TL;DR:
- Flower, a startup co-founded by Daniel Beutel from the University of Cambridge, addresses AI’s limitations due to reliance on public data.
- Flower’s platform employs federated learning, allowing AI model training on distributed data without direct data access.
- FedGPT, a new offering from Flower, enables large language model training on diverse data sources, maintaining data privacy.
- Partnership with Brave’s Dandelion project aims to create an open-source federated learning system across millions of Brave browser clients.
- Flower gains traction with 2,300+ developers and Fortune 500 companies like Porsche, Bosch and Nokia, along with esteemed institutions.
- Backed by investors, including First Spark Ventures and Hugging Face CEO Clem Delangue, Flower secures $3.6M in pre-seed funding.
Main AI News:
The realm of AI research finds itself constrained by its dependence on public data sources, predominantly drawn from the web. This impediment, as asserted by Daniel Beutel, a luminary tech entrepreneur and researcher hailing from the venerable University of Cambridge, is the crux of a burgeoning challenge in the AI domain. Beutel, who stands as a co-founder of the startup Flower, has embarked on a mission to alleviate this obstacle and redefine the landscape of AI research.
Speaking in an exclusive email interview with TechCrunch, Beutel elucidates, “The realm of publicly accessible centralized data represents a mere fraction of the vast data reservoirs that saturate our world.” This is in stark contrast to the far-reaching potential of distributed data—the troves that reside within devices such as smartphones, wearables, IoT devices, and corporate silos. Alas, these invaluable sources of information remain beyond the grasp of contemporary AI, thus prompting the establishment of Flower.
Conceived in 2020 by Beutel and his esteemed Cambridge compatriots Taner Topal and Nicholas Lane, a former head honcho at Samsung’s AI Center in Cambridge, Flower endeavors to revolutionize the AI training process. The platform orchestrates an ingenious decentralization of AI training by enabling developers to sculpt models using data emanating from myriad devices and locations. The underpinning technique, dubbed federated learning, grants developers a method to train models without direct data access, thereby fostering privacy and compliance in precarious scenarios.
Beutel envisions a future where this approach transcends its current status to become the crux of AI training, asserting, “Flower envisions a scenario where the advantages of distributed data render this paradigm the prevailing norm.”
While the concept of federated learning isn’t novel, its implementation spans academia and practical applications. Flower, however, leverages this technique to maintain data integrity. “In the Flower ethos, data remains ensconced within its source—whether that’s a device or a corporate enclave—throughout the training process,” explicates Beutel. The platform ushers in a transformative reality where computations gravitate towards data, eradicating the need to relocate the data. Instead, partial training occurs in various locations, with only the training outcomes exchanged, amalgamating into the grand tapestry of results.
As a testament to its innovation, Flower has recently unveiled FedGPT—a groundbreaking approach to training large language models (LLMs), akin to OpenAI’s heralded ChatGPT and GPT-4. This avant-garde creation is currently in preview, granting organizations the power to construct LLMs fortified by internal, sensitive data, sans the obligation to share them with third-party providers.
A strategic liaison with Brave, the open-source browser heavyweight, has spawned the Dandelion initiative. This audacious endeavor aims to erect a federated learning system spanning Brave’s extensive user base, a staggering 50 million strong. This, Beutel asserts, is part of the larger picture where AI navigates a future fraught with regulatory scrutiny, warranting vigilant data management.
In a span of mere months, Flower’s growth trajectory has been meteoric. Boasting a developer community exceeding 2,300, Beutel highlights the inclusion of “dozens” of Fortune 500 titans and esteemed academic institutions among Flower’s adopters. The likes of Porsche, Bosch, Samsung, Nokia, and a constellation of prestigious institutions, including Stanford, Oxford, MIT, and Harvard bear testimony to the platform’s allure.
Fuelled by this triumph, Flower, an esteemed member of Y Combinator’s 2023 cohort, has garnered the favor of investors. Noteworthy backers include First Spark Ventures, Hugging Face CEO Clem Delangue, Factorial Capital, Betaworks, and Pioneer Fund, culminating in a pre-seed investment of $3.6 million.
Beutel envisions a vibrant future where collective effort ushers in a pantheon of open-source federated techniques, accessible to the global community through Flower’s all-encompassing framework.
Conclusion:
The emergence of Flower signifies a transformative shift in AI training methodologies, harnessing the power of federated learning to overcome data limitations. This pioneering approach not only elevates privacy and compliance standards but also empowers businesses with the potential to leverage distributed data for advanced AI models. The collaborations, partnerships, and investments underscore Flower’s strategic positioning in reshaping the AI landscape, a market that demands innovative solutions to maximize data-driven insights while upholding regulatory concerns.