quick rant
My last two posts have been about dictation and text-to-speech, respectively. I feel like now is a good time to also speak on the more narrow tool that is an AI voice mode. What makes a good AI voice mode and how and when can it be useful?
The tension with an AI voice mode is latency and response quality. As of recent, there has been a push by the major AI labs, including Google and OpenAI, to move towards lower latency approaches. Since I have a ChatGPT subscription, I will now talk about their offering which in the latest edition would be GPT-Live-1 in their web UI. With this, the thought is of keeping a capable model in the background to do decent research at an effort level that I'm not sure is configurable, while also putting a secretary in front that is remarkably stupid, like one, two years ago kind of stupid, but extremely quick to respond and capable of good turn-taking and handling interruptions.
Now, I want to acknowledge from a technical perspective, it's quite impressive to do quick, accurate turn-taking, as well as responding in 100 milliseconds. As well as even interjecting with an active listening simulation. This in AI is kind of irritating to me because when I want to prompt it, I don't particularly care for affirmations that it hears me, certainly not in a natural language kind, but it is impressive from a technical standpoint.
Now the tragedy is that this takes a lot of the stuff that is great about chatting with an AI model and just tosses it out, removes it. If you talk to or rather write to something like GPT-6 Astra, you get multiple paragraphs of detailed, researched information. You get a sort of mini-lecture for each of your prompts, certainly by default. You can tweak this, if you prefer to be more concise or to be quicker and use less reasoning.
So, what's the issue? The issue is, there was something called "Standard Voice". This inside of ChatGPT was, to my knowledge, the earliest voice mode: a very thin wrapper, where you have text to speech, as well as speech to text and some very thin layer of turn taking. So you have your main model, the model you actually want to talk to, and you can just kind of talk to it. And admittedly, sometimes it will research something and then you got to wait 30 seconds, right? Or sometimes you get a long response and then you're, you know, kind of busy listening to it for five minutes and you can interrupt it, but it's not as graceful as with the new iteration.
So, I'm not claiming that the standard voice was perfect from a technical perspective, but it was perfect from a product perspective. And so, my point in publishing this rant is is mainly to encourage anyone and everyone who either works in AI labs or is also a power user to bring this kind of experience back and to inform me where I can get it. Because that's really where voice mode shines is when I know I have something that's kind of detailed, that needs a lot of research that I don't fully grasp yet, such as maybe complicated tax law or accounting or maybe I want to research a specific historical period. And then I can enter a phone call with an expert on a topic that can give me a kind of mini lecture and where I can kind of explore the topic but keeping it in a detailed well-reasoned high-quality format.
So, in a short conclusion that is perhaps not very satisfying, I advise you in the current date to probably stay away from voice mode. Maybe wait another year, see where the labs are taking it. But due to decisions of product managers that perhaps are not too keen to follow the advice of end users and would rather cater to business needs, for now, the only good voice mode is out of stock.
When it returns, remind me and give it a shot. Until then, who knows, maybe it's better to lay off the AI every once in a while.