We use cookies on this site to enhance your experience. Visit our Privacy Policy for more info.

How Lukas Biewald grew Weights & Biases from a side project to the AI revolution’s must-use service

As a child, Lukas Biewald used to devour NOVA’s science documentaries when they were broadcast on PBS. The technology that excited and captivated the young Biewald the most? Artificial intelligence. 

“The idea that computers could learn to do things on their own just seemed amazing to me,” he says. “It always kind of felt like humanity’s last project.”

That early inspiration would end up shaping Biewald’s professional arc, planting the seed that ultimately led him to launch Weights & Biases, an AI developer platform that’s now beloved by hundreds of thousands of data scientists throughout the world.

Throughout his teenage years, Biewald continued to fuel his excitement about the potential of AI and devoted himself to math and science in school. He eventually went to Stanford University, joined its AI lab, and earned a master’s degree in computer science under legendary researcher Daphne Koller

But while at Stanford, he realized just how far from reality the AI-powered technologies he’d dreamed about as a child still remained.

“I had this hard lesson: Almost nothing actually worked in practice,” he recalls.

He remembers feeling crushed by the results of DARPA’s driverless car competition, which challenged entrants to autonomously navigate 150 miles. No robot vehicle even came close. 

“I thought, ‘Wow, there’s not really many applications, outside of stock picking and ad ranking, that really work using AI,’” he says. 

After graduation, he joined Yahoo as an engineer, working on machine translations to improve search results, and started becoming fixated on one of the foundational components of AI systems: well-labeled data. 

Unsexy but crucial AI

To create a functional AI system, you need to train it on gobs of data. But for the models to understand that data, it first needs to be accurately labeled.

Precise and error-free data labeling is an unsexy but deeply important part of building artificial intelligence, and Biewald saw an opportunity to make the process more efficient. 

People labeled data slowly, but automated systems often made ruinous mistakes, so he launched CrowdFlower in 2007 to combine the two into a human-in-the-loop data labeling platform. 

Biewald ran the company for more than a decade, rebranding it as Figure Eight and selling it to Appen in March 2019 for $300M. Figure Eight had become an institution in the machine learning sector, hosting regular conferences where the leaders of AI would congregate. But he equates that time to running a marathon where you start out too fast. “We landed early adopters of AI very quickly, like eBay and Google,” he says. “But there weren’t a lot more places to go.” 

Building “something that people use and like” at OpenAI

In the final few years that Biewald ran Figure Eight, big strides had happened in artificial intelligence.

“Suddenly, all this stuff was kind of starting to work,” he says. The dream-like visions spurred by those childhood documentaries started to come back into focus for him. “I got really excited about building a company that would help behind the scenes in making stuff reliable,” he says.

“I got really excited about building a company that would help behind the scenes in making stuff reliable”

The next logical step would be to start building. Instead, he called up OpenAI, which at the time was transitioning from a non-profit research lab, and asked for an unpaid internship. 

“I was starting to feel like I was a bit out of date and a little less technical,” he says.

He also started teaching online classes about machine learning to force himself to re-learn the material he needed to stay at the cutting edge of AI. (The classes also helped Biewald pay his way: He was thinking about buying a house in San Francisco and having a child, neither of which are cheap.)

Both the internship and the classes helped Biewald in different ways. 

The classes highlighted the points of friction machine learning programmers faced when trying to tackle deep learning.

“Watching 100 people in a room – really smart people – struggle with this stuff was eye-opening,” he says. “Like, ‘Wow, this is actually really hard.”

And the internship at OpenAI, where he was assigned to a robotics project, helped him understand the issues AI developers faced when creating and adjusting models. 

Biewald started building a machine learning operations tool to solve these issues, and would bring prototypes to a bar in San Francisco’s Mission district where the OpenAI team met to tinker every Wednesday night. 

“What I really like to do is just build something people use and like,” Biewald explains. “I was making it for my friends, and I really cared a lot, so I would show it to them.” 

While the OpenAI team wasn’t initially interested in using his tool, it began catching on.  Eventually, its popularity grew to the point that employees started complaining about issues and bugs – which Biewald would quickly fix.

Weights & Biases is born

Biewald’s tool aimed to make life easier for machine learning practitioners like his friends at OpenAI. 

He’d realized through his internship and classes that AI systems require a completely new style of coding, which meant that developers needed a different model of software and best practices to be successful.  

In 2018, alongside his CrowdFlower co-founder Chris Van Pelt and Google engineer Shawn Lewis, Biewald founded Weights & Biases. They developed their first product – Experiment Tracking– in an Airbnb they rented to host a hack day.

The tool did something simple but revolutionary for AI developers struggling to stress-test and check their models: It showed them exactly how those models worked.

The Weights & Biases founding team made what many would see as a rookie mistake, but which paid off in the long run. They almost singularly developed their product to meet the needs of one customer. 

“We felt a little insecure,” Biewald says. The startup was hostage to fortune based on OpenAI’s success. 

But it paid off, in large part because OpenAI was operating on such a large scale – and at a more advanced level – than everyone else in the market at the time. That meant that when competitors caught up, Weights & Biases was ready to serve them. The company now counts Meta, Samsung, and Spotify among its customers. 

For years, Biewald continued teaching his AI class while building Weights & Biases. It helped him appreciate what programmers at both ends of the market would need from his tools, from beginners to advanced users, like W&B’s corporate customers.

“I started to use it myself more and more in the classes,” he says. “I really felt the customer pains so deeply.”

The uniting desire with both beginners and enterprise clients? They “both want things to be clear and simple,” according to Biewald. “There is more overlap there than you might think.”

The investor POV

George Mathew, Investor and Managing Director at Insight Partners

“Why aren’t there better product experiences for machine learning practitioners?”

The company began getting noticed. Unlike with CrowdFlower – where Biewald would pitch 100 investors, 99 would say no, and one would say maybe – he was now facing inbound requests. 

“We just kind of hit it off,” says George Mathew, managing director at Insight Partners. “I think that he – as a boundary builder – saw his path, and I – as someone who scaled data and analytics companies – saw mine.” 

Mathew and Biewald had met before. The former had spoken at conferences organized by Biewald during his time at CrowdFlower.

Mathew also had extensive knowledge of the frustrations that machine learning developers felt when building their products – frustrations that Biewald had built his company to solve. 

“Why aren’t there better tools? Why aren’t there better product experiences really targeted at machine learning practitioners? And what I found was that most of the software experiences were not that terribly delightful, to be frank,” Mathew says. That was until Weights & Biases appeared.

“It was kind of the first company, product, founder, and team that I saw that was really targeting the developer persona with a delightful experience,” he says. Insight Partners led a $45 million Series B round for the company in February 2021.

From one customer to 700+

The funding and support from Insight Partners has helped Weights & Biases grow enormously. 

Before its Series B, the company was mostly engineers, but started adding roles in finance, R&D, and sales – alongside managers to help grow the business sustainably. Biewald, meanwhile, had to stop taking shifts on the company’s customer service chatbot (which he admits he misses).

“That’s the moment we flipped from being a startup project to a scaling, growing company,” says Biewald.

“We flipped from being a startup project to a scaling, growing company.” 

Since Insight’s investment, the company’s ARR has increased 20x.

“Anyone who’s building their models for scale and deployment is using Weights & Biases as a de facto tool,” says Mathew. 

And that loyalty is likely to grow as the company meets the shifting demands of its customer base. The firm recently announced it was supporting large language model (LLM) application development, recognizing the outsized role it currently has in the AI revolution. 

“The surface area of Weights & Biases is going from a targeted product that really handles the core of how machine learning practitioners build their models, to a platform that serves the entire lifecycle of machine learning,” says Mathew. “At scale.”

And Biewald is committed to following his north star of listening to customers and adapting to their evolving needs. 

“I really do want to make stuff that people use and like,” he says. “That sustains me.” He calls picking a target market that you like and admire an “underrated way to do business.”

“I think it’s wise as a CEO to start with a group that you actually like, because you have to spend time with your customers,” he says. “So pick ones whose company you enjoy, that you want to get a beer with. Because if you have a connection with the customers, you can learn a lot more from them.”

Interviewed by Chris Stokel-Walker for Insight Partners.

How a casual conversation over whisky became Fermyon, a business that’s revolutionizing cloud computing

The idea for Fermyon, a startup that is defining the next wave of cloud computing, was born during a Microsoft offsite in British Columbia, after-hours and over maple-infused whisky. 

It was 2018 and Matt Butcher and his team found themselves chatting over drinks about the challenges of cloud computing. 

“We were sitting there, on this not very good whisky, and just got into this creative zone,” says Butcher.

He and several colleagues joined Microsoft in 2017 when it acquired their company, Deis, which created popular tools for building Kubernetes apps. Butcher had become a principal software development engineer on a de facto open-source R&D team inside Microsoft Azure.

“We had two jobs when we were there,” he recalls. “The first was to help Microsoft repair its reputation in the open-source ecosystem. And the other was to go out there, talk to customers, and find out what was hindering them as they tried to adopt cloud technologies. And then build the right tools to solve these problems for people.”

The team discovered a long list of problems but a short list of solutions.

That was because the current systems and how they worked in the cloud weren’t suited to the needs of the present day. Broadly, there were two kinds of cloud computing in use – virtual machines and containers – and neither solved the interaction problems raised by customers.

Dreaming up a new kind of cloud computing

So over the too-sweet whisky, Butcher and his colleagues began sketching out a third cloud computing model. They came up with a set of characteristics they felt it needed to have — instant startup speed was near the top of the list, as was an ability to span architecture across time.

Because Fridays on the team were devoted to researching new areas, the group dedicated time each week to figuring out this new model of cloud computing. 

Butcher had recognized, from years of meeting Microsoft’s cloud computing users, that the biggest hardship wasn’t grid management or power consumption but capacity. Users can rack servers as fast as possible and fill them quickly – but most of those servers aren’t actively running workloads. 

They’re being “used,” technically, but aren’t doing anything. “What we discovered was the way we have been using computing is fairly inefficient,” says Butcher. “Most of our systems sit around idle.” 

What would become Fermyon solved that inefficiency. A serverless and faster cloud, it is architected from the ground up to compile and ship code in WebAssembly binaries. WebAssembly can execute code at near-native speed in the browser or on servers, making it faster, with higher performance, than JavaScript for building web applications. 

Rather than relying on servers that sit idle and need time to spin up, Fermyon can instead execute code extremely fast. This then frees up resources for another application to use servers to run at a much higher efficiency instead of sitting idle for long periods of time. 

Microsoft’s reaction to the idea planted the startup seed

Around 10 out of Butcher’s 14-strong team were working on the idea at its peak, eventually reaching a point in early 2021 that they felt ready to pitch it to Microsoft higher-ups. Butcher’s audacious plan was to overhaul his team to break out of their current mandate and focus on this new paradigm.

But the reaction within Microsoft was lukewarm at best. 

Azure was scaling enormously and providing great results for the business. And this serverless option would not only compete with it but potentially nullify the need for it. “The meeting ended with them going, ‘You know, the current technology is rocketing upward. It’s our fastest-growing service. And you’re suggesting you want to go off and build a competitor to this? I think it’s premature to do that.’”

However, Butcher was taken aside by his boss – Brendan Burns, the creator of Kubernetes – after the meeting, who asked whether he had ever considered starting a company. “I have no idea to this day if he was ribbing me,” admits Butcher. “He has a very droll sense of humor.”

As part of his role at Microsoft, he consulted with venture capital firms. So he tapped that network for advice on launching a company.

“I thought it was going to take years to do this,” says Butcher. “It always seemed like a big, insurmountable hill to build a company.” Instead, Butcher talked to three VCs and rapidly received two term sheets.

He and his co-founder, Radu Matei, decided to approach their Microsoft team.

They did so on a Friday evening around 7 pm. “We laid out the vision for the company and all that,” says Butcher. “And we all agreed we were going to shop it out to VCs.” 

“I thought it was going to take years to do this. It always seemed like a big, insurmountable hill to build a company.”

Matei and Butcher asked their team members at Microsoft to take the weekend to decide on whether they’d join Fermyon or stick with the tech giant.

“The last thing I said was, if you’ve got any questions, reach out to Radu, reach out to me. Otherwise we’ll reconvene at 7 pm Monday night and see how things are going,” recalls Butcher.

His announcement was met with silence. The two co-founders began texting each other, wondering if they had made a massive mistake. 

But then, by Monday at 7 pm, almost everybody said yes. “We had ten people who left Microsoft all on the same day, which I think caused my boss near cardiac arrest,” says Butcher. Fermyon was born on October 21, 2021.

Things moved quickly from there. The company closed its $6 million seed round by the end of that year.

Winning over the Hacker News crowd

Assembling and trusting a team allowed Butcher to focus on the important elements of building the business. On March 31, 2022, the company released its first open-source product, Spin, a developer tool for building WebAssembly microservices and web applications. 

Spin was designed with a specific audience in mind:developers who wanted to build serverless applications but were thwarted by the complexity of doing so on AWS, GCP, or Azure. The Fermyon team found many of these folks were quite active on Hacker News. 

“They’re obnoxiously happy about something when they like it and obnoxiously antagonistic when they don’t,” says Butcher. “But they’re always trying cutting-edge stuff. So we knew if we could find a foothold in that crowd and win them over, we’d have a nice path forward.”

“If we could find a foothold in the Hacker News crowd and win them over, we’d have a nice path forward”

Fermyon launched its blog about Spin in February 2022, and hype ensued. It gained traction and peaked at number three on the most-read list. 

“Our core user story for 2022 was that, as a developer, you could go from a blinking cursor to a deployed application in two minutes or less,” says Butcher. By December 2022, he and the team had managed to get that down to 66 seconds and launched a cloud platform.

The investor POV

Michael Yamnitsky Insight Partners

“There was a really strong connection and common bond over how WebAssembly could be the next wave of cloud computing”

Fermyon also attracted the attention of Insight Partners. 

Michael Yamnitsky, managing director at Insight Partners, was tracking WebAssembly since his time at Datadog, a company that ran one of the largest Kubernetes clusters in the world. 

“I was blown away by WebAssembly’s potential to simplify serverless application development and allow developers to build polyglot applications,” says Yamnitsky.  

Fermyon and Insight Partners first met in March of 2022. Fermyon had a bare-bones skeleton of its Spin framework to show. 

“We have a shared vision for where serverless development could go with WebAssembly. We also both really like a good loaf of rye!” says Yamnitsky. Knowing Butcher was a fan, Yamnitsky brought him a loaf from Tartine bakery in San Francisco, carrying it with him on the plane to Boulder, Colorado. 

“There was a really strong connection and common bond” 

By May 2022, the pair met up in Valencia, Spain, at KubeCon. They went out for dinner in the Spanish city, sketching out plans for the future of the business on napkins. “It made it clear that there was a good fit here to work together on this business,” says Yamnitsky. “A couple of weeks later, we proposed a Series A investment, and the rest is history.”

Insight Partners team led a $20 million funding round in October 2022.

Gambles, roadmaps, and the future of Fermyon

Since its investment, Insight Partners has helped Fermyon find a lucrative new target audience by introducing them to developer teams at companies of all shapes and sizes.  

Fermyon wooed the Hacker News crowd, but it needed to zero in the broader cloud computing market. So, the company began building out a paid tier that offered a full platform and support for teams in all sorts of industries to harness the power of serverless computing — with the ease of going from blinking cursor to fully deployed application in under 60 seconds.

“We wanted to make it possible that the day people started using these kinds of features, they would have the option to upgrade to a paid tier,” says Butcher. The upgraded, serverless cloud option was designed to coax people who were testing WebAssembly to put their applications into production.

But the company is confident, as is Yamnitsky. “This was a team that has real expertise in WebAssembly and in cloud computing,” he says. 

“This team has real expertise in WebAssembly and in cloud computing”

Throughout its development, Fermyon has drawn a hopeful roadmap for growth and has met or exceeded every milestone. 

That team of ten ex-Microsoft engineers who left on the same day now nears 30. The plan is to grow to a team of 32 to 34 by the end of 2023. “And we’ll probably have to hire engineering managers next year, which is a big thing,” says Butcher. 

The aim is to hit a significant revenue milestone by Q4 2023 and to start seeing the fruits of that labor – exponential growth – by early next year. 

“The team’s pace of development is extremely fast,” says Yamnitsky. “They have managed to assemble a very strong team and have outpaced our own internal expectations for how quickly they could bring it on.”

Interviewed by Chris Stokel-Walker for Insight Partners.

[REPORT] How top banks are experimenting with generative AI and how it’s reshaping financial services

This post is a special feature of Insights Distilled, a weekly tech-focused email briefing read by 1,000+ financial services executives. Learn more and sign up here. 

We’re on the brink of an AI revolution.  

Generative artificial intelligence – or AI models trained on enormous datasets that can create new content resembling human output – is set to disrupt financial services as we know it.  

There’s been a whirlwind of news, experiments, and progress over the past year that has shown how this technology could help FinServs increase operational efficiency, improve personalization to serve customers better, and uncover unique insights hidden within gobs of data in record time. 

Based on close coverage of the topic, Insight Partners’ $4 billion investment in artificial intelligence since 2014, and interviews with execs, analysts, and experts, Insights Distilled has compiled the top areas of potential for financial services – and how those use cases are already being explored by innovative firms.  

For example, Morgan Stanley, JPMorgan, Goldman Sachs, and Zurich Insurance are already experimenting with ways to use GenAI in wealth management, customer service, code creation, and more.  

How top finservs are using generative AI for efficiency, unique insights, and backend technology.

“There’s a mad scramble right now to figure this out,” as the CEO of Evident AI, Alexandra Mousavizadeh, put it, including pressure from board members and senior leaders.  

But the process isn’t easy.  

Evident created an index earlier this year that ranks the 23 largest banks on their AI preparedness, and Mousavizadeh warns that firms eager to use generative AI need to deal with their fundamentals, first. 

“If you don’t have your content management really squeaky clean, it’s really hard to run a large language model on it,” Mousavizadeh said. “You may be able to identify lots of use cases, but if your content management isn’t in order, if your data isn’t labeled and tagged appropriately, if people aren’t trained correctly, it’s very hard.”  

To that point, an Insight IGNITE survey of more than 300 execs from earlier this year found that generative AI ranked below other artificial intelligence and data categories.

Enterprises are still prioritizing their data and analytics tools, over generative AI.

To take advantage of the enormous potential, executives should lay out a clear technology strategy to prepare for generative AI that includes data engineering and pipelines, MLOps tools, and finding (or upskilling) the right talent. 

For much more context on how FinServs like Deutsche Bank and BNY Mellon are experimenting with generative AI, download the full report here.