A new class of reasoning-focused large language models are arriving—and VC investors see a new battlefield opening within AI, one where AI agents and reliable models will flourish.
These new models are different from Anthropic‘s Claude or OpenAI‘s GPT. Reasoning models are built upon more complex architecture and algorithms—often with wholly different parameters and datasets—to handle difficult computations and tasks. The models mimic how humans think and process information—how they reason—albeit faster and with greater accuracy.
Venture capitalists are bullish that investments in these models will be lucrative, especially now that the segment has been thrust into the spotlight with the release of OpenAI’s o1-preview model, code-named “Strawberry.”
“There’s this whole new room for competition,” said Chris Kauffman, a principal at General Catalyst whose portfolio includes Symbolica, a mathematics-based LLM developer. “Reasoning opens the door to more creative mathematical approaches. ... OpenAI’s o1 is a super-important formal step in the AI community’s approach to reasoning—but it is far from the only step.”
A new crop of startups have emerged to tackle reasoning. Fei-Fei Li, the researcher known as the “AI godmother,” launched World Labs with $230 million in funding in September to develop spatial reasoning technology. Harmonic, a startup co-founded by Robinhood CEO Vlad Tenev that is developing models to solve complex math equations, just raised a $75 million Series A led by Sequoia Capital.
The broader AI and machine learning space continues to grow, according to PitchBook’s Q2 2024 Gaming Report. Deal value reached $30.8 billion, up from $23.2 billion quarter-over-quarter.
Reasoning models could further boost AI investment because of the lucrative opportunities they create.
“It’s going to unlock new capabilities that we started to see signs of but weren’t fully realized,” Kauffman said.
Agent activate
One of the use cases Kauffman refers to are AI agents, which can carry out multistep services from a single request. Examples include travel agents capable of booking and scheduling entire vacations and personalized shoppers that can order gifts within defined parameters. The usefulness of what emerges from reasoning-based AI depends on users’ confidence in the quality and accuracy of the output.
Startups in this segment include Otto, an AI travel agent that raised a $6 million seed round led by Madrona in August and Sierra, a conversational shopping and customer service specialist founded by Bret Taylor, former Salesforce co-CEO and OpenAI board member. The Information reported in January that Sierra was raising an $85 million round that would value the startup at $1 billion.
These startups have been hamstrung by earlier large language models.
Models like GPT are built on a large series of networks fed huge amounts of data that process requests in sequences called Transformers. These models are good at understanding context because they process multiple sequences at the same time. But they’re less reactive and fall short at real-time decisionmaking that is critical to tasks agents must perform. They also can’t remember details and preferences over time.
“We saw a bunch of wonderful demos, but I haven’t seen any come out that are in the hands of users, being used by millions of people without any major issues,” Kauffman said. “Reasoning is a critical intermediate step to really achieve useful products.”
Opening the toolbox
User confidence in quality and accuracy is where AI has struggled to get a foothold, particularly with large enterprise clients who are enthusiastic about deploying AI but wary of quality. Industries like biomedicine, applied materials research and aerospace engineering are best suited for these more complex models, said former machine learning engineer Jonathan Choi, a partner at Buckley Ventures, which is another of Symbolica’s investors.
He said that reasoning-based models can help solve AI’s reliability problem. These models can be trained to adhere to complex mathematical concepts where proving solutions to validate accuracy is the standard. Applying these methods might help large language models get a foothold in industries with strict regulations.
“We’re not exploring the broader toolset. We’re using the same tool and trying to use it for different cases,” he said. “Not all AI models are the same thing—having more of a math foundation allows for more control over actual performance.”
The landing zone
As this next phase of AI development gets underway, investors are still cautioning startups on where to build.
Battery Ventures principal Brandon Gleklen, whose portfolio includes the AI agent startup Lindy, has a warning for startups attempting to ride on OpenAI’s coattails: “Any company who succeeds or fails based on the next OpenAI launch—I think you’re in the wrong business.”
Gleklen said that startups focused on improving reasoning models through niche fixes are setting themselves up for failure.
“People who are too focused on the current limitations of models and doing small technical solves for them, I think that’s a tough place to land.”
Update: This article was updated to reflect Harmonic’s Series A. (Sept. 23, 2024)
Featured image by Daniel Grizelj/Getty Images
