. In retail and fast food, self-checkout and online or kiosk ordering is popular for a reason.

I've found myself feeling those same feelings when being forced to have conversations with a chatbot. Chatting with ChatGPT just for fun can be fun, but it can be just as painful as a call center if you need to get something specific done. In order to cancel a hotel reservation, I was having to chat with some bot. It made it into a whole conversation, with brief pauses. It should have been 3 clicks.

I don’t think I’m naively expecting LLMs to magically fix things. The reason voice is currently a bad interface is that the agents are dumb, mapping verbs to hand-coded action trees. The missing pieces are first intelligence and second long-term adaptability/personalization.

LLMs can plausibly solve both.

I can see a demand for increasing or expanding LLM context without summarizing. We're not there yet though. Limits are tough.

Also, I don't want to engineer a prompt to convince my assistant to do my bidding. That's just hand coding action trees with extra steps.