Critical Look at AI Interaction Models  🔗 **Source:** [Critical Look at AI Interaction Models](https://share.snipd.com/episode/42050c41-e60d-446c-bc27-decd84b58343) ## Highlights - **Full Duplex Removes Voice Agent Latency** - Thinking Machines' Interaction Models aim to make AI listen while speaking to remove the unnatural latency in voice agents. - They implemented a full duplex system that processes input and generates output simultaneously, mimicking human conversational overlap. Transcript: Jaeden Schafer So they're trying to solve this by essentially having the AI model listen and actively taking information while it's also talking and be able to do this kind of simultaneously like A human would. We're going to get into all of this. It's fascinating also because this is coming out of Thinking Machine, which is from Miriam Mirati, who's raised a ton of money and hasn't had a ton of products come out. So anyways, a lot to get into here. But before we do, I wanted to mention, if you haven't already joined our AI Hustle School community, we would love to have you as a member. Every single week, we record a bonus video where we break down a tutorial on the tools, strategies, and things we're using to grow and scale our businesses with AI. We show you the actual products and projects we're actively working on. We show you revenue numbers. We show you deep dives. Everything we can't share publicly, we show you the tools we're using. And so it's an amazing place. You can go join. It's only $19 a month. We try to make this affordable for everybody. We'll keep this price low. And if you lock in this price, if we raise it in the future, it won't be raised on you. There's over 200 members in here. So lots of people that are actively building and sharing what they're working on. And we'd love to have you be one of them. So I'll leave a link in the description to that. Jamie, kicking over to this story with Thinking Machine Labs, I guess, what was the most interesting part of this software to you? Because I know you personally are working on and you have tools and you help businesses with kind of these AI receptionist kind of things. So for you, what are your first thoughts on this? Other Speaker I mean, I think this is a great idea. I think that's a real problem that they need to solve because, you know, even if you have really quick latency with, let's say, an AI receptionist, there's still a little bit of unnatural Lag that happens between the time you finish speaking and then it starts talking back to you. And sometimes even if you pause, but it's a natural pause, like I just stopped talking for a moment, it will start responding even though you haven't finished your thoughts. I you know i think there's up to this point this ai you know receptionist ai voice agents have gotten really good but they're still not quite to that that human level yet where you know As i'm talking right now you could be think you're you are thinking and processing what i'm saying Uh, a traditional, you know, agent will, it has to first turn all of your, your, what Your sentence into a text, then figure out the context and respond. And it's, it's kind of like a, a chain process where this is, this is trying to make it more conversational. ([Time 0:00:30](https://share.snipd.com/snip/9afe2d64-b4c2-494e-8fd6-2f37545663f0)) - **Design For Overlap And Interruptions** - Test human-like overlap handling in voice agents for high-interruption scenarios like heated calls or negotiations. - Prioritize milliseconds-level responsiveness to handle people talking over each other naturally. Transcript: Jaeden Schafer So I'll leave a link in the description to that. Jamie, kicking over to this story with Thinking Machine Labs, I guess, what was the most interesting part of this software to you? Because I know you personally are working on and you have tools and you help businesses with kind of these AI receptionist kind of things. So for you, what are your first thoughts on this? Other Speaker I mean, I think this is a great idea. I think that's a real problem that they need to solve because, you know, even if you have really quick latency with, let's say, an AI receptionist, there's still a little bit of unnatural Lag that happens between the time you finish speaking and then it starts talking back to you. And sometimes even if you pause, but it's a natural pause, like I just stopped talking for a moment, it will start responding even though you haven't finished your thoughts. I you know i think there's up to this point this ai you know receptionist ai voice agents have gotten really good but they're still not quite to that that human level yet where you know As i'm talking right now you could be think you're you are thinking and processing what i'm saying Uh, a traditional, you know, agent will, it has to first turn all of your, your, what Your sentence into a text, then figure out the context and respond. And it's, it's kind of like a, a chain process where this is, this is trying to make it more conversational. Anyways, I think it's a great idea. ([Time 0:01:36](https://share.snipd.com/snip/7d40738b-fba0-4201-927a-ddad2d3bddae)) - **Host's Real Frustration With Current Voice Modes** - Jaeden recounts frustrating experiences with current voice modes freezing mid-conversation and cutting users off. - He compares awkward pauses to hearing a response already answered in the first half and then the AI asking "What did you say?" Transcript: Jaeden Schafer But as far as the tech of what they've released, I actually love it. I think I was, you know, I was really curious what she was going to come out with. Because obviously, if she was just trying to make another competing LLM to chat GPT, which they very well might, who knows. But like, it didn't seem like it was, I don't know, you know, you go leave OpenAI and just make another OpenAI clone. This is cool because it's something completely unique, which is kind of what I expected her and even Ilya is working on something and they're making these kind of completely unique Platforms and solutions that didn't exist before. And so specifically what they have done with this, it's called a full duplex. That's the technical term. And like you mentioned, voice assistants right now, they run on this, what's called a sequential loop. So I'm talking, it listens, it transcribes it, it sends it back. But when you have an actual conversation with someone, you'll know it's quite a bit different, right? Like, especially imagine like a heated argument. I'm thinking of like a podcast, but maybe you have a heated argument with someone, you know, and like, you kind of talk over each other and they say something and you say no, and then whatever, Right? Or like a negotiation that's like really active or people like talking really fast and they're all talking over each other that is incredibly hard slash impossible for an ai model i'm Just thinking about myself having conversations with chat gpt on their voice mode which of course the voice mode it sounds natural it sounds good but the problem with chat gpt voice Mode is it's like giving me a long response and it kind of already answered the question in the first half of its response and so i'm like okay like, I don't really need that. Like, move on. Tell me more about this. As soon as I start talking, it's like, it like freezes. It's like, sorry, I missed that. What did you say? Like, you know, it just like cuts itself off. ([Time 0:05:23](https://share.snipd.com/snip/fcf8e40f-d106-4336-bb94-86cd93695b61)) - **Midresponse Pivoting At Millisecond Latency** - Interaction Models bring dynamic mid-response pivots to voice like Claude CoPilot's mid-write edits but at millisecond latency. - This allows AI to adapt instantly when a user interjects, avoiding freeze-and-reask cycles. Transcript: Jaeden Schafer And so specifically what they have done with this, it's called a full duplex. That's the technical term. And like you mentioned, voice assistants right now, they run on this, what's called a sequential loop. So I'm talking, it listens, it transcribes it, it sends it back. But when you have an actual conversation with someone, you'll know it's quite a bit different, right? Like, especially imagine like a heated argument. I'm thinking of like a podcast, but maybe you have a heated argument with someone, you know, and like, you kind of talk over each other and they say something and you say no, and then whatever, Right? Or like a negotiation that's like really active or people like talking really fast and they're all talking over each other that is incredibly hard slash impossible for an ai model i'm Just thinking about myself having conversations with chat gpt on their voice mode which of course the voice mode it sounds natural it sounds good but the problem with chat gpt voice Mode is it's like giving me a long response and it kind of already answered the question in the first half of its response and so i'm like okay like, I don't really need that. Like, move on. Tell me more about this. As soon as I start talking, it's like, it like freezes. It's like, sorry, I missed that. What did you say? Like, you know, it just like cuts itself off. And it's like has this awkward moment where if I was just talking and I had kind of maybe able to finish its sentence and listen to me at the same time and then immediately start talking To me throughout it. And it kind of reminds me of what we're actually getting out of something like Claude Cowork, or maybe even OpenAI is kind of doing this too, where mid-response, you can actually ping It with new information as it's kind of writing out its response, and it will change dynamically the response, which for quick little responses isn't that important because it does Them pretty fast. But when it's doing something like Claude Cowork, it's doing like this massive multi-step project or building a website or a web page for you. And, you know, it takes five minutes and halfway through you send it a message and then it can kind of pivot and change. I don't know. It feels like the same concept, but they've done this in a technical way where it's not seconds or minutes. ([Time 0:05:52](https://share.snipd.com/snip/5ae6404d-f5cb-48a6-83fe-4825e3542bc7)) - **Latency Is The Adoption Bottleneck** - Solving conversational latency is the missing piece for mass adoption of AI receptionists. - Until dialogue feels seamless (no awkward pauses), users will recall bad past experiences and resist AI receptionist use. Transcript: Other Speaker Yeah. I mean, what do you think the implications are for this? Because right now, I think people are still turned off by the idea of, for example, AI receptionists. So it takes a lot of ability. It takes a lot of... You have to be very creative in how you present that to somebody as far as something that they would want for their business, especially because people are thinking back two years to The last time they called tech support somewhere and they had to talk to like a very robotic answering service. And they're like, I absolutely don't want that. It's changed a lot already, but this is kind of going to take it to the next level. So, you know, how do we, what are your thoughts as far as like long will this take to be adopted by the masses or at least by companies so that they feel comfortable with this? Jaeden Schafer So yeah, I mean, I think there is a limited research preview that's coming in the next few months. They're going to have a wider release later this year. I think until this piece of the puzzle gets solved, and I'd be so curious if OpenAI just clones this, which very possibly they can, or if people integrate with them to help them do it. ([Time 0:07:45](https://share.snipd.com/snip/52e5c0f0-99fe-42e0-98f6-898ca31d3239)) - **Pitch AI As Capability Expansion Not Replacement** - Pitch AI receptionists as enabling capabilities rather than replacements to reduce resistance. - Emphasize tasks now possible (faster refunds, quicker support) and improved authorizations to solve problems end-to-end. Transcript: Jaeden Schafer So yeah, I mean, I think there is a limited research preview that's coming in the next few months. They're going to have a wider release later this year. I think until this piece of the puzzle gets solved, and I'd be so curious if OpenAI just clones this, which very possibly they can, or if people integrate with them to help them do it. I'm not sure how technical it actually gets. But I think until this piece of the puzzle is solved, people are still going to feel like it's quite robotic. Because if you're talking and it's talking and then it pauses and you have this glitch, it's just this awkward moment. It just feels like a robot and we don't like it. I think at the end of the day, the AI receptionist idea will be fully realized and people won't mind or care if it sounds perfect, acts perfect. It accomplishes everything an AI receptionist could accomplish. Maybe it has all of the same authorizations to issue refunds or whatever. Like for me, when I talk to chat bots, um, on a company's website, if it can issue me my refund or help me with my problem or, you know, whatever the issue is, if it has authorization to actually Solve my problem for me, I don't care if it's a human. In fact, I'd rather it not be a human. I'd rather just talk to a bot that can solve my problem. But the problem is a lot of times we talk to these bots and they ask all these questions. They're like, let me root you to a human. And then it's just annoying. You just feel like you kind of wasted your time like there was a screener kind of like screening you so i think that's probably part of the issue um and then there's like the whole latency Problem which is what they are basically solving here and i think once this rolls out i'm with the right authorization that these these bots will be awesome the other thing that i think Is interesting because you mentioned this a couple times of how people don't really like the concept of ai. And even when you talk about a receptionist, they have these bad experiences they think back to. One thing that I think would give people a better, more positive connotation to all of this is less of replacing a person with the AI. So like replacing your receptionist with the AI and more of, hey, you can do more than you did before. Like, you know, before your customer support people, they would get around, you know, within six to 12 hours to customer support requests, or maybe you're getting super bogged down, You get days behind, like, you know, that's a very bad experience. Or maybe you're really bad at calling people back, and it just doesn't happen. So I think pitching things as like, you were unable to do this before, and now you're able, like, you can, you know, now I just can replace that person or do it a little faster. Like a little better isn't as good as like you are dropping the ball in this whole area and now this is going to fix that problem. So I don't know. That's a kind of a perspective shift that I think might be a better pitch for people. ([Time 0:08:33](https://share.snipd.com/snip/227728c8-4649-4c59-a166-071a0ad7cc78)) - **Watch For Partnerships Or Big Tech Clones** - Expect partnerships with specialized audio companies if the solution is technically unique; otherwise expect big players to clone it. - Monitor whether 11 Labs or OpenAI/GCP roll native versions or integrate via partnership. Transcript: Jaeden Schafer With. It's incredibly hard and technical and like they solved something that's very difficult for other people to solve then i would see companies like 11 labs 100 partnering with them on This um if it's just like they figured out something cool but they figured out at first and everyone else can kind of figure it out which is what it feels like happens so often in ai right Like as soon as opening eyes like we have a reasoning model and it's just running through a whole bunch of prompts. Everyone could kind of see how the reasoning worked. They reverse engineered it. All the Chinese labs cloned it. And then now we all have the reasoning models because they weren't like a crazy technical thing to figure out. If that's the case, then I think you'll probably see companies like OpenAI and Google roll their own version of it. You might see OpenAI and Google roll their own version regardless because they don't really like integrating or working with other people. They can just try to figure out how to figure it out. But yeah, for a lot of smaller companies, I mean, OpenAI, 11 Labs is getting to be one of the bigger companies now. But for ones that aren't OpenAI and Google, I think you'll see a lot of partnerships with them to be able to get stuff down. ([Time 0:11:44](https://share.snipd.com/snip/3546fe64-c4d4-4f33-9876-38678909db39))