Voice In, Visuals Out: The Agony and the Ecstasy

Talk transcript

1. 0:01

   All right. I'm Allen Pike, and today I'm gonna be sharing some of what we've learned building voice-in, visuals-out

2. 0:11

   experiences using AI. Here we go. So this

3. 0:18

   is Andrej Karpathy, and he made an argument last month

4. 0:24

   that voice is the human preferred input for AIs, but that we prefer visuals as the output.

5. 0:36

   And he knows a thing or two about this stuff, but that's not how we have been, for the most part, building or using AI. We've been typing to it.

6. 0:52

   It's been typing back maybe with some markdown. But there's been breakthroughs over the last few months where both this visuals-in and audio-- or the audio-in and visuals-out experience is now, um, really feasible, and we can create really delightful experiences with it.

7. 1:12

   The visuals-out side is pretty intuitive, right? Of course, a third of our brain is dedicated to processing visual information. We love looking at things.

8. 1:24

   And models have recently got to the point where they can generate rich HTML, tool calling, so we can have these experiences where there's visualizations that come back, explain things, help us understand, communicate the, the responses from these models.

9. 1:42

   They can give us interactive, um, controls that allow us to explore and understand and modify and change and, and, and direct the models. And they can even respond with beautiful illustrations and images, right?

10. 2:01

    And so the, the, the ceiling on the, the visuals-out piece has really lifted, and we have a lot more capability of what we can do in terms of responding from these models.

11. 2:15

    The more controversial half of Karpathy's argument, though, is this idea of voice as the preferred input.

12. 2:24

    We've long sort of idealized and fantasized about speaking to an AI and having this real-time conversation where it understands what we need and it, and it reacts, uh, appropriately in real time.

13. 2:39

    But the experiences that most people have had so far with voice interfaces have been more like trying to get Siri to turn the lights on, and it's not working, or, like, this guy [chuckles] that is, like, trying to get ChatGPT voice mode to do things, but it keeps being, like, awkward and confused, right?

14. 3:01

    The models that we have so far, the experiences that people-- most people have seen are both slow and dumb, which is, like, not a great combination. And so a lot of people are down on voice as an input.

15. 3:18

    But speaking is the ultimate way that as humans we communicate. When we're speaking, we have more words per minute than when we're typing,

16. 3:31

    but we also convey more with each word. There's a huge difference in between if you say to me, "Okay,"

17. 3:42

    versus if you say, "Okay!" And that's why when we have something that is really important that we need to communicate or get through to somebody, we will jump on a call,

18. 3:59

    and we will-- or we'll speak in person so that we can get through that high bandwidth, uh, communication.

19. 4:08

    Um, and so what we've done at Forestwalk is we have built an agent that is in our calls that can help us out in real time. And so the other day, I mentioned to my co-founder in a call-- on a call that I'd seen a bug with, uh, the Slack integration, or at least I thought I had.

20. 4:29

    And she responded that she had seen the same bug. And so I just said, "Okay, well, let's file that as a linear issue." And the voice agent within a second responded that it had done so.

21. 4:46

    And that feels perfectly natural when you get it really dialed in, that you ha-- are speaking either incidentally or, you know, with a purpose, intentionally speaking to the AI, and it responds in a way that is not interruptive v- of you.

22. 5:04

    It doesn't need to be voice. It is just taking action on your intent. And so this is-- We're gonna see this experience more and more in the coming months and years.

23. 5:16

    But there's a huge barrier, a huge challenge to making this actually work and feel good.

24. 5:22

    And that is the tyranny of latency. [chuckles] It's really difficult to get a response through that whole chain fast enough. We, we've known since the sixties that

25. 5:40

    to have a computer react to us fast enough that it feels instant, it needs to react in about a hundred milliseconds, a tenth of a second.

26. 5:51

    And so of course, we are always aspiring to get our products to react that fast, and sometimes we can achieve that. But with networking and everything, it can be challenging depending on what work needs to get done.

27. 6:04

    And so sometimes we might flex up to a thousand milliseconds.

28. 6:08

    Right? You get a full second. It's about the limit before people start to lose their train of thought. They ask Siri to do something, takes more than a second, we're off to something else mentally, right?

29. 6:19

    And so we're always in that play in between those two kind of human limits of trying to respond within a second, ideally within a hundred milliseconds.

30. 6:30

    But if we wanna have a seamless voice conversation with a, an AI or another human, the limit is much more aggressive. We need to have a 200-millisecond latency or less if we want to have a fully conversive, people are verbalizing something, and they are interrupting or interjecting or

31. 6:54

    agreeing and forming that sort of connection for a full voice-in, voice-out, uh, conversation. Um, and so imagine trying to do that where you have a network request, uh, maybe you take-- turn speech to text, uh, and then you do some sort of model inference on that text and then send it back on the network.

32. 7:15

    It's like a ridiculous, uh, amount of, of work to do in 200 milliseconds. Um, there are some clever approaches of trying to work around that. Uh, Thinking Machines and NeoLab has, uh, a really thoughtful architecture where they were demoing recently, um, just a few weeks ago, a-- they time sliced into 200-millisecond chunks like that 200-millisecond goal

33. 7:40

    and, and have a model that's doing continuous inference in and out on 200-millisecond slices. Um, so there, there are ways around it for voice-in, voice-out, but there's also just we don't need to wait for novel architectures.

34. 7:57

    We can just switch to having voice-in, visuals-out, and then we benefit from the more forgiving visual response envelope that people have. That if you have something that appears on screen within a second of what you've, you've said, it feels like it has met that, it's keeping-- it's within your attention span.

35. 8:20

    It's reacting to you in a way that feels seamless. Um, and so we've been building for that over the last few months, and it's a lot of fun, and we've learned a lot doing that.

36. 8:30

    And so I wanna finish up just by sharing three of the things that we found, um, to be really important for getting within that latency envelope and making the experience feel really delightful, um, for, for people who are using it.

37. 8:47

    So the first one is that in order to respond in a way that feels seamless, you have to use a really fast model.

38. 8:57

    So that's not just a model that is so small, um, that it can respond fast enough in a few hundred milliseconds, so of course, it, it does need to be.

39. 9:05

    Um, but it also has to be on an inference platform that prioritizes latency. So when GPT-5 Mini came out, we were excited that, um, that, okay, well, yeah, we get this more intelligent model, and presumably it'll be fast.

40. 9:20

    But then in practice, we were seeing latencies of 5,000 milliseconds, 7,000, P95, sometimes 10,000 milliseconds [chuckles] latency for the small model. It was cheaper than the bigger model, but we weren't consistently or even ever seeing it respond fast enough.

41. 9:38

    Um, Haiku is much better, um, in terms of, of that, that P95 latency. So you re- but you really need, uh, kind of a Haiku class model as the one that's responding in this real time and, and, or one of the smaller open source ones.

42. 9:54

    And then if there's a larger chunk of work that needs to get done, then that model then hands off or sends off an asynchronous message to a larger model that can think, and then the continuous real-time model or soft real-time model that's responding quickly then can we interleave in those responses if it's doing something heavier.

43. 10:14

    Um, so that's number one, you need to have a fast model, obviously with a, a short enough context that you're providing it that it can respond in a few hundred milliseconds.

44. 10:22

    Um, the second thing that's really critical if we're going to have this feeling of it being really responsive is to have short intervals of how quickly we're sending out for inference.

45. 10:35

    And so traditionally for like a voice application, you might listen to the user for a few seconds, and then they stop. You listen for a second of silence, and then you have, uh, some sort of inference goes on, and then you're, you've blown your budget by a, a pretty wide margin just for waiting for silence.

46. 10:55

    So if we're gonna have our-- something up on screen within a second, um, then when we wanna have our inference pretty eagerly responding as the person is talking, even if we're not entirely sure they've stopped talking, being willing to send inference every one or two seconds as they speak, 'cause, you know, we sometimes have a sentence where,

47. 11:14

    you know, we ask, "Oh, hey, we're gonna wanna change this," and actually also list you this other thing. It feels much more seamless if you're doing those things as people talk.

48. 11:24

    Um, so you have to have a, a model and, and an infrastructure where you're able to get these fast turns going.

49. 11:32

    And then finally, in order to actually make all of that work, you need to have a stable caching regimen. So we have this huge improvement in the way that we work with LLMs over the last year, where we have on the different platforms, different ideas of this prefix caching, where if the beginning of the context you send

50. 11:54

    to the model is the same each time, then you can get up to ninety percent cheaper, faster inference, um, depending on the conditions. And so you need to lean heavily into this architecture.

51. 12:06

    Really, I think from, for most applications, we're moving towards this, whether it's like a long-running agent or it's a frequently running agent. It's the same principle applies that we want the first ninety percent of the context window, if we can, to be the same from request to request and then just use that final ten percent, and then

52. 12:23

    also, of course, minimize the number of output tokens so that you can get these really fast and relatively affordable inference turns to create that, that de- delightful experience.

53. 12:36

    Um, so these are some of the, the techniques we've found really useful. Um, I would love to hear from anybody who, um, has been exploring and experimenting, whether it's with real time or any way that people are pushing the boundaries of creating delightful experiences, uh, with these models.

54. 12:55

    Uh, I'd love to chat, share what we've been learning. Uh, and I hope some of this inspires some of you to go out and, uh, build something great. Thanks.