Agemo.
- Agemo has exited stealth with $4 million to build AI that turns text prompts into software.
- Founders Aymeric Zhuo and Osman Ramadan want to tackle the problem of LLMs plateauing.
- The startup bills itself as Europe's answer to incumbents such as Poolside and Magic.
When Aymeric Zhuo and Osman Ramadan met while working at a startup in 2022, they realized they had something in common: they both see themselves as underdogs.
Now, the pair are taking on incumbent AI companies that have raised hundreds of millions of dollars with their new venture, Agemo, which has just exited from stealth.
Agemo, which raised a $4 million seed round earlier this year, has built a platform that aims to allow non-developers to autonomously develop software from text prompts.
Its funding was led by Firstminute Capital, backers of Mistral and Wayve, and Fly Ventures, along with capital from DeepMind executive Mehdi Ghissassi and Datadog CEO Olivier Pomel.
The founders met while working at Shift Lab, a startup, and realized they had an instantaneous synergy.
"When we look at AI researchers, they're good at researching but not building solutions. We're both very technical but in a complementary way," Zhuo told Business Insider.
Zhuo, the startup's CEO, and Ramadan, the CTO, started working together on Agemo while still at Shift Lab. The pair say they even turned down an opportunity to work at OpenAI in 2023 in favor of building their upstart.
A key problem that the duo wanted to solve was programming large language models to orchestrate software. Essentially, this required them to train their AI models to reason like a team of engineers.
Agemo finds itself up against the likes of Poolside, which raised $500 million in October, and Magic, which raised $320 million in August.
But the pair are no strangers to adversity.
Ramadan grew up in Sudan during a time of civil unrest. As a teenager, he pored over his brother's computer science textbooks and taught himself how to code, eventually securing a place at Cambridge University to study engineering. After graduating, he completed a nearly four-year stint at Microsoft as a research software engineer.
Zhuo was the first in his family to graduate from high school and made his way to France's most prestigious engineering school — École Polytechnique — before securing roles at Activision and TikTok.
Agemo is tapping into AI reasoning
Zhuo told BI that one key issue in the AI ecosystem is that large language models are "hitting a plateau in incremental improvements."
This means that companies might be pouring a lot of money and resources into training these AI models but with proportionally less to show for it in terms of output.
To combat this problem, Agemo is building AI systems that can "reason" in software. It requires AI to make inferences from a dataset and generalize it to a set of problems.
Typically, an LLM would solve a problem by combing through its training data instead of learning the addition skill, Zhuo told BI.
"The implications are that our system can solve problems it had not been trained on before, whereas existing LLMs such as GPT-4 can't," Zhuo added.
Agemo.
For example, instead of ChatGPT providing users with steps on how to integrate a machine with Snowflake, the Agemo founders say its AI system can generate the tool to do it, based on a user's high-level intent.
Europe's answer to Poolside and Magic
To differentiate it from competitors such as Poolside and Magic, the startup says it has developed a neurosymbolic AI system for software reasoning. Neurosymbolic AI is a technique that combines neural networks and rules-based symbolic AI to address the weaknesses in each approach.
While current incumbents in the space deploy AI to help developers build code, Agemo wants to build a platform that actually understands software.
"As far as we understand, Magic, Poolside, and Cognition are building this for developers," he said. "But we've spoken to enough developers to know they don't need another dev tool."
Ramadan added that their startup differentiates itself by going beyond just providing a model — instead, creating a "system" that gives users the model and the environment.
"A nice analogy is that Poolside is designing a robot to drive a car. You plug it into your computer and it does what a human would be doing," Ramadan explained. "This isn't efficient because the LLM needs to be as good as a human."
"We want to improve the environment so the LLMs can work better. So we're building an autonomous car — you'll never trust a good enough robot to drive you," he added.
The startup is based in London, which it sees as a hotbed for AI talent. "Silicon Valley productionizes a lot of research. But there is a lot of talent in Europe that's hungry for ambitious projects, but they don't have the opportunity to do so," Zhuo noted. "We want to push on research and product development — and a lot of it's happening in London, even with ChatGPT and Llama."
A nimble fundraise
Agemo's $4 million funding is a drop in the ocean compared to the huge sums raised by other AI firms.
"I'm confident that if we were in the Bay Area, we could have raised more money," Zhuo said. "But we've been very capital-efficient in how we've been deploying that money."
He pointed to existing juggernauts such as OpenAI and Meta burning through cash while training huge models. Instead, Zhuo insisted that raising a modest sum from European investors has enabled Agemo to build a sustainable business.
The next step is to go to market — eventually in the US. The team is gearing up to release a platform called CodeWords, and focus on building their software incrementally so that it can be enterprise-ready.
"From a numbers perspective, from the amount raised, we're underdogs," Zhuo said. "We've been underdogs all our lives. We have good credentials — but ultimately, a benefit is that there isn't as much risk with us as with someone raising $500 million and releasing a product."