Imbue, formerly known as Generally Intelligent, may now be enjoying a $1 billion valuation — and unicorn status — after this week’s funding round of $200 million led by Astera Institute, NVIDIA, Cruise CEO Kyle Vogt, Notion co-founder Simon Last, and other investors. But the AI research lab, which focuses on building custom, reasoning AI agents, doesn’t see itself in direct competition with OpenAI, Anthropic and other foundation model companies. Cofounder Josh Albrecht told VentureBeat that “we’re quite bullish on the diversity and to be more of an ecosystem.”
And the idea of a diverse ecosystem where different companies provide different models for different needs is exciting for Imbue, one of the very few woman-led AI unicorns, added the company’s other cofounder Kanjun Qiu.
“It feels like we’re at the very beginning of something huge,” she said. “This is the first time computers have had intelligence. That’s so crazy. So what we’re really excited about is like, how can we make that accessible to everyone so that everyone can imbue intelligence and be able to use that intelligence.”
That speaks to Imbue’s M.O., which is developing large language models (LLMs) optimized for reasoning abilities. “We build foundation models, large foundation models optimized for reasoning,” said Qiu. “We believe, essentially, that reasoning is the core blocker to agents that work really well.”
The company believes that the ability to reason is crucial for effective AI agents. Robust reasoning allows agents to handle uncertainty, adapt their approaches, gather new information, make decisions, and deal with the complexities of the real world.
Imbue adopts a “full stack” approach to develop reasoning models, which involves training foundation models, prototyping experimental agents and interfaces, building robust tools and infrastructure, and studying the theoretical foundations of deep learning.
Imbue trains its own very large models optimized for reasoning, with over 100 billion parameters. This allows them to iterate rapidly on training data, architecture, and reasoning mechanisms. Powering Imbue’s “wide” approach to training datasets is a ~10,000 strong Nvidia H100 GPU cluster.
The goal is to train models with strong reasoning capabilities in order to build “AI agents that are able to help us accomplish bigger goals in the world,” said Qiu.
Currently, Imbue focuses on building agents for everyday tasks like writing code or analyzing policy submissions.
As Albrecht explained, “We have some of those and but for the most part, they’re really bad. And, you know, I think you’ll see that when you look at where agents are today.”
While Imbue’s ultimate objective is to enable anyone to build their own AI agents, the company’s initial focus is on developing reasoning models for internal enterprise uses. Imbue specifically concentrates on agents that can code, as coding improves reasoning and provides a practical test-bed for evaluating the effectiveness of their models. According to a blog post, Imbue believes that coding agents are strategically important and can significantly enhance research and engineering capabilities.
A key difference from models like ChatGPT, according to Qiu, is Imbue’s focus on “making models explain their reasoning and give some references.” This “unpacking the black box,or explainability, improves transparency and trust over models that just output answers.
As AI systems grow more powerful, understanding what’s happening inside the “black box” becomes increasingly important. Hence why Imbue is working to shed light on model reasoning.
“Making it not a black box is a good user experience,” said Qiu. Imbue’s goal is developing AI where “you can check [it]. It’s just a better experience when you can really understand what’s going on.”
Albrecht noted this focus on explainability is key to building trust: “We want these things just to be software tools that you use. It just does what you expect.”
Imbue is also advancing scientific understanding of different models and approaches to training neural networks. As Albrecht explained, “We want to understand what’s happening inside those weights, what’s happening inside those black boxes inside of deep learning.”
When asked about Imbue’s business model, Qiu and Albrecht signaled an open approach. Qiu noted, “it’s going to really depend on what capabilities of the model are easy and hard to build with.”
If building applications directly on their models proves difficult, “it might make more sense for us to go direct to the end user and build agents that they can use,” Qiu explained. However, “If it’s the case that it’s actually much easier to build on top, then we might enable other people to build on top of them.”
Albrecht agreed Imbue’s focus is on “tools for empowering other people to build on top of” the company’s work.. By making its reasoning-focused models accessible, Imbue aims to serve both businesses and individuals. As Qiu said, “probably both will happen just like Apple in their App Store.”
“It really harkens back to what was the dream of the personal computer, the whole idea is that it was personal,” said Qiu. “Hopefully in the future, I have all of my own custom software, and you have all of your own custom software. It’s all different.”
Rather than direct competition with other AI labs, Imbue envisions “an ecosystem where different people can actually really change their own computer to what they want it to be,” Albrecht said, “We very much want this to be democratized, individually driven. The user is the one who’s in control.”
In its first year since coming out of stealth, Imbue has made progress experimenting with internal agents built on its models. “Now we’re training models and we’re experimenting with agents internally and that’s where we are going to keep going on that and continue pushing the models on reasoning,” said Qiu. Their goal remains developing systems that can reliably help users through an emphasis on robust and “street smart” AI.
While some question high valuations in the sector, Imbue remains focused on the foundational work. As Albrecht noted, “We’re going to keep doing the same thing that we’ve been doing and just trying to understand how these things work and build things that are useful.” The vision is that reasoned, trustworthy AI tools will unlock vast potential when widely accessible.
Head over to our on-demand library to view sessions from VB Transform 2023. Register Here
Imbue, formerly known as Generally Intelligent, may now be enjoying a $1 billion valuation — and unicorn status — after this week’s funding round of $200 million led by Astera Institute, NVIDIA, Cruise CEO Kyle Vogt, Notion co-founder Simon Last, and other investors. But the AI research lab, which focuses on building custom, reasoning AI agents, doesn’t see itself in direct competition with OpenAI, Anthropic and other foundation model companies. Cofounder Josh Albrecht told VentureBeat that “we’re quite bullish on the diversity and to be more of an ecosystem.”
And the idea of a diverse ecosystem where different companies provide different models for different needs is exciting for Imbue, one of the very few woman-led AI unicorns, added the company’s other cofounder Kanjun Qiu.
“It feels like we’re at the very beginning of something huge,” she said. “This is the first time computers have had intelligence. That’s so crazy. So what we’re really excited about is like, how can we make that accessible to everyone so that everyone can imbue intelligence and be able to use that intelligence.”
That speaks to Imbue’s M.O., which is developing large language models (LLMs) optimized for reasoning abilities. “We build foundation models, large foundation models optimized for reasoning,” said Qiu. “We believe, essentially, that reasoning is the core blocker to agents that work really well.”
Event
VB Transform 2023 On-Demand
Did you miss a session from VB Transform 2023? Register to access the on-demand library for all of our featured sessions.
The company believes that the ability to reason is crucial for effective AI agents. Robust reasoning allows agents to handle uncertainty, adapt their approaches, gather new information, make decisions, and deal with the complexities of the real world.
Imbue adopts a “full stack” approach to develop reasoning models, which involves training foundation models, prototyping experimental agents and interfaces, building robust tools and infrastructure, and studying the theoretical foundations of deep learning.
AI agents with reasoning can open the black box
Imbue trains its own very large models optimized for reasoning, with over 100 billion parameters. This allows them to iterate rapidly on training data, architecture, and reasoning mechanisms. Powering Imbue’s “wide” approach to training datasets is a ~10,000 strong Nvidia H100 GPU cluster.
The goal is to train models with strong reasoning capabilities in order to build “AI agents that are able to help us accomplish bigger goals in the world,” said Qiu.
Currently, Imbue focuses on building agents for everyday tasks like writing code or analyzing policy submissions.
As Albrecht explained, “We have some of those and but for the most part, they’re really bad. And, you know, I think you’ll see that when you look at where agents are today.”
While Imbue’s ultimate objective is to enable anyone to build their own AI agents, the company’s initial focus is on developing reasoning models for internal enterprise uses. Imbue specifically concentrates on agents that can code, as coding improves reasoning and provides a practical test-bed for evaluating the effectiveness of their models. According to a blog post, Imbue believes that coding agents are strategically important and can significantly enhance research and engineering capabilities.
A key difference from models like ChatGPT, according to Qiu, is Imbue’s focus on “making models explain their reasoning and give some references.” This “unpacking the black box,or explainability, improves transparency and trust over models that just output answers.
As AI systems grow more powerful, understanding what’s happening inside the “black box” becomes increasingly important. Hence why Imbue is working to shed light on model reasoning.
“Making it not a black box is a good user experience,” said Qiu. Imbue’s goal is developing AI where “you can check [it]. It’s just a better experience when you can really understand what’s going on.”
Albrecht noted this focus on explainability is key to building trust: “We want these things just to be software tools that you use. It just does what you expect.”
Imbue is also advancing scientific understanding of different models and approaches to training neural networks. As Albrecht explained, “We want to understand what’s happening inside those weights, what’s happening inside those black boxes inside of deep learning.”
AI ecosystem primed for a PC-like revolution
When asked about Imbue’s business model, Qiu and Albrecht signaled an open approach. Qiu noted, “it’s going to really depend on what capabilities of the model are easy and hard to build with.”
If building applications directly on their models proves difficult, “it might make more sense for us to go direct to the end user and build agents that they can use,” Qiu explained. However, “If it’s the case that it’s actually much easier to build on top, then we might enable other people to build on top of them.”
Albrecht agreed Imbue’s focus is on “tools for empowering other people to build on top of” the company’s work.. By making its reasoning-focused models accessible, Imbue aims to serve both businesses and individuals. As Qiu said, “probably both will happen just like Apple in their App Store.”
“It really harkens back to what was the dream of the personal computer, the whole idea is that it was personal,” said Qiu. “Hopefully in the future, I have all of my own custom software, and you have all of your own custom software. It’s all different.”
Rather than direct competition with other AI labs, Imbue envisions “an ecosystem where different people can actually really change their own computer to what they want it to be,” Albrecht said, “We very much want this to be democratized, individually driven. The user is the one who’s in control.”
In its first year since coming out of stealth, Imbue has made progress experimenting with internal agents built on its models. “Now we’re training models and we’re experimenting with agents internally and that’s where we are going to keep going on that and continue pushing the models on reasoning,” said Qiu. Their goal remains developing systems that can reliably help users through an emphasis on robust and “street smart” AI.
While some question high valuations in the sector, Imbue remains focused on the foundational work. As Albrecht noted, “We’re going to keep doing the same thing that we’ve been doing and just trying to understand how these things work and build things that are useful.” The vision is that reasoned, trustworthy AI tools will unlock vast potential when widely accessible.
VentureBeat’s mission is to be a digital town square for technical decision-makers to gain knowledge about transformative enterprise technology and transact. Discover our Briefings.
Author: Bryson Masse
Source: Venturebeat
Reviewed By: Editorial Team
