Imbue secures $200M to build AI models that can reason
A research lab previously known as Generally Intelligent, Imbue, has recently secured $200 million in a Series B funding round. The company is now valued at over $1 billion, and notable investors include the Astera Institute, Cruise CEO Kyle Vogt, Nvidia, and Notion co-founder Simon Last. With this new funding, Imbue has recently raised $220 million, placing them among the top-funded AI startups. AI21 Labs currently holds the top spot with $283 million, followed by Cohere with $435 million and Adept with $415 million.
In a blog post published this morning, Imbue announced that their latest funding will speed up the development of AI systems capable of reasoning and coding. The ultimate aim is to create practical AI agents that can safely and effectively assist us in achieving more significant goals in the real world. Despite the focus on innovation, the objective remains to build AI that can be trusted to work for us safely and practically.
Last October, Imbue emerged from stealth mode with an ambitious objective: to explore the essential elements of human intelligence that machines lack. As explained to TechCrunch then, the company’s approach was to convert these “fundamentals” into various tasks to be solved. Imbue then designed different AI models and tested their ability to learn and solve these tasks in complex 3D environments created by the Imbue team.
The company has changed its approach since then. Instead of using AI to explore 3D worlds, Imbue is now focused on developing internal models that are useful for various purposes, such as coding models similar to GitHub Copilot and Amazon CodeWhisperer. Imbue’s models stand out because they can reason robustly, according to the company’s claims.
According to a blog post by Imbue, effective AI agents are often hindered by a lack of robust reasoning. To be able to take effective action, it is essential to tackle uncertainty by being flexible and willing to change approaches when necessary. This involves asking questions, gathering new information, exploring different scenarios, and making decisions based on the available facts. To deal with the complicated and unpredictable nature of the real world, it is also crucial to be able to generate and discard hypotheses as needed. These abilities are vital for success in any field and help us navigate the complexities of the world around us. In order to improve the effectiveness of AI agents, it is crucial to focus on developing these reasoning skills.
According to Imbue, code has significant value beyond just facilitating the development of AI applications on a large scale. In their recent blog post, the company argues that code can enhance reasoning abilities and is a highly effective approach for models to perform tasks on a machine.
According to a recent statement by the company, an agent that writes a SQL query to extract data from a table is more likely to meet a user’s requirements than an agent that tries to assemble the same information without using any code. Furthermore, the company believes that models trained with code learn to reason better, while models trained without code tend to claim poorly. This philosophy is similar to Adept’s, which aims to create AI capable of automating any software process. Google DeepMind has also explored ways to train AI to control computers, such as observing people’s keyboard and mouse commands when completing tasks like booking a flight.
On the other hand, Imbue claims that its models are specifically designed for reasoning by training on data to reinforce good reasoning patterns. Imbue uses techniques that spend more computing time during inference to arrive at robust conclusions and actions. The company introduces “extensive” models with over 100 billion parameters optimized to perform well on its internal benchmarks for reasoning. Imbue conducts this training on a compute cluster co-designed by Nvidia, which contains 10,000 GPUs from the document_number_1 series.
Imbue is investing in creating its own AI and machine learning tools, such as prototypes for debugging and visual interfaces on top of AI models. The company is also researching to understand the learning process in large language models. However, Imbue plans to put only a little of its current work into production. Instead, it views these tools and models as a way to enhance future, more versatile AI and lay the foundation for a platform individuals can use to construct their customized models.
According to Imbue’s blog post, the company is creating computers that can comprehend our objectives, communicate proactively, and work in the background when developing AI agents. Ultimately, the company aims to release systems that allow anyone to construct robust, custom AI agents that empower individuals worldwide with the productive power of AI. Imbue believes this recent funding will expedite the development of AI systems that can reason and code, enabling individuals to achieve larger goals globally.