Hermes Frangoudis (00:07)
Welcome everyone to the ConvoAI world podcast where we interview the leaders pushing the voice AI space forward. Today, I'm very proud and excited to have Diana Zhu from RiseLink with me today. Welcome Diana.

So Diana, before we really dive into the tech piece of it, one of the things that we really like to talk about is like the origin things. So please take me back. Tell us the origin story. What was the problem that the company was really trying to solve from the beginning?

Diana Zhu (00:36)
Yeah, I started 20 years ago with the founder, Dr. Pengfei Zhang. He was a pioneer of a Wi-Fi dual bandwidth, fully integrated Wi-Fi chip. He invented this chip.

And that was a historic moment in wireless connectivity chips. And then he later founded a company in Silicon Valley and successfully sold it. In 2005, he decided to start Beacon Corporation, which then had been around for 20+ years, holding over 1,200+ patents. And we think it's time that the US builders also have access to the super-low power, highly intelligent edge AI chips that we

produce

and hence, this year we started our US office so that we can make ourselves more accessible for builders and intelligent on-edge devices.

Hermes Frangoudis (01:26)
Super exciting and you guys have an awesome SDK and developer platform. Like really excited for people to experience that. Yeah, excited for it. In terms of your own path, like you've had a really interesting path, right? From Harvard to Yale, what led you to the world of AI and chips and wireless technologies?

Diana Zhu (01:45)
Yeah, it was kind of magical. A little bit about me, I started off in Applied Math at Harvard. And what I was most interested in is applying research to real-world applications. So I first spent a year in finance in New York developing algorithms for foreign-exchange. And then I decided it was more interesting, the healthcare applications of economics that's more policy oriented to actually improve people's lives. So I started my PhD

at Yale studying questions like how do we leverage AI and Machine Learning to match patients to the right doctors and also what kind of drugs work best for people who have all different characteristics so the same drug may affect different people differently so developed an algorithm to decide that. I was very fortunate and received funding to bring one of my research projects to life as

startup after a PhD that focuses on patient provider matching. So bringing technology and latest research to real application has always been a passion of mine throughout. And the story to wireless actually, it was started planting seed very early on when Dr. Zhang showed me ⁓ the BK7258 devboard. And back then I was still working on my health tech company.

And, but instantly I was like, wow, what if like my toys, like my little horse from Jelly Cat can finally talk? Cause that's a dream growing up. As a child, I'm like, I always give all these personalities and have these virtual, like not-real conversations with my toys. But it seems like with all the AI technology now, this is totally something that's possible. So like that started the itch in me where I was like,

Oh my gosh, so nice. I probably need to make the transition to really dive into this world. So when the opportunity came up to lead the US office here, it was a pretty quick decision to start jumping into this new world. And yeah, it's been so rewarding just learning with the builders in US who are building really cool applications. And we bring them the ecosystem, the place for the, I guess, the platform for them to

really achieve very cool applications that are intelligent and connected to the cloud.

Hermes Frangoudis (04:05)
No, that's super cool. And you bring up like, brought up earlier that the RiseLink has been around, you know, global company. And this is, as you said, you're here now leading the US office and opening up kind of these US operations. So, so many like US developers are just kind of like starting to discover RiseLink. And I think you hit upon it really interestingly, like

the area that the company is really like trying to push forward is this connectivity, this like smart device, right? Like that's the message you want developers to hear.

This is not only a really cool piece of technology, it's something that you can customize and bring to life yourself.

Diana Zhu (04:43)
Yeah, exactly. I want to like the messages that we want to want US developers to hear is that we understand the time-to-market is extremely important. And just being in hardware the last year made me realize it's actually like ⁓ I came from the software background. Like on top software, like making hardware work is just so difficult because you also have the manufacturing piece to make sure all the connectivity like all the powers optimized. There's so many other concerns on top of

software that you need to take into consideration and we want to provide the most support to developers here, not only in terms of the hardware developing, making it easy to develop on our chips, but also the software support to make it, we have like the full engineering support if you have a cool idea, we want to work with you in order to make that a reality.

Hermes Frangoudis (05:35)
So cool supporting the developers like huge for the community and you bring up a very interesting topic around the lift it takes to do hardware, right? And RiseLink chips are known for like the really ultra-low-power consumption and

how did your team kind of hit that? That's something that's worth pointing out. The power consumption is very important for hardware.

Diana Zhu (05:58)
Yeah, exactly. So that's actually something that we're really proud of and devoted a lot of R&D resources into. So this is kind of like a world-record level of like achieving 50 milliamps and Wi-Fi keep-alive. And then really it comes down to really aggressively power  saving at every level. For example, we developed the hardware, the custom wireless radio.

so that it can hear, it can optimize how the chip hears signals and still wake up while not consuming a ton of power. And that also really gate where power is sent to on the chip. So when some peripherals are needed, when some like the AI processor isn't needed, we just completely gate away the power so that it's not even used in those parts.

And also optimizing on the firmware level with algorithms. How do we like, what's the scheduling for the best time to listen for signal from the router? And I'm definitely not like a technical person super deep in the semiconductor industry yet. But yeah, I know this is something our team is really proud of and they have to do all this energy thinking about this problem.

Hermes Frangoudis (07:09)
I mean, it's huge because that's just, it adds to the lifetime of the device that's in people's hands, right? Like the longer someone can have it in their hands, the closer of a connection they can have with it, right? Exactly.

Diana Zhu (07:14)
Right. Yeah.

Exactly.

Exactly.

Think about all the wearables like for a healthcare device, for example, like one charge you want it to last for days. And then also the smart door locks, you don't want to always want to change battery, etc. So it's really important. And even for the toys, like we don't want the toy to have all have to always be charging, but you want to maximize the time you're interacting with these devices. And what if a camera, IP camera also. So yeah, for all the

devices like power is always ⁓ one of the top concerns.

Hermes Frangoudis (07:51)
Completely make sense and aside from power, one of the other big concerns is performance, but balanced with efficiency, right? So these chips from my understanding are really designed to kind of

do that really difficult balancing act, right? And so could you tell us a little bit more about how you get the AI to run like on the edge like that?

Diana Zhu (08:05)
Yeah, yeah, exactly.

Yeah, so you made a really good point. When we are using a lot of compute power, it automatically means that we need to use more

power to support that processing. I guess one way we're trying to solve it, for example, with now our latest edge-AI chip BK7259 is that we architect the chip with a dual-core kind of layout, which means that when you're processing every day more like easier kind of tasks, we use a core that's it's called M52 from ARM,

that uses less processing, less power. But once you switch to something that requires more power, more processing, we switch to a different core to process it. It's called M55 Cortex. And we also have NPU on the chip as well, which means that it can accelerate the Machine Learning and AI processing ⁓ speed for things.

And so with this dual-core mode, we can see when you're doing mundane tasks, then we just allocate power to one type of the core so that it doesn't consume a lot of power. And then when it's processing needs more processing power, then we switch to another one of the cores to optimize.

Hermes Frangoudis (09:29)
Sorry, I'm just like blown away by the thought that probably has gone into this, right? Like it's a multi-core architecture that has thresholding, right? So it understands, okay, here we're gonna use this core, but then we're gonna dynamically switch over here.

Diana Zhu (09:34)
Right.

Yeah, and on top of that,

there's like an NPU that accelerates AI and Machine Learning algorithms when needed, and they wake up intelligently when these like different demands arise. It goes beyond me as well.

Hermes Frangoudis (09:56)
That's wild. I was like, I was just gonna ask. Did I hear you right? Did you say there's an NPU on the board? And that's on the BK7259, right? Yes, that's right.

Diana Zhu (10:02)
Yes, NPU.

Yes, that's right.

And if we're talking more about BK7259, that incorporates more edge intelligence, which means you can do a lot more things on the edge without sending data to the cloud. So what that means for

builders is that now you can run small language models on the edge so that your toy can talk to you without having to wait for the cloud large language models to respond. Also, it can recognize your face, recognize your voice, and understand voice commands locally without sending anything to the cloud.

Hermes Frangoudis (10:43)
So it could do like the voice fingerprint, right? Because that's an NPU process and then store that data locally. So it does the animal. So, so it offloads those little like those pain points. Like that would be so painful if you had to wait for that. Right. But because it happens on device, it feels probably like frictionless.

Diana Zhu (10:48)
Right.

Exactly.

Yeah, exactly.

Exactly.

Hermes Frangoudis (11:04)
That's so cool. Really now, because voice is becoming such a dominant interface with these devices, how is RiseLink positioning itself at the intersection of hardware and conversational AI? I know part of that is our partnership with Agora, but I guess, what else are you guys doing?

Diana Zhu (11:18)
Yes.

Yeah, so we are sitting in a very interesting nexus of things because we build a hardware foundation to enable.

devices to be voice-ready. And we want to be the go-to voice device for developers when they're making new things. examples are we design our SoCs with features like integrated audio ADCs so that it can connect to microphones very easily. It can end enough on-chip horsepower to enable the chip to

wake up and also connect to the large language model in the cloud or do edge processing. On top of that, we also have on-device wake-word detection and noise suppression, which means for the chip when someone is saying, "Hey, Agora", "Hey, buddy" like it can be named anything, the chip will like hear for the signal and other times just

stay very put and also have the super-low Wi-Fi keep-alive current. And as it hears words from the person interacting with it, it suppresses background noise so that we can hear better when interacting with the user. And all these are built into the chip so that the user doesn't have to worry about it.

Yeah. And like, also like to mention is the partnership with Agora, of course, because we understand for people building on chips, it's the software part. That's also incredibly important. So like being able to partner with an amazing company like Agora allows even lower ultra-low latency when people are interacting with the device. So that feels more like a real-time conversation. And also we support like barge-in

so that people can intervene and talk over the device when it's talking so that people don't feel like, I'm just talking to a robot and I have to wait for it to finish before I ask them this next question.

Hermes Frangoudis (13:12)
It has that human feeling of conversation where it's very like dialogue driven and you can interrupt the person, right? No, that's super cool. And thank you for bringing up Agora. Like we're excited for this partnership. It's been really cool to work with your team on this and the ability to bring voice to such like a small device, right? Like I have one here. It's tiny like, and I think

Diana Zhu (13:15)
Exactly.

Right. Right. Right. Exactly.

Yeah.

Hermes Frangoudis (13:36)
This whole bottom piece is like the speaker. Like the board itself is this, right? That's super cool. It's such a small form factor.

Diana Zhu (13:41)
Yes.

Exactly.

And we can also, when you're making different devices, you can also disassemble it and assemble it so that different electronic parts fall in different pieces. And actually right now, I know I'm learning a lot more about hardware and also computational AI because I'm building my passion project. That's a voice device for early childhood development, a toy that can interact with children and introduce them words. And through interactive stories, very much dialogue-driven because language, I guess, is like just

the natural way people are interacting with the world. So building that language component into intelligent devices is just a trend that we're seeing over time.

Hermes Frangoudis (14:22)
For sure, I mean, it's the most frictionless way to interact with a device, right? I mean, when I'm building demos, my kids come in and they see me talking to the computer and they laugh, but they also talk to it, you know, they'll jump in and, it just makes you realize like, it's that simple. There's no typing. You don't need to know how to read. It's, the most natural way of communication. It's so cool.

Diana Zhu (14:34)
Yeah. Exactly.

Right,

Right. And now Imagine a world where lots of devices can be connected. So for example,

I was just talking to a company that uses our chips ⁓ to build connected and intelligent two-wheelers. So Imagine an e-bike that can now hear voice prompts and just guide you and navigate you to places. So That will be just fantastic because then I don't have to use my phone to navigate me and put that voice on. And instead, Everything is just interacted so seamlessly through people's voices.

Hermes Frangoudis (15:12)
That's so interesting. I wonder, does it connect to your phone so that your phone can stay in the pocket, right? Like it can, yeah, that's super cool. And so you're saying you're seeing this moving not only into toys and wearables, but like connected devices in general. So like everything's gonna be able to see and hear and interact at this like natural level. You're not gonna have to like, "let me learn."

"What's that code that I need to push in or combination of buttons," right? To get this to work. Exactly. That's so cool. And at the embedded chip level, is that all with this one? This one chip like the BK7259, like this is the powerhouse. It can kind of go anywhere, do anything at that level. That's so cool.

Diana Zhu (15:41)
Exactly.

Yeah, we have a series of, our company's chips are generally divided into three categories. There's BK7259, which enables a lot of the edge AI applications. So you can run vision models locally and language models locally. Models up to a few hundred thousand parameters, they can all be run locally now instead of connecting it through cloud. So we also have other categories like the BK7258.

which is a dev board that you just showed that carries this chip. So that's a multimedia SoC, which means that it can connect to audio, video very easily, and that also can receive information. There's a microphone connection very easily accessible to developers. And there's the connectivity, outside of the multimedia chips, there's connectivity chips, which allows for Wi-Fi, Bluetooth connectivity very easily at ultra-low power. So one interesting application for the

connectivity chips that I've been seeing is that as the, I guess, AGI is coming, there's a lot more power need at the infrastructure level. So we're seeing now our chips like the BK7239 series being used in terms of these mesh Wi-Fi network to support these power plants.

Hermes Frangoudis (17:16)
Super interesting. So

they connect to the mesh network and then can intelligently kind of communicate between.

Diana Zhu (17:20)
Yeah,

Yeah, exactly. It can support up to 250+ devices to allow them to talk to each other and also report back if there are problems.

Hermes Frangoudis (17:30)
That's super cool. So you'd have like one centralized device that can communicate with all of them, get the status, aggregate that data because we think about that. That's a lot of data, right? Exactly.

Diana Zhu (17:33)
Yeah.

Right.

Exactly.

Yeah. And it's interesting because for the, like on one side, we're seeing all the connectivity chips, but on the other side, We are also moving in a direction where AI is becoming more multimodal, which means that we not only process visual information, but also like audio information.

Hermes Frangoudis (17:42)
⁓ yeah.

Diana Zhu (18:00)
So that everything can be even more intelligent. So I can recognize your face, also interact with you in terms of audio very intelligently, and also remember things about a user that I'm interacting with.

Hermes Frangoudis (18:15)
Interesting like on-board memory of ⁓ the previous interaction so you can kind of ask about the last time we spoke and the last thing we did. That's huge for like auditing for even just like personal use like being able to recall what you did previously without having to sit there and take detailed notes

Diana Zhu (18:18)
Yeah. Right.

Yeah, exactly.

Yeah, exactly. And then this is pretty big. So in like toys that can

keep us company like the company that we're usually supporting Fuzozo, I think that's a huge like if a true friend will remember shared moments with us and we want the like I guess the virtual friend that we're building here the Smart Choice to also be like that and then I think memory is a huge part of that and then in the personal passion project that I'm building with the Smart Choice for early childhood learning. There's also the

memory piece that enables adaptive learning, which means that I'll remember which words the kid has learned and which kind of scenarios or what kind of stories they like the most so that I'll tell more stories in that direction so it's faster for them to learn. And then if they learn words really fast, then I can boost up the speed of learning and pumping new information to them. So yeah, a lot of areas to explore.

Hermes Frangoudis (19:30)
That's super cool. Cause like I've built a voice AI agent that I let my kids play with, right? Very, very guard-railed, but piece is like one of the things that they constantly have to remind it is like, "Hey, I like this kind of stuff." or "I like that kind of stuff." Cause it doesn't have built in memory but if that could happen on device, it's huge. So they're not like working their way to where they want to go,

Diana Zhu (19:45)
Yes.

for sure.

Hermes Frangoudis (19:52)
they kind of start where they want to go and move forward.

So you talked a lot about on-device, which kind of covers a lot on the privacy side of it, right? And in terms of moving from like on-device to the cloud, there's going to be a, obviously for developers, way that they can keep stuff on-device, choose what goes up to the cloud, right? Like that's programmable within, in this pack. So how do you balance kind of like the latency on that,

Diana Zhu (20:13)
Yeah, for sure.

Hermes Frangoudis (20:18)
what's the edge?

Diana Zhu (20:19)
Yeah, so maybe it'll be helpful to just walk through the normal process of what happens when we want to run, for example, a Large Language Model in the cloud

while keeping some stuff on the edge and what happens on the edge versus in the cloud. So let's just take the BK7258 dev board that you showed just now. What happens is when there's a wake word, like, "Hey Agora", "Hey, Armino", it hears it in the Wi-Fi keep-alive mode. And then it does local processing. So then it tries to filter out the noise that it hears, the background noise

and keeps your voice as the thing that it wants to take information from. And then locally, it does audio processing so that it translates the audio to text.

And if audio is too complicated, sometimes then we will make a decision to compress the audio locally and then securely send it to a cloud location so that it can be processed over the cloud. And then it can connect to the LLM models, which then sends a response back. Then on the device there, the TTS, which is the text- to-speech happening, then the user hears the voice

back

to the information provided by the LLM model in the cloud. And to reduce latency, there's lots of chip on the device kind of things that needs to happen. For example, how do we get rid of background noise so that it's most clear what message we're delivering to the device? And then also, how do we most efficiently decide which information is sent to the

cloud so that it can reduce the latency the most.

Hermes Frangoudis (21:59)
Okay, that makes sense and it's really a lot of that is programmable on the board in terms of, that's super cool and I totally make sense having so much of it happen really on the edge, on that board. You're just doing what you you absolutely need to so only incurring costs, I guess.

Diana Zhu (22:05)
Yeah.

Yeah.

Hermes Frangoudis (22:20)
You know,

When you think about, from a developer perspective, you're only incurring those API costs when it's absolutely kind of critical to close the gap.

Diana Zhu (22:28)
Right, exactly. And then with the BK7259, for example, we are delegating even more stuff on the edge because the processing power is bigger. So there's more stuff that can be

processed on the edge, even with the smaller language models. That's not like a ChatGPT-5 in the cloud, but still very powerful. It can recognize not just faces, like the smart door lock made with 7259. It can unlock using someone's face under 200 milliseconds, which that's like one-fifth of a second, which is no time. And then there is like, you can interact with the voice and then all the information

is on the edge, give a voice command, open the door or like lock the door or whatever. So this all can happen on the edge instead. And then there's like, for example, health kind of wearables. We can process the sensor information and detect if there's anything that's abnormal also with models running on the edge.

Hermes Frangoudis (23:26)
So it happens right there, right? Like, cause sensor data, not only is it sensitive, but it's like time-critical. So if you're talking about like going off to the cloud and you lose critical timing on that. But that also be interesting. That's also interesting to think about like the wearable that supports you from like the health perspective. Like I remember my grandparents growing up, they had medications, right? And there's so many of them at some point.

Diana Zhu (23:28)
Yeah.

Right. Exactly.

Yeah.

Hermes Frangoudis (23:51)
And keeping, making sure like you get this many of this and that many of that, the device can see that, recognize it without, you know, going up into the cloud and kind of speak to them naturally instead of some weird like buzzer or something that, feels awkward. It now becomes like an assistant. Right.

Diana Zhu (23:55)
Right. ⁓

Yeah, yeah,

Yeah or even like, for example, the device is detecting your heart rate or your like glucose level, and then it can in real time give you feedback of like what's happening and this can be lifesaving in some scenarios.

Hermes Frangoudis (24:22)
And it's got memory. Hey, maybe you should go eat that apple and get your sugars back up or something.

Diana Zhu (24:24)
Yeah.

All right.

I'm still in the morning. ⁓

Hermes Frangoudis (24:33)
You talk about your passion project, which is very interesting, right? Like toys, kid-safe devices, from a design standpoint, aside from, just like processing on device. Are there any other, features within the boards for enabling kind of like these guardrails or defining certain, maybe prompts or things

that could be more for safety than anything else.

Diana Zhu (24:56)
Yeah, definitely. So I would think about like safety and design from like two perspectives. There's like the hardware component and there's a software component. For hardware component, for anything that interacts with children and humans, in general. There's like no sharp edges, no small parts so that they don't choke on it. But still like enables rich interaction. And what we're thinking is like for this toy, because we hear so many parents complain about because like after a long day they don't have

like energy or ideas to interact with their kids even though they really want to. They end up like iPads or like screens, their phones end up like filling in the gap which research has shown that this is like quite bad for  children's growth. And so like can we have a more next-generation of toys that are very intelligent but still foster dialogue? When your child is like seeking play

can like the education piece can both be fun and also like help the children learn something. So like on the software piece, what we're thinking is, we do have like a little screen that helps the children see the picture book. So it acts like an interactive reading engine where with this like little elephant, for example, it will ask questions about the elephant and relate it back to the children's lives. It's like "This elephant has big ears. Do you have big ears?".

And then also like teaching abstract concepts like "This elephant looks sad" and then "This squirrel is trying to comfort the elephant". So it talks about feelings so that and describes actions to help boost how they express themselves and how they describe feelings, how they relate to characters in the stories all through dialogue. So yeah, these are all things that we put into consideration when we're developing this toy.

And

Then we're developing it, because I did my PhD at Yale, my advisor actually works on early childhood development. And then one day when we were talking about what I'm working on now these days, he was like, wow, what if we can make an AI? Because they have this training program for like 10 years. And it's been like award-winning and really shown to improve language ability within a short span of like 4 to 6 weeks.

so,

Like right now, the training program is very expensive because parents or caretakers need to spend at least like three days in their workshops, and especially for working parents, parents of lower-income, they don't have that time to spend, like, just training how they interact with their children. So, but now, like with all the technology, we can make that a possibility. And with the chip, it can enable very low latency natural conversations, especially with

with Agora's add-on, even lower ultra latency real-life conversations that can happen just like when they're talking to their peers.

Hermes Frangoudis (27:38)
No, I think that makes, that makes a huge difference because a child doesn't really understand the signals going over through a processor and getting analyzed. They're just like, why isn't talking back to me fast enough? But when you're talking about processing on-device, offloading minimally to the cloud with ultra-low latency deliveries, like that gives it that natural feel, the kind of feel that makes them want to keep interacting

Diana Zhu (27:52)
Right. Exactly.

in a way that's not like difficult. Right. And because they want to, it will adapt to their kind of interests with today's AI models. They can remember what they already learned and what area prompts the most response. So then it will go more into that direction to explore and help them learn even further.

Hermes Frangoudis (28:04)
in a way that's not like difficult.

It's really blurring the line on what can be considered AI-enabled. Soon everything's gonna be able to be just powered and enabled by AI in a way that will only help improve us, right?

Diana Zhu (28:39)
Yeah, exactly.

Hermes Frangoudis (28:40)
So from a developer perspective, I do want to help promote some of the really cool things that not only RiseLink's doing, but we're doing together, right? We said we have the BK7258 dev board. If a developer wants to get started, where should they go? Do they go to your website?

Diana Zhu (28:53)
Yes.

Yeah, yeah, well, obviously, I would love to hear from developers. So reach out to us and then we'll definitely respond. But I think a great place to start will be like from the SDKs and also the developer kit, because, for example, this dev board is ready to be tested if you want to.

With the little eyes, it can already serve as like if you're making a little toy, it can already get you started very quickly. And also it can be disassembled and you can just order like the SoC itself and then connect to your audio, video like yourself as well. We have GitHub online and also engineers waiting to hear from you. So yeah, I would start there.

Hermes Frangoudis (29:40)
And, you mentioned to reach out to you. You're actively supporting developers in building physical devices. That's really cool as a company. Can you tell us more about that?

Diana Zhu (29:48)
That's right.

Yeah, definitely. So we are actively supporting local builders and startups that are building all kinds of like really fun things. I just talked to a startup that does Rubik's Cube, which they for each little tiny cube, they're making it intelligent and they have like pictures on it and then you can play games with it. So instead of the static ones, now everything is electronic. And then like very soon,

this Rubik's cube will have a personality inside it and can interact with you. So that's very fun. And part of the reason I was building the AI toy is also because I wanted to go through this process of what are the challenges a developer might face when they're developing on our chips. And yeah, it's been really just seeing how supportive the team members are. I think also understanding more of the challenges of building hardware

to make sure like how like sometimes you have to pay the first upfront cost. How do you, like even to build the prototype on like software now it's so easy to like code up an app and very low cost but for hardware is still kind of costly. So how do we help developers support them in achieving their ideas. I think that would be a big topic for us and we're hoping to we're also talking to high schools, universities to

bring up more events so that people can get that hands-on experience developing on the chips with the support from us. There are hackathons that we're planning in the next year so that students, not only hobbyists who are already engineers in this space, they can get access to building stuff around their household.

Hermes Frangoudis (31:26)
No, you bring up a really good point in terms of like there's a certain investment you need to make just to like, kind of get into this world. And the fact that your team offers something so well put together where the developer doesn't have to sit there and piece together, right? Like from a bunch of Arduinos with separate hardwares and you got to go find custom chips and things like that.

Diana Zhu (31:44)
Right. Yeah.

Hermes Frangoudis (31:50)
This chip is custom. This is perfect for that developer. Don't even bother trying to do it yourself. Get one of these, right?

Diana Zhu (31:57)
Yeah, exactly.

Yeah, at the Maker Faire the other day, actually there was a developer that's now developing like a...

intelligent spider. So you just 3D print the spider and then put our chip inside and then the spider can talk to you and also like be connected to Wi-Fi and you can program it to do different like gestures for Halloween. I think I feel like people's imagination just goes beyond my, you know, what I thought could be possible and fun little things can all be built now and want to support that community growth here.

Hermes Frangoudis (32:26)
That's amazing. You've lowered the barrier so much that they're building even like novel household decorations with AI enabled boards. It's super cool. And I love the fact that, you know, dogfooding is so important in this industry. The fact that you're using the chip and going through the developer experience is huge for, you know, your developer community because that

Diana Zhu (32:33)
Yeah. Right. Exactly.

Hermes Frangoudis (32:51)
instills that trust. You understand the same pain points that they're having when dealing with it because that's how software is like. Most people, they have a new idea. It's something first time being applied to this use case. So how do you support and kind of get them over their own hurdles of just like getting their idea across the line.

Diana Zhu (33:09)
Yeah, exactly. Especially because building these days requires both the AI piece, the software piece, for example, like how do we build the AI agent that would interact with you in a way that you desire? How do I choose the model? Like, which kind of model do I choose? And then how to optimize my token use. But this is like later down the line. And then coming to the hardware piece, can it perform well in a noisy environment? Which I think

is like a huge topic for both Agora and RiseLink because like a lot of the applications are in noisy environments and then how does it distinguish words that people are saying from, like little children is different from like teenagers versus like adults and then all these like little concerns that you're probably like I just didn't think about before I actually hands-on building this yeah and then like how do you

3D print, what kind of character would you want to 3D print and how do you lay out the electronics components inside so that the toy can actually hear you when the electronic components are inside its belly, for example. So yeah, there are so many like little details that you wouldn't think about as you're actually building something until you actually build it.

Hermes Frangoudis (34:20)
No, you're right. It's one of those things like the deeper you go, the more you learn and it's super cool.

Diana Zhu (34:25)
Right, Yeah.

Hermes Frangoudis (34:27)
So looking ahead, where do you see conversational AI and hardware kind of evolving to over the next 1 to 3 years, we'll call it, because 3 to 5 years is way too long.

Diana Zhu (34:39)
Yeah, well, I think it's going to be more and more intertwined.

So like, in terms of conversational AI and chip design, because conversational AI models, they're advancing really rapidly in quality and decreasing in size as we're optimizing it. On the other hand, chip design for AI is focusing on making it more feasible to run surprisingly powerful models locally so that we, like again, continue to decrease the latency and also improve privacy so that nothing needs to be uploaded to cloud if we don't choose to.

I think pretty soon in a few years, we'll be able to see chips that can run surprisingly powerful models on the edge that will enable much like without having to connect to like cloud LLMs. And then also like something we talked about earlier

in our podcast, there's the multimodal AI becoming more popular that can process both visual information and also audio information together. So when we're interacting with our door lock, even we wanted to hear naturally and also recognize our faces when we are building a baby monitor, for example, we wanted to taking all these different information and be able to

recognize faces and do pattern recognition and all that. And then building choice similarly, we want to, I guess, the natural interaction just both visually and also in terms of audio. So I think we're going to see a big advancement in multimodal AI becoming more prevalent as well as our GPU, CPU advance. So that's another direction. I guess one other thing is

which is, since we talked about memory, I think memory is also a huge part and the bandwidth on the chips. So to make it more personalizable and then make it as memory tech is growing very fast.

It may be more common that little chips can store a lot of memory information so that your device can truly be personalized and know you specifically.

Hermes Frangoudis (36:26)
you

Adjust over time learn learn kind of from your behaviors and learn to support you in the way you need. No, that's super cool and I'm gonna agree with you. The future is multimodal. Yeah,

Diana Zhu (36:46)
Yes,

that's how humans interact or are used to interact.

Hermes Frangoudis (36:50)
Yeah, we are multimodal, we interact on so many different levels and who knows, one day,

these things will probably also have a sensor for touch, right?

Diana Zhu (37:00)
No, exactly. like when we're,

I guess, talking to the child development experts and also parents, I think for babies, also for adults, I think touch is a very important aspect for things. Like when the baby touches something, can we recognize them and then also use that as a source of input to say for example, baby touches the elephant's ear and then we say something about the ear because we know they are intrigued by this. So yeah, you're absolutely right.

Hermes Frangoudis (37:24)
Super cool, using that as a learning tool right? Because now screens are touchscreen, you can understand where they're applying pressure, what's around them, what's the environment, the mood. Yeah, makes sense like that multimodal future is going to bring it all.

Diana Zhu (37:36)
Yeah.

Right, right. And that all comes down to how much processing power you have on the chip and also balancing it with how much power it uses.

Hermes Frangoudis (37:47)
This is just the infancy. And if you have this sort of power efficiency at this point, it's going to be awesome to see.

Diana Zhu (37:53)
Yeah, we're

excited to have this feature together.

Hermes Frangoudis (37:56)
So do you think there's gonna be a world where like everything has a voice and maybe eyes or the ability to like, like everything we touch will be interactive?

Diana Zhu (38:08)
Yeah, I think that's the future. Well, at least I'm really intrigued by. the intro.

Intelligent two-wheelers I was talking I can't wait to have that actually connected, be actually connected intelligence so it can respond to me via voice instead of like I have to very artificially like scan my app and also like type in my address so I can free my hands and actually like interact with all the little everyday objects. Like a real living thing if we put it that way. And then also like for example like coffee,

sensors built-in coffee machines to have it tell you when your cup is filled, for example. And also smart fridges that tell me when my leftover is going bad so that I don't food poison anyone. Yeah.

Hermes Frangoudis (38:53)
No, no, that's

a huge one. It's like, wait, when did I make this? And the fridge could be like, you put this in here on this night.

Diana Zhu (39:00)
Yeah, and don't buy more milk because we already have like three bottles.

Hermes Frangoudis (39:03)
That's the other one is when you come home

and you start putting away the groceries, you're like, ⁓ I forgot to cross this off.

Diana Zhu (39:10)
Yeah,

exactly. And they can send you a reminder and be like, get more vegetables because you don't have any left.

Hermes Frangoudis (39:17)
That's huge. I've been enjoying this conversation. It's been going so well, but we have blown through our time and I want to be cognitive of your ⁓ time.

Diana Zhu (39:26)
Yeah, was so

It was so fun.

Hermes Frangoudis (39:28)
It's been a great conversation. I have one question though, and this is kind of our wildcard that we ask at the end. If you weren't at RiseLink, what aspect of the AI industry or what would you be doing? Where could you see yourself?

Diana Zhu (39:31)
Yes.

Yeah, I would definitely think more and more the  education space is like AI education space is like really becoming really intriguing. And also because of like parents dislike of screens and also it's like actual objective, really how it's bad for children, like more fun devices, toys, ways to play for children that don't involve like, you know, playing with an iPad or an iPhone. I think like the toy industry

is ready for a revolution. Also like these days in children's education, how do we help them using like dialogues, conversational AI to help them learn new things? Yeah, I think that's a space that will be very interesting, just like what I'm doing with my passion project.

Hermes Frangoudis (40:28)
So you'd still find yourself right in here. There's no other space for you right now. That's true. I love it. Well, thank you so much for your time and thank you to all our viewers following along. And we'll, if you're on socials, do that social thing, like, subscribe, tweet, retweet, reshare, and we'll catch you on the next one.

Diana Zhu (40:31)
Exactly, yes,

Thank you.