Nearly any generative AI can write an essay at this point.
Input the topic, a few key points and the word count — a moment later, the AI tool will spit out an essay. Granted it might not be the most eloquent or well-researched piece of writing, but that is where the human behind the keyboard comes into play.
The process of evaluation — where the person figures out what questions to ask next or what sections to cut and rearrange — is part of reasoning that, right now, generative AI can’t do.
That won’t be the case for long if Imbue, an AI research company formerly known as Generally Intelligent, keeps growing the way it has over the last year. Just last month, Imbue brought in $200 million in series B funding.
“Our growth has exceeded my wildest expectations,” said Technical Staff Member Bas van Opheusden. “Since I joined the company around 6 months ago, we doubled the size of the engineering team, we have hired several recruiting and operations staff and we have raised $200 million at a $1 billion valuation.”
Opheusden added that Imbue has made improvements on the technical side, as well: the team has gathered pretraining data, trained several large language models and developed abstractions for robust reasoning agents.
“Although I have been confident in our mission from the start, what surprised me the most is the speed at which we have been able to go from idea to execution,” he noted. “This is especially striking for me personally, as I have a background in academia, where usually not much changes over any six-month period — and certainly not the exponential growth I’ve seen at Imbue.”
“Although I have been confident in our mission from the start, what surprised me the most is the speed at which we have been able to go from idea to execution.”
The company is building a future where AI agents have reasoning capabilities and can safely work in the real world.
“We believe reasoning is the primary blocker to effective AI agents,” Imbue wrote in a September 2023 blog post. “Robust reasoning is necessary for effective action. It involves the ability to deal with uncertainty, to know when to change our approach, to ask questions and gather new information, to play out scenarios and make decisions, to make and discard hypotheses and generally to deal with the complicated, hard-to-predict nature of the real world.”
An introduction by CEO Kanjun Qiu on the company’s website goes into detail explaining how the company hopes to essentially change how personal computers are used. She envisions a future where technology will be able to do more work for humans.
Some examples of how personal computers could take on complex tasks from humans might include planning a family vacation or making a slide presentation. Today, AI models can generate text or images after receiving a prompt. However, the process of working on complex tasks is still lacking in today’s systems. Where Imbue hopes to step in is to improve the reasoning capabilities of AI agents to make them deal with complex tasks autonomously.
Imbue gives the example of a personal computer taking on complex tasks like planning a family vacation or making a slide presentation. Right now, many generative AI models can generate something — an essay, an image, a button — but the process of looking at what it’s generated and reasoning through what to do next with it, that is where Imbue hopes to step in.
The treatise of Imbue thought leaders is to let AI take on tech-related tasks and free people up to take on more human pursuits, like art, creation and idea generation. The technical team at Imbue takes a very thoughtful approach in ensuring the safety of the AI agents, though. The company conducts research on the state of ethical decision-making by large language models — studying things like the process of failure and the assessment that humans go through when they fail.
Pushing themselves to think bigger and expand into new ideas is a key part of the culture.
“At Imbue, we are deeply committed to personal growth and skill acquisition,” said Opheusden. “This is part of our culture and we cherish it. We have a number of tools to facilitate learning and to encourage our team members to seek out opportunities for personal growth.”
Opheusden shared the example that every week, many of the technical teams jointly read and discuss recent studies. “We’ve even organized reading groups where we jointly read through a textbook on topics like software engineering or mathematics,” added Opheusden.
Jamie Simon, a research fellow, also noted that the culture at Imbue revolves around encouragement and support for “self-betterment of all kinds.” Simon explained how impactful it is when there is a continuous push for evolution and education.
“People are constantly learning new things — hardware things, software things and machine learning things,” said Simon. “People are encouraged to share their learnings with others. We function very well as a collective brain.”
“People are encouraged to share their learnings with others. We function very well as a collective brain.”
Back when the company first started, one of its foundational elements was a focus on conducting and sharing new research — all while developing the company and tech.
“Our growth has been fast but exceptionally targeted,” said Head of AI Safety and Policy Matt Boulos. “We have been able to find people who contribute meaningfully to both our culture and to the domains that align with their passions. Our team has a remarkable culture that encourages autonomy and self-growth, and it has been inspiring to see it actually get better as the team has gotten larger.”
Boulos added that the company’s growth is lining up perfectly today with where the team hoped it would several years ago. “It’s exactly on track,” said Boulos. “We’ve almost doubled since I joined, and that was our target.”
Opheusden added that because of the nature of Imbue’s work, it’s even more important to keep everyone in the company hungry to learn.
“As transformers and large language models are becoming increasingly capable, it is obvious that the economic impact of AI will be tremendous, and an entire industry is developing to develop, implement, train and run these models,” noted Opheusden. “However, language models are often unreliable, and the biggest opportunity is for someone to build agents that can leverage the capabilities and integrate them in a robust system that actually works and is useful.”
The sheer size of these models require massive computing hardware and many engineers, Opheusden pointed out.
“Language models are often unreliable, and the biggest opportunity is for someone to build agents that can leverage the capabilities and integrate them in a robust system that actually works and is useful.”
“This is why now is the moment to maximize the opportunity and build robust agents that are useful to the real world,” Opheusden added. We believe we have the necessary infrastructure, tools and ideas to take on this challenge — but it’s going to take a lot of hard work.”
Boulos added that the team at Imbue treats mutual growth, personally and as a company, as its number one priority.
“No one is ever on their own, and as both our personal and team needs evolve, we step up to support that,” concluded Boulos. “The opportunity in front of us is tremendous and we have the progress and resources to actually realize it — this is an exciting moment for us.”