As a key figure in the PC revolution, late Microsoft co-founder Paul Allen understood the power of new technologies to create world-changing companies.

Later in life, Allen saw the same promise in artificial intelligence. So in 2013, when he recruited University of Washington computer scientist and serial entrepreneur Oren Etzioni to launch and lead the Seattle-based Allen Institute for AI (AI2), Allen was intent on producing innovative startups in addition to cutting-edge research.

That ethos lives on at the AI2 Incubator. As a branch of the 200-person research institute, the startup incubator leverages AI2’s technical expertise and industry connections to help entrepreneurs build “AI first” companies.

“Paul wanted to see impact, and impact in the commercial world translates to up-rounds, revenue, exits,” said Etzioni, the AI2 CEO, in a recent interview in his office at AI2 headquarters on the northern edge of Lake Union. “And believe me, he’d heard a lot of hype. So I could just see Paul saying, ‘Very nice. … Show me the numbers.’ “

So here are the AI2 Incubator’s key numbers so far:

• Fifteen startups have been created in the AI2 Incubator since 2016, including seven startups last year alone.
• Two of the startups have been acquired: computer vision company Xnor.ai by Apple in 2020; and conversational language company Kitt.ai by Baidu in 2017.
• AI2 Incubator startups have raised total venture funding of $100 million, including $56 million last year.
• The collective value of AI2 Incubator startups now exceeds $500 million, half of which was added last year.

The AI2 Incubator is now preparing to raise a new pre-seed fund to create more “rocket fuel,” as Etzioni calls it, for early stage AI startups. He says this second fund will be “substantially larger” than the first, a $10 million fund raised in 2020 from well-known venture capital firms and high-profile individual investors.

Like Allen before him, Etzioni is thinking big, and pushing the team to achieve more.

“I feel like the incubator is on this tremendous growth path,” Etzioni said. “And the goal is for them to raise additional funding, and frankly, to get off my P&L,” he added with a smile and laugh, referencing a desire for the AI2 Incubator to become financially self-sustaining.

This next phase will give AI2 a chance to cement Allen’s legacy for a new era of technology, and reinforce the Seattle region’s status as a hub of artificial intelligence. It will also require the organization to overcome several key challenges.

Analyzing the incubator

From interviews with more than a dozen people in and around AI2 — startup founders, investors, computer scientists, AI2 board members, and staffers — GeekWire pieced together a picture of the startup incubator’s strengths, weaknesses, opportunities, and threats, in the spirit of a classic business analysis.

Atop the list of strengths …

• The incubator has an exclusive partnership with the AI2 research institute, a unique organization of world-class researchers known for advancing the fields of natural language processing, computer vision, common sense reasoning, and other key components of artificial intelligence.
• An experienced in-house team of businesses and technical experts helps entrepreneurs in the incubator connect with co-founders, investors, and researchers; test ideas; build teams; identify and develop customers; create a product strategy and business plan; and develop and train their AI models; among other activities.
• Entrepreneurs who build startups in the AI2 Incubator get more autonomy and give up less equity than those who go through the alternative models of startup labs and incubators, with a variety of options for financial support.
• One of the intangible benefits for entrepreneurs is the network of AI2 startup founders, now numbering more than 40 people, who connect and support one another in person and via a private Slack workspace.

In the realms of weaknesses and threats …

• AI2 is a non-profit research institution competing for AI researchers and engineers against heavily funded tech companies; and for entrepreneurs against an array of startup incubators, accelerators, studios and labs.
• It’s competing for the limelight against well-known research rivals, such as San Francisco-based OpenAI, of GPT-3 fame, fueled by a $1 billion investment from Microsoft and partnership with the Redmond company.
• And it’s competing for scarce computing resources against many of the same players. Training AI models requires immense amounts of processing power. Even in the cloud era, access to the GPUs commonly used in the field has become more difficult and costly due to global supply chain issues.
• To reach its full potential, some investors and entrepreneurs say AI2 needs to further open the pipeline between its research arm and incubator, including more startups led by researchers themselves.

The opportunity is huge, with estimates putting the potential global economic impact of AI between $13 trillion and $15.7 trillion by 2030. On the conservative end, that equates to an additional 1.2% of GDP growth per year.

Then there’s the big wild card: the ongoing uncertainty over the fate of Allen’s massive investments and projects following his death in October 2018 at age 65 following a recurrence of non-Hodgkin’s lymphoma.

Under a secretive process determined by Allen’s will and led by his sister, Jody Allen, as executor and trustee, many of his far-flung holdings in real estate, pro sports and other areas are expected to be sold or spun off.

However, AI2 is widely considered a fundamental part of Paul Allen’s legacy, along with the separate Allen Institute that focuses on brain science, cell science, and other biosciences.

The Allen estate is providing more than 95% of AI2’s total of $100 million in annual funding for 2022, and there’s a commitment from the estate to support AI2 in perpetuity, Etzioni said. The specifics are still being worked out through the creation of a 10-year plan that will define the exact amounts for 2023-2033.

“The plan assumes that the estate funding will increase beyond $100M over time, and that our own fundraising will increase as well,” Etzioni explained via email.

A representative of Vulcan Inc., the late Microsoft co-founder’s holding company, referred GeekWire back to AI2 in response to inquiries on these questions.

While there may be “a few little structural tweaks” between Vulcan and AI2, they won’t impact the incubator or startup founders, said Matt McIlwain, a Madrona Venture Group managing director with close ties to AI2 through the firm’s investments and a longtime friendship and partnership with Etzioni, who also serves as a Madrona venture partner.

“I think it’s a non-issue from the perspective of the long-term success of the incubator, and the entrepreneurs and founders that are working on projects there,” McIlwain said.

Madrona led the incubator’s first $10 million round for pre-seed startups, with participation from Silicon Valley’s Sequoia Capital and Kleiner Perkins, and New York’s Two Sigma Ventures. It’s expected to play a similar role in the upcoming second pre-seed fund.

McIlwain cited numerous instances in which the AI2 Incubator has made connections and incubated startups to position them for larger success.

One was Ozette, a biotech company that spun out of the Fred Hutchinson Cancer Research Center. Madrona led a $6 million investment in Ozette last year, with participation from AI2 and Vulcan Capital, the former Paul Allen investment firm that has since spun out on its own under the name Cercano Management.

The incubator, McIlwain said, is “really off to a great start.”

Long-term returns from early stage startups are unpredictable by nature, so AI2 can’t count on proceeds from its minority stakes in incubator startups as part of its 10-year plan. However, Etzioni said, if the incubator continues on its current trajectory, it has the potential to generate significant revenue for AI2 in the future.

Inside the incubator

After getting its start in a single room inside AI2 headquarters, the incubator is now housed a 5-minute walk away, in a 7,250-square-foot space on the other side of Seattle’s Burke-Gilman Trail. On a recent visit, we found AI2 entrepreneurs, like many in the tech industry, rediscovering the benefits of working in a shared space.

“There’s certainly camaraderie amongst all of us,” said Jamien McCullum, an AI2 entrepreneur in residence and CEO of digital health startup Measure Labs, describing the natural challenges of building any startup during a group interview in an AI2 Incubator conference room. “It’s helpful to have someone you can lean on, and cry with.”

The AI2 Incubator has a full-time, six-person team of startup, business, and engineering experts who work with founders in the program for 6 to 18 months. Another four top AI2 leaders, including Etzioni, work on the incubator part-time.

Startup and venture capital veteran Bryan Hale, one of the AI2 Incubator’s managing directors, said support from entrepreneurs who’ve been through the incubator was key to the survival of fledgling AI2 startups early in the pandemic — contributing to the run of success in fundraising and other key milestones over the past year.

The AI2 incubator team will be able to tell “some pretty bad-ass ‘we knew them when’ stories,” Hale predicted.

“It’s our fervent belief that some of these companies are going to be really big someday,” Hale said. “And we can tell stories about knowing them and helping them before they found co-founders, before they incorporated their companies, and having played a small part in that ourselves. That’s really the big draw.”

The AI2 Incubator last year expanded its prowess in biotech and healthcare with the addition of bioengineer Jenny Cronin as a principal on the incubator team.

Cronin, whose doctoral research at the University of Washington focused on brain-computer interfaces, said she’s excited to see startups apply AI to solve problems for scientists, researchers, patients, clinicians and others.

“These early stage companies are where you actually make a difference,” she said.

The AI2 Incubator’s technical director is Vu Ha, a two-time world math champion who previously co-founded AI2 Semantic Scholar, used by more than 7 million monthly users; led DARPA research projects at Honeywell Labs; and was a principal development lead for Microsoft Bing and AdCenter.

He explained that startups in the incubator are able to take advantage of cloud credits from Amazon Web Services, Google Cloud and Microsoft Azure, as well as workstations at the incubator.

But without the bankrolls of big tech companies or larger institutions, they need to be creative and scrappy in their approach, starting with pre-trained AI models in creating their minimum viable products, for example, before moving on to custom training data.

“There are certain things that are difficult if you don’t have those kinds of resources, but there’s still quite a lot of interesting problems you can solve,” Ha said.

Current and former entrepreneurs in the program say the AI2 Incubator provides resources, connections, guidance, and technical expertise that would have been difficult or impossible to access otherwise, accelerating the progress of their startups.

Many were introduced to co-founders and key investors through the incubator.

“AI2 as a brand helped me do a lot of things much, much faster than I would have done them entirely on my own,” said Gaurav Oberoi, CEO and co-founder at Lexion. The AI2 Incubator spinout makes a contract management system that uses AI to analyze corporate agreements, tracking and alerting companies to key milestones and terms.

In addition to getting access to top AI practitioners, the ability to name-drop the Allen Institute for AI helped to open doors when cold-calling potential customers, recruiting employees and pitching investors, Oberoi said.

A former SurveyMonkey vice president, Oberoi’s experience as an entrepreneur in the AI2 Incubator convinced him to put his own money into AI2’s first pre-seed fund as an investor.

AI2’s most surprising “superpower” for incubator startups may be its ability to help with hiring — not just recruiting but vetting technical candidates — said Alessya Visnjic, CEO and co-founder of WhyLabs, an AI2 Incubator spinout that develops technology to identify and prevent problems in machine learning models.

“Having the AI2 team jump in to add another voice in our interview loops has been really, really helpful,” said Visnjic, a former Amazon senior technical program manager.

Incubators vs. studios and labs

But entrepreneurs and AI2 Incubator leaders caution that it’s important to understand the inherent benefits and limitations of the incubator model. Even with all the help and support from the AI2 Incubator team, the entrepreneurs are ultimately the ones responsible for coming up with startup ideas and building their own companies.

In many ways, the AI2 Incubator itself is the in-house team’s startup, said startup and VC veteran Jacob Colker, an AI2 Incubator managing director who works with entrepreneurs on non-technical aspects of their businesses.

“Our focus is building an incubator to create that safe space, this community of people who can do that innovation work,” Colker said. “We are with those folks for a period of time on their journey, and often the catalyst for some really great teams coming together, but it’s up to the entrepreneurs to drive those startups and execute.”

This differs from startup labs or studios, where in-house teams do much of the hands-on work to conceive and build startups. Examples in Seattle include the Pioneer Square Labs startup studio and Madrona Venture Labs.

The economics can also differ substantially.

• AI2 typically receives a 9% stake in incubator companies, according to Colker. In contrast, startup labs and studios commonly get a stake of 30% or more in startups they develop.
• Entrepreneurs in the AI2 Incubator have the option to take a stipend of up to $10,000 a month, as an advance against a convertible security. This is commonly used to cover personal expenses or provide a small amount of capital to test and develop early startup ideas.
• The incubator’s pre-seed fund makes investments of up to $500,000 after the formation of a strong founding team and indications that a team is making meaningful progress. (This is the “rocket fuel.”)

“The deal here is the best deal you’ll get in the entire Seattle area,” said Jordan Ritter, co-founder and CEO of Augment AI Corp, a recent AI2 Incubator spinout. A serial entrepreneur, startup advisor, and co-founder of the original Napster music service, Ritter previously founded and ran the startup studio Ivy Softworks in Seattle.

Numerous early stage AI2 projects haven’t worked out over the years, but Colker differentiated those from fully realized startups that received outside capital.

Out of the 15 startups that received outside funding, 12 are still independent and active, and two have been acquired. The one that shut down, ClusterOne, was “unable to achieve the customer proof points necessary to raise additional funding beyond their seed round,” Colker said via email.

“While a single failure out of 15 efforts that raised meaningful venture capital is great, we acknowledge that the law of startups says that more failures will happen over time,” he added. “Of course, the real measure is what happens when these companies exit and only time will tell.”

Institute meets incubator

But a need for a stronger connection between the AI2 research institute and incubator came up repeatedly in interviews with entrepreneurs who’ve been through the incubator, and investors who’ve put money into AI2 startups.

Heather Redman, co-founder and managing partner at Seattle’s AI-focused Flying Fish Partners VC firm, recently invested in Augment AI, and lauded AI2 as a “powerhouse research institution.”

Redman, who is also on the board of the Institute for Systems Biology, said Seattle needs more indie research institutions, without the entrenched bureaucracies of universities. It’s an area where Seattle lags compared with Boston and San Francisco. Paul Allen doesn’t get enough credit for starting AI2, she said.

But like many other research institutions, she said, AI2 would benefit from more of its own researchers launching startups through the incubator. So far, most of the AI2 incubator startups have been launched by entrepreneurs brought in from outside. The ideal, she said, is when technologists are inspired to turn their own work into companies.

Redman said she hopes to see “more of those researchers becoming entrepreneurs themselves — having that fluidity between research and entrepreneurship, and really getting that flywheel going.”

Xnor.ai, the AI2 spinout acquired by Apple in 2020, is the best example thus far. Xnor was started by Ali Farhadi and Mohammad Rastergari, who were AI2 employees who wanted to commercialize their own research, before the incubator was formalized inside the institute. 

AI2’s leaders acknowledge this isn’t the norm, but they say this possibility is part of their pitch to AI2 research recruits who might be interested in turning their work into companies. There are currently two very early stage projects in the incubator led by researchers in the area of natural language processing, Colker said.

Etzioni, the AI2 CEO, has extensive experience as a AI researcher and startup founder, co-authoring more than 250 technical papers and co-founding companies including price comparison service NetBot (acquired by Excite in 1997); text mining and analytics company ClearForest (acquired by Reuters in 2007); travel price prediction company FareCast (acquired by Microsoft in 2008) and electronics price prediction startup Decide (acquired by EBay in 2013).

Some entrepreneurs who went through the AI2 Incubator earlier in its evolution expected more free-flowing brainstorming with AI2 researchers on potential startup ideas, based on their understanding when they were pitched on joining the incubator. But they soon came to realize that wouldn’t be the case.

Others, including Visnjic of WhyLabs and Oberoi of Lexion, said they knew going in that it wasn’t the researchers’ jobs to help them come up with startup ideas. Even so they were able to use the researchers for a reality check and feedback on the technical possibilities and limitations of their own startup concepts.

Colker said the AI2 Incubator experimented initially with open-ended explorations between entrepreneurs and researchers on startup ideas, but moved away from approach.

Instead, he said, the AI2 Incubator team now collaborates with entrepreneurs on a “product memo” that asks key questions to create a framework for their startups.

While Seattle has many incubators, including some with outstanding track records, the AI2 Incubator’s focus on AI is a distinguishing factor, said Ed Lazowska, a UW computer science professor and member of the AI2 board.

“Just as a consumer products startup would be crazy not to try to work with Maveron,” Lazowska said, referring to the consumer-focused VC firm, “an AI-intensive startup would be crazy not to try to work with AI2.”

On a related note: One of the entrepreneurs at AI2 who is working on a consumer-oriented AI startup, Varan Puri of AI-powered speech coach Yoodli, said if he could change one thing, it would be having more consumer AI startups to compare notes and commiserate with in the enterprise-heavy incubator.

Even so, he said, the best thing about the incubator is the existing community of entrepreneurs and mentors.

A contrarian AI approach

Underlying all of this is a philosophical and technical bet. At a time when OpenAI and its popular GPT-3 language model are spotlighting the potential of massive neural nets and transformers, AI2 is going in a different direction.

“AI2 would say it’s unlikely that scale alone will lead to general artificial intelligence; that there’s some other things that we’re not quite understanding yet about common-sense reasoning,” said computer scientist Peter Lee, Microsoft’s corporate vice president for research and incubations, and a member of the AI2 board.

Whereas giant neural networks — artificial systems modeled after the structure of the human brain — can help a self-driving car recognize when there’s a barrier in the road, for example, it’s also important for AI systems to have the common sense to deal with “fuzzier” situations like a construction zone or a street fair, he said.

Lee also works regularly with OpenAI as a result of Microsoft’s $1 billion investment and partnership with the San Francisco-based company, first announced nearly three years ago.

“It’s one of the fundamental questions in AI today: what is essential in the setup of a computing system to achieve this emergent concept of general intelligence?” Lee said. “You could almost put AI2 and OpenAI on opposite ends of a philosophical spectrum. … It’s super important to be doing both.”

“Honestly, I think Oren is out to prove a point about what’s really valuable,” Lee said. “He really believes that focusing on big classifiers, big recommendation systems, big neural nets, isn’t going to be the only place where there’s high impact and high value in the commercial space or in society.”

For AI2, he said, “the incubator has been an important place to try to prove that out.”

RELATED CONTENT

GeekWire’s coverage of the 14 active or acquired AI2 Incubator startups.

• Augment AI: Stealth startup focused on AI for productivity.
• BirchAI: Streamlining customer support for healthcare companies.
• BlueCanoe: Language learning technology.
• KITT (Baidu): Conversational language technology.
• Lexion: AI analysis of business contracts.
• MajorBoost: Streamlined communication between providers and insurers.
• Measure Labs: Remote care technology for healthcare.
• Modulus: AI-engineered immune cells.
• Ozette: Immune monitoring platform.
• Panda: AI customer relationship management.
• WellSaid: AI-generated synthetic voices to work.
• WhyLabs: Monitoring of machine learning models.
• Xnor (Apple): Low-power AI in edge devices.
• Yoodli: AI-powered speech coach.

Read More