The paths of Ricursive Intelligence’s co-founders, Anna Goldie and Azalia Mirhoseini, appear to have been fated to converge.
Anna Goldie, CEO, and Azalia Mirhoseini, CTO, hold such esteemed reputations within the AI sector that they were among the select group of AI engineers who “received those unusual emails from Zuckerberg with extravagant proposals,” Goldie humorously recounted to TechCrunch. (They declined these offers.) Both individuals previously collaborated at Google Brain and were among the initial team members at Anthropic.
Their notable achievements at Google included the development of the Alpha Chip, an artificial intelligence tool capable of generating robust chip layouts within hours, a task that typically demands over a year for human designers. This innovative tool played a crucial role in the design of three iterations of Google’s Tensor Processing Units.
This impressive background clarifies why, only four months post-launch, Ricursive recently secured a $300 million Series A funding round, valuing the company at $4 billion, spearheaded by Lightspeed. This achievement followed closely on the heels of a $35 million seed round led by Sequoia a few months prior.
Instead of manufacturing chips, Ricursive develops AI-powered tools for chip design. This distinction sets them apart from most other AI chip startups, positioning them not as an aspiring rival to Nvidia, but rather as a collaborator; indeed, Nvidia itself is an investor. Major chip manufacturers such as Nvidia, AMD, and Intel, among others, represent the startup’s primary clientele.
“Our aim is to facilitate the automated and significantly accelerated creation of all types of chips, whether custom or conventional. We leverage AI to achieve this,” Mirhoseini informed TechCrunch.
Their initial encounter occurred at Stanford, where Goldie was pursuing her PhD while Mirhoseini instructed computer science courses. Subsequently, their professional trajectories have mirrored each other precisely. Goldie elaborated, “We commenced our roles at Google Brain simultaneously. We departed Google Brain simultaneously. We joined Anthropic simultaneously. We exited Anthropic simultaneously. We returned to Google simultaneously, and then departed Google once more on the same day. Finally, we launched this company together on the same day.”
While at Google, the duo maintained a close bond, even participating in circuit training workouts together. Jeff Dean, the renowned Google engineer and their collaborator, recognized the clever wordplay, coining their Alpha Chip initiative “chip circuit training” in reference to their fitness regimen. Internally, the pair was also affectionately known as A&A.
While the Alpha Chip garnered significant industry attention for them, it also sparked contention. According to Wired, a Google colleague was dismissed in 2022 following years of attempts to undermine A&A and their chip design contributions, despite this work being integral to some of Google’s most critical, strategically important AI chips.
The Alpha Chip project, developed at Google Brain, served as conclusive evidence for the core concept behind Ricursive: leveraging AI to vastly expedite the process of chip design.
The Complexities of Chip Design
The challenge lies in the fact that computer chips incorporate millions, often billions, of logic gate components onto their silicon wafers. Human designers typically dedicate a year or more to meticulously positioning these components to guarantee optimal performance, efficient power usage, and other specific design criteria. Precisely orchestrating the digital placement of such minuscule elements is, as one might anticipate, exceedingly difficult.
The Alpha Chip was capable of “producing an exceptionally high-quality layout in approximately six hours. What made this method particularly remarkable was its ability to learn through experience,” Goldie stated.
The fundamental principle of their AI chip design methodology involves employing “a reward signal” to evaluate the quality of a design. Subsequently, the AI agent utilizes this assessment to “adjust the parameters of its deep neural network, thereby improving its performance,” Goldie explained. The founders noted that after processing thousands of designs, the agent achieved remarkable proficiency and increased its speed as it acquired more knowledge.
Ricursive’s forthcoming platform will advance this concept significantly. The AI chip designer under development “will learn across a variety of chips,” Goldie noted, implying that each design iteration will contribute to its enhanced proficiency for subsequent chip projects.
The Ricursive platform additionally integrates Large Language Models (LLMs) and will manage the entire design pipeline, from initial component placement to final design verification. Their target market encompasses any enterprise involved in electronics manufacturing that requires custom or specialized chips.
Should their platform successfully demonstrate its capabilities, as appears probable, Ricursive could contribute to the ambitious objective of attaining artificial general intelligence (AGI). Ultimately, their profound aspiration involves designing AI chips, which would entail AI itself conceiving its own computational architectures.
“Chips represent the vital energy source for AI,” Goldie declared. “I believe that by engineering more potent chips, we pave the most effective path forward for this domain.”
Mirhoseini further commented that the protracted nature of the chip design process currently impedes the pace of AI advancement. “We believe we can foster a rapid co-evolution between AI models and the chips that underpin them,” she articulated, thereby allowing AI to progress and become intelligent more swiftly.
While the notion of AI autonomously designing its own computational capabilities at accelerating rates might evoke images of Skynet or the Terminator, the founders emphasize a more constructive, immediate, and, in their view, probable advantage: enhanced hardware efficiency.
Once AI laboratories are capable of designing significantly more efficient chips (and eventually all foundational hardware), their expansion will not necessitate the depletion of such vast quantities of global resources.
“We possess the capability to engineer a computer architecture specifically optimized for a given model, potentially yielding an almost tenfold enhancement in performance relative to the total cost of ownership,” Goldie explained.
Although the nascent startup refrains from disclosing its initial clientele, the founders affirm that they have been approached by virtually every prominent name in chip manufacturing. Predictably, they have the luxury of choosing their inaugural development collaborators.