CoreWeave Fully Connected 2026 | Lawrence Liu, Willow

Clips
More from CoreWeave Fully Connected 2026

Lawrence Liu

Co-Founder & CTO

Willow

play_circle_outline Willow Voice: AI voice dictation app enabling speech-to-text across desktop apps

play_circle_outline Voice-Created Codex/Instinct Agents and WillowScribe Caps-Lock Dictation for Auto-Composed Communications

play_circle_outline Robust On-Device ASR for Real-World Audio with Low WER and Compact LLM for Post-Processing and Learning

play_circle_outline Personal AI Executive Assistant for Knowledge Workers — Freemium Bottom-Up Growth Replacing Keyboards

Info
Transcript

Lawrence Liu, Willow

Lawrence Liu

Co-Founder & CTO Willow

In this interview from Fully Connected 2026, Lawrence Liu, co-founder and chief technology officer of Willow Inc., joins theCUBE's John Furrier and Dave Vellante to discuss how CoreWeave's infrastructure is powering Willow's AI-driven voice dictation platform at production scale. Liu explains how Willow lets knowledge workers prompt agents and compose communications by voice across tools like Cursor, Claude, Gmail and Slack. He details the company's reliance on CoreWeave's reinforcement learning infrastructure, which abstracts away the underlying compute so W... Read more

explore Keep Exploring

What is Willow Voice, how do people use it, and how does it differ from ChatGPT's voice interface? add

What are the primary ways people use Willow? add

What makes your solution better than those from large frontier-model companies? add

- What is your vision for the product/company? - How do companies engage with you and how do you commercialize the product (your business model/sales motion)? - Do you provide infrastructure or development support (e.g., must customers host an MCP server themselves)? add

bolt Powered by CUBE AI

Lawrence Liu, Willow

search

(INTRO)

John Furrier

>> Welcome back, everyone, to theCUBE's livestream here in San Francisco for CoreWeave's Fully Connected Conference. We should call it the Fully Connected Intelligence Conference. A little bit off-brand there, Dave. Dave Vellante is my co-host. Lawrence Liu, he is co-founder and CTO of Willow Voice. You guys are doing some cool stuff with voice and you've got the agentic piece. Welcome to the show. Appreciate it.

>> Yeah, thank you guys for having me.

John Furrier

>> Yeah, we love the AI native. And what I love about CoreWeave is that they've really identified the system, the entire system that's kind of made for the AI native. It's not just cloud compute, it's got everything. You guys are a customer using it for your app. Explain what you're working on. Explain what you guys do, how you do it, what's the value?

>> Yeah. So Willow Voice is an AI voice dictation application. A few hundred thousand people use us and they use us to dictate into apps like Cursor, Claude, Gmail, Slack, basically anywhere on your computer there's a text box, you can now use your voice. Have you guys tried anything similar

John Furrier

>> before? ChatGPT has that beautiful voice

Dave Vellante

>> interface.It's sick,

>> right?It's more of

John Furrier

>> a—

Dave Vellante

>> Is it?Oh, yeah. You talk to it all the

John Furrier

>> time.I gotta do that. I use ChatGPT

>> voice.

John Furrier

>> Yeah.Input. But that's input.

John Furrier

>> Yeah.I don't think it's

Dave Vellante

>> dictation.So what's different? So I'll talk to Instinct and say, hey, send me— send an email to John and Jim about XYZ. Use the email template that I used for the last

John Furrier

>> one.

Dave Vellante

>> Boom.And it'll come back.

>> When people use Willow, they use it for two primary purposes. The first being prompting agents, or imagine you're talking to Codex or Instinct. Willow has a great interface. It does formatting correctly, has really low word error rate. So what you say is exactly what comes out. And I'm sure you've probably felt some of the pain before with ChatGPT voice mode.

John Furrier

>> Latency and just screwed up the audio.

>> Exactly.

Dave Vellante

>> Yeah, yeah. It's not a smooth experience.

>> Exactly. And with more and more people adopting agents, We find that voice is becoming the primary modality for how most people interact with their computer. The second part in which Willow is really useful is people love using us for any communication, whether that be in Slack or in Gmail. And a big part of that is this feature in our product called WillowScribe. The way WillowScribe works is you hold down the Caps Lock key and you just say whatever you want top of mind, and we're able to compose it for you. And so what is composition? Have you guys tried dictating into ChatGPT and then having it create an email for you?

Dave Vellante

>> Yeah, this

John Furrier

>> is—not really.

>> Right?

Dave Vellante

>> Not happening.

John Furrier

>> The old tools have been crummy, basically.

Dave Vellante

>> I think that's what Dots is supposed to solve, right?

>> And it sounds like AI-generated, but the big reason why people typically do this is they want to offset a lot of the composition step, right? Any communication you do with other people typically involves some amount of composition. And Willow does a really great job of that. It understands your voice, your tone, what your previous edits are. It's able to output something as if you had taken the time to write it well.

John Furrier

>> Now, can I, can I use Willow to make agents? Do I just tell it to make an agent?

>> Yeah, you can, you can use Willow to dictate into Codex, have Codex create agents for you. Exactly.

Dave Vellante

>> Now, to get to the secret sauce, why are you so much better than these rich frontier model companies?

>> Yeah, 100%. it's two main things. The first is we train our own ASR model. It's a fine-tuned version, but we've optimized it for very quiet audio, so very real scenarios. Now imagine you're in a coffee shop, right? There's going to be a lot of background noise. You're probably going to be trying to whisper so people don't think you're like a freak, right? And this is the exact data that we're able to fine-tune our model on. And so we work really well in these real-world scenarios. The second, which is I think the main part of our secret sauce, is we have this really, really small LLM that post-processes everything that you say. And now it's able to format everything correctly, take your inputs and change it to exactly what you would've written.

Dave Vellante

>> Okay. So, and then it improves over time so it can learn from the edits that I make.

>> Exactly.

Dave Vellante

>> Reasoning traces of the human.

>> Exactly.

John Furrier

>> So CoreWeave talks about RL a lot, reinforcement learning. They have serverless. How is that working? What does that mean for you? How is your engineering team using—

>> Yeah, 100%.

John Furrier

>> The product.

Dave Vellante

>> Take it, Lawrence.

>> So for some background, we've been working with CoreWeave for about a year now, and it's since our company's inception, and they've done a lot of, essentially all of our deployed model training for us. Now, the great thing about CoreWeave's RL infrastructure is we don't have to worry about anything under the hood. The only thing we really need to focus on at this point is eval alignment. And I'm sure you guys know the importance of getting your evals aligned.

John Furrier

>> It's hard.

>> It's important. Exactly. And, for us, a lot of our time was, even with all the infrastructure put aside, we spent a ton of time on eval alignment. And if we didn't have, if we had the overhead of dealing with infrastructure, it would've made our team a lot less efficient. And at this point, with all of the infrastructure abstracted away, we can put checkpoints into production almost on a weekly basis.

John Furrier

>> Every team now.

>> So it's enabled us to do really, really great things.

Dave Vellante

>> So, okay, so the obvious why CoreWeave question, was it availability of compute? Were there other factors that went into that decision to use CoreWeave?

>> For us, this might sound like a cop-out answer, but it really was their support and their forward deploy team. So they partner with this company called OpenPipe, which they had acquired recently, and these guys have just been absolutely amazing. They help us understand training. They help do any sort of eval alignment tasks as well. I think that's really been the highlight.

Dave Vellante

>> I don't think that's a cop-out, by the way. I think that's consistent with other customers.

>> That's really great.

Dave Vellante

>> They say that the quality is better. It's obviously focused on AI.

>> Yeah.

Dave Vellante

>> And they're more mature than NeoCloud.

John Furrier

>> Explain the value of OpenPipe to CoreWeave. What were they doing that was so compelling I think a lot of people got this wrong. Explain more because the looping of the data is key in machine learning. Why is OpenPipe important for CoreWeave?

>> So I don't want to pitch too much on OpenPipe's behalf, but what they used to do was they did serverless SFT, right? So we could just give them the input and the output data and they'd give us a checkpoint and model which we could download the weights and deploy on any infrastructure that we wanted. They recently evolved that to RL. This was as recently as 9 months ago and they released their own sort of RL methods. And I think what CoreWeave really wanted there was they wanted a team that was able to do a lot of this forward deployed work and a team that was able to innovate on a lot of the techniques that you see a lot of companies using now when doing reinforcement learning.

John Furrier

>> On the eval, we also hear rewards are huge. Yes. How do you think about rewards? Because that's gonna be key on the task generation. You got evals, you got all the ways, and now the reward models come in. How do you think about reward models, the reward model?

>> So reward modeling, what I'll say is for us, it's a relatively easy problem. Well, for us it's just a transcript in and a rewrite out, right? And so we don't have to worry about these really complex environments or very long user trajectories. But the one thing that came to mind every single time was price. And so if you're training a reinforcement learning model, if you're using RL and you're training a model and you're using, let's say, Opus 4.1 as one of your reward models, you're going to be spending like $20,000 to train a new model.

John Furrier

>> Right.

>> And so getting lightweight models to be able to align was relatively difficult.

Dave Vellante

>> I'm paranoid about somebody stealing my voice. So talk.

John Furrier

>> Yeah, we're all over the internet, right? We're probably deepfake machine training too.

Dave Vellante

>> So, but okay, so let's say I'm not all over the internet.

John Furrier

>> That's true.

Dave Vellante

>> It's easy to steal my voice. Talk about security, how you deal with that. I've been writing a lot about the old cloud shared responsibility model. It's different now. The stakes are much higher. How do you think about security? What role do you play? What role does CoreWeave or other partners play?

>> So for us, I think dictation is one of those few things where you have to be extremely, extremely concerned about privacy, right? People are dictating into ChatGPT, dictating to their parents, they're dictating in millions of sensitive scenarios in which we cannot possibly log this. And so one of the biggest engineering challenges on our end has been how do we improve a model when we actually can't see end-user data? It ends up being a lot of research into how synthetic data gets produced. And we've come up with a pretty good pipeline that has pretty pseudo-realistic input. That's probably been the biggest engineering challenge. In terms of how we deal with security, a lot of what CoreWeave has done, which is great, is they by default have everything opt-out. So they don't train on any of your data that passes through. Everything is encrypted in transit, nothing is stored. And that's been a security guarantee that's been really helpful.

John Furrier

>> I ask these questions about the input/output about OpenPipe because you guys have transcripts, you have data, data in, data out. There's a lot of companies out there and developers who are sitting on a massive corpus of RAG data.

>> Yes.

John Furrier

>> Okay. It's been indexed. It's just sitting there.

>> Yes.

John Furrier

>> They don't know what to do with it. What they then want to do is say, okay, how do I get this into agents to identify things from that, whether it's voice or text or other things. But they all kind of have the same pattern. That's where I think the magic is coming in. What you guys are doing is interesting because you're doing that. You've kind of done it. What's your advice to people who are sitting on this rag corpus that want to come in and turn on CoreWeave or turn on an AI agent, start producing value?

>> 100%. So if you're either trying to own your own intelligence at this point, or if you're building something very domain specific, Creating these fine-tuned models is going to be your top priority, right? Frontier models typically are going to perform, but they're very expensive, and larger corporations want to own everything end-to-end. The biggest tip that I have is breaking down subjective problems into very objective truths. Now imagine you're a company that's doing vacation booking, right? How do you objectively— how do you create some sort of reward model for what a good vacation is? right? It's a very difficult problem. But if you break it down into, okay, is this spot something the user would likely go to? Is this cost within the user's range? Is the user going to most likely invite their family members? And is this something we've aligned on, right? If you can break a subjective problem down to objective truths, you're typically gonna do very well in terms of reward modeling.

John Furrier

>> Because you're tasking them out, they could be identified.

>> Exactly.

John Furrier

>> All right.

Dave Vellante

>> What's your mix on CoreWeave between training and inference? How, how can you Give us a sense of that. Is it mostly inference, mostly training, 50/50?

>> Yes. Most of our work with them has been through training. We use them as one of our inference providers currently. And so under the hood, they're, we have a few models in production that we swap between occasionally. And yes.

Dave Vellante

>> Okay. So it's predominantly training.

>> Predominantly training.

Dave Vellante

>> And presumably that's growing. And I presume inference is growing as well.

>> Yes.

Dave Vellante

>> Is inference growing faster? Is that mix shifting or is it pretty consistent?

>> Inference is growing faster. Yes.

Dave Vellante

>> Yeah. Do you see that over time as flipping? if you're 90/10 today, do you see that flipping to 10/90 or no? Or, in the fullness of time, 50/50 or don't know?

>> For us, training comes at a pretty periodic pace, which is we're updating checkpoints in production maybe a week at a time.

Dave Vellante

>> Yeah.

>> For inference, the amount of demand for voice dictation has just been growing exponentially. So that's definitely going to surpass.

Dave Vellante

>> Yeah. Okay.

John Furrier

>> What's the status of your company now? Explain. You guys started working with CoreWeave. How far along are you? Funding, team size?

>> So we're a relatively big team now, around 10 people. We had started, we raised, I like to say, a relatively lean team.

Dave Vellante

>> I love that.

John Furrier

>> From an AI standpoint, that's—

Dave Vellante

>> A valuation, that's $80 billion.

John Furrier

>> The last generation was, we have $100 billion doing 10x.

>> Yeah, we just raised around $4 million and we're seeing around 30 to 40% month over month growth. So it's been very—

Dave Vellante

>> it's been a lot.

John Furrier

>> What's your vision?

>> Yeah, so our end vision is that every single knowledge worker has their own executive assistant. And this is an executive assistant who's seen all of your communication, is super experienced. Now imagine you're trying to dispatch some work, right? Like a project. Instead of having to assign it to every single person, right, you could just tell Willow and Willow is able to dispatch that work for you. More importantly, it works conversely as well. Now imagine someone needed information for you. Willow is able to answer on your behalf and only if necessary is able to escalate as such that there's human in the loop.

Dave Vellante

>> And so I mean.

>> And so we see if every single knowledge worker has this sort of executive assistant, we can eventually become the communication layer for entire organizations.

John Furrier

>> Yeah, you connect to their Slack, you get their brain. All right, so how do companies engage with you? What kind of, how are you guys doing business today on customers? Let's just say, hey, I wanted to work with you guys and help us build the biggest CUBE app on the planet, voice activated. Give me the best clip from Dave or John or what was the main thesis of the show? You can actually create an agent for what we're doing as an example, but how would we engage with you? Is it a service?

Dave Vellante

>> You didn't know you're going to get a prospect out of this interview, did

John Furrier

>> you?We want to barter. We don't pay. We just do barter deals. No, but as a use case, as an assumption, okay, How would we engage if someone wants to build an app? How do you guys commercialize the value?

>> Yeah, so for us, most vendors engage with us on the voice dictation level. And the way that looks is maybe you guys independently install Willow Voice and start using it, right? It's a really great voice dictation product. Eventually, you guys are the executives, but you eventually, we see this being very bottom-up growth, right? And so people adopt us very naturally that way. We don't have to do any explicit outbound. Most of our leads for enterprises

John Furrier

>> happen.You get the freemium, they get usage, and then you're going to probably go buy an OpenAI-like model subscription.

John Furrier

>> Exactly.And do you guys do any development if I need to put an MCP server together, or do I not have to worry about any infrastructure?

>> You guys don't have to worry about any infrastructure. So if people want to, for example, retrieve their transcripts and use Codex to retrieve their transcripts, right, we have an ability for you to use an MCP server locally.

Dave Vellante

>> Okay. And so I presume you're, well, for AI anyway, are you exclusive to CoreWeave? I'm sure you're using Gmail, Google Workspace, whatever, but is your AI effort all on CoreWeave? Are you using hyperscale resources as well? Are you doing any of your own stuff on-prem? I don't know how to speak in quiet.

>> A lot of our work with CoreWeave has been limited to the training aspect.

John Furrier

>> Okay.

>> In the future this might change, but as of now it's mainly just been training.

Dave Vellante

>> Okay. So you've got other, obviously, infrastructure to run your company.

>> Yes.

Dave Vellante

>> Yeah. Okay.

John Furrier

>> All right. What are your goals now going forward after this event? What's on your to-do item?

>> To-do item? In 6 months or so, we want every single laptop that you see in Silicon Valley, someone speaking to it. In a year or so, we want it such that most F500s are running with Willow as their communication engine. Within 2 years, we want it such that the world, you don't even need keyboards anymore. You can just speak and your computer understands.

Dave Vellante

>> And we'll be speaking in a low voice. Exactly.

John Furrier

>> Exactly.Lawrence, congratulations. Love your venture. Thanks for coming on theCUBE.

>> Thank you guys so much. Yeah, thank

John Furrier

>> you.

>> Thank you.It's been awesome.

Dave Vellante

>> Great to have

John Furrier

>> you.Again, the startups are coming. What's happening is a 1,000 flowers blooming in this AI era. And what's happened is the innovation around a new architecture. The system is the new unit of value. It's the AI factories. It's the cloud, AI clouds like CoreWeave. And then entrepreneurs come in, kind of the old days of AWS, Dave. Yeah, very smart. We were there when it started there too. This is theCUBE. Thanks for watching.