Why the newest $1 billion AI startup isn’t aiming to outperform OpenAI.

Fri Sep 8, 2023 - 3:11am GMT+0000

Imbue, formerly known as Generally Intelligent, has recently achieved a valuation of $1 billion and attained unicorn status following a funding round of $200 million. This round was led by Astera Institute, NVIDIA, Cruise CEO Kyle Vogt, Notion co-founder Simon Last, and other prominent investors. However, Imbue, an AI research lab specializing in creating custom AI agents with reasoning capabilities, doesn’t view itself as a direct competitor to companies like OpenAI and Anthropic, which focus on foundation models. Co-founder Josh Albrecht explained that they are more inclined towards fostering diversity within the AI ecosystem.

Imbue’s vision revolves around creating a diverse ecosystem where various companies offer different AI models to cater to various needs. It’s worth noting that Imbue stands out as one of the few woman-led AI unicorns. Co-founder Kanjun Qiu expressed excitement about being at the forefront of a groundbreaking era where computers possess intelligence, emphasizing their commitment to making this intelligence accessible to everyone.

Imbue’s primary focus is on developing large language models optimized for reasoning. According to Qiu, they believe that reasoning is a fundamental hurdle to creating highly functional AI agents. Effective reasoning enables these agents to handle uncertainty, adapt, acquire new information, make decisions, and navigate the complexities of the real world.

Imbue employs a comprehensive approach to develop reasoning models, involving training foundation models, creating experimental agents and interfaces, building robust tools and infrastructure, and researching the theoretical underpinnings of deep learning.

To achieve its goals, Imbue trains its own extremely large models, boasting over 100 billion parameters. This enables them to quickly iterate on training data, architecture, and reasoning mechanisms. Their training process is powered by a formidable Nvidia H100 GPU cluster comprising around 10,000 units.

Their ultimate objective is to develop AI agents capable of assisting in various tasks, such as coding and policy analysis. However, Imbue acknowledges that current AI agents for these tasks often fall short of expectations.

While their long-term goal is to enable anyone to create their own AI agents, they initially focus on developing reasoning models for internal enterprise applications, with coding agents being a particular emphasis. Imbue believes that coding agents can significantly enhance research and engineering capabilities, setting them apart from models like ChatGPT by emphasizing explainability and transparency.

Imbue’s commitment to explainability extends to building trust with users. They want their AI tools to be user-friendly and understandable.

Imbue is also dedicated to advancing scientific understanding of various models and approaches to training neural networks, with a focus on understanding the inner workings of deep learning.

Regarding their business model, Imbue is open to different possibilities depending on what capabilities of their models are easy or challenging to build upon. They may offer applications directly to end-users or enable others to build on top of their models.

Imbue envisions an ecosystem where individuals have the freedom to customize their own AI tools, rather than engaging in direct competition with other AI labs. Their goal is to democratize AI and put users in control.

In their first year since emerging from stealth mode, Imbue has made significant progress in experimenting with internal agents based on their models. Their ongoing mission is to develop reliable and “street-smart” AI systems that can assist users effectively.

Despite questions about high valuations in the AI sector, Imbue remains focused on foundational research, with the belief that reasoned and trustworthy AI tools have the potential to unlock vast possibilities when widely accessible.