As we previously reported, enterprise CIOs have been slow to embrace generative AI. One reason is that AI doesn't fit into existing software engineering workflows because it literally doesn't speak the same language. For example, LLM (aka Large Language Model) requires a lot of deception to deliver valid JSON.
That's where Dottxt comes in, a US-based startup with the promise of “making AI talk computers.” The company is led by a team that helps outline open source projects. This allows developers to get what they want from ChatGPT and other generative AI models without resorting to crude tactics like injecting emotional blackmail into prompts (like “write the code or write the kitten”). You can get it. Understood! ”).
Using software libraries such as Outline, Python libraries, Microsoft guidance, or LMQL (also known as Language Model Query Language), you can use an approach called structured generation (or sometimes You can guide the LLM in a different way. constrained generation).
As the name suggests, the focus of this technique is on the output of the LLM rather than its input. In other words, teach the AI model how to answer, says Remi Roof, CEO of Dottxt.
This approach “allows us to return to traditional engineering workflows,” he told TechCrunch. “I will refine my grammar until I understand it correctly.”
Dottxt is model-agnostic, offers more features, and provides better performance than the open source project it is based on (Outlines), making it a great choice for building powerful structured generation solutions. That's what I'm aiming for.
Louf is a French national with a PhD and multiple degrees, and like the rest of the Dottxt team, has a background in Bayesian statistics. This basis in probability theory is thought to have opened our eyes to the possibility of structured generation. Beyond AI, my IT savvy also influenced my decision to start a company focused on helping others take advantage of generative AI.
It is no coincidence that Roof mentioned grammar. Dottxt is based on the premise that most of the text we interact with is highly structured. Of course there is code, but there are also many other templates for LLMs to follow to actually be useful in a working environment.
GPT creator OpenAI recently introduced a form of structured generation called Structured Outputs. And we paid tribute to Outlines as part of that “inspiration.”
Meanwhile, Louf believes Outlines' popularity is a sign that a different, more flexible approach with more bells and whistles is needed. And investors seem to agree. Dottxt raised $11.9 million in a few months.
The startup secured a $3.2 million pre-seed round led by deep tech VC firm Elaia in 2023, followed by an $8.7 million seed round led by EQT Ventures in August of this year. In the meantime, Roof and his co-founders have focused on proving that their approach doesn't impact performance. During this time, demand for open source outlines exploded. They say it has been downloaded over 2.5 million times. This inspired them to think big.
Raising more money made sense for another reason. Dottxt's co-founders wanted to use the money to hire more people to meet the growing demand for structured generation tools. The startup's fully remote team will reach 17 employees by the end of this month, up from eight in June, Louf said.
The new staff includes two DevRel (developer relations) experts, reflecting Dotxt's ecosystem-building priorities. “Our goal over the next 18 months is to accelerate adoption beyond the commercial side,” Louf said. However, he also said that commercialization is expected to begin within the next six months, with a focus on enterprise customers.
This could be a potentially risky approach if the AI hype has died down by the time Dotxt seeks more funding. But the startup believes there is substance behind the bubble. The hope is precisely to help businesses derive real value from AI.
Julien Chaumond, CTO of Hugging Face, which is also an investor in Dotxt, recently called Structured Generation “the future of LLM.” So the hype extends to this area of the GenAI technology stack as well.
With other tailwinds, such as the rise of AI agents and small-scale AI models, Dottxt's bet on adoption could pay off. “Everyone will be using structured generation within a few years, there's no doubt about that,” Louf predicted.