Artificial intelligence and LLMS models that work on many useful applications, but despite their promise, they are not very reliable.
Nobody knows when this problem will be solved, so it is logical that we see startups find an opportunity to help institutions to ensure that the applications that LLM pay that are paying for work as intended.
Starting in London Composo It feels that it contains Headstart in trying to solve this problem, thanks to its custom models that can help institutions assess the accuracy and quality of the applications that LLMS operate.
Similar to the company factorFerblay, Humanloop and LangsmithWhich all claim to provide a more solid alternative to LLM for human test, review lists and current note tools. But composo claims to be different because it provides the option not to symbol and API. This is noticeable because this expands the possible market range – you don’t have to be developed for its use, and the experts and executives can assess the artificial intelligence applications of contradictions, quality and accuracy themselves.
In practice, composo Gather The bonus form trained to output the person prefers to see from the AI application with a specific set of CRITERA specific for this application to create a system mainly evaluates outputs from the application against these criteria. For example, Chatbot Medical sorting can have instructions for the customer group to verify the symptoms of red flag, and Composo can record how continued the application.
The company recently A public app programming interface launched To align composo, a model for evaluating LLM applications on any criteria.
It seems that the strategy is somewhat working: it contains names like Accenture, Palantir and MCKINSEY at the base of its customer, and has recently raised two million dollars in financing before the seed. The small amount collected here is not uncommon for the start of operation in the project climate today, but it is noticeable because this land of artificial intelligence, after all – financing such companies is abundant.
But according to the co -founder and CEO of Composo, Sebastian Fox, the relatively low number is that the startup approach is not thick capital.
“Over the next three years, we do not expect to raise hundreds of millions because there are many people who build institution models and do so with great effectiveness, and this is not our USP,” said Fox, former McKinsey Adviser. “Instead, every morning, if you wake up and see a news piece of Openai a great progress in their models, this is a good thing for my work.”
Through new criticism, Composo plans to expand its engineering team (led by co -founder and CTO Luke Markham, a former Graphcore), and gets more customers and enhances research and development efforts. Fox said: “The focus of this year is about scaling the technology that we have now through these companies,” said Fox.
British Fund for AI Twin track projects He led the seed tour, which also witnessed participation from Jvh Ventures and the prison (The latter has supported the start of the operation through its acceleration program). “Composo treats the critical bottleneck in the adoption of the AI,” said a TWIN Path spokesman in a statement.
Fox said the bottleneck is a major problem of artificial intelligence movement in general, especially in the institution’s sector. “People are more than the noise of excitement and think now,” in fact, does this really change anything about my work in its current form? Because it is not reliable enough, and it is not sufficiently consistent. Even if so, you cannot prove to me how much it is. “
This bottleneck can make composo more valuable for companies that want to implement artificial intelligence but can bear the reputable risks to do so. Fox says that this is the reason that his company chose to be inappropriate for the industry, but it still has an echo in compliance, legal care and security.
As for its competitive trench, Fox feels that the research and development required to reach here is not trivial. He said: “There is the structure of the form and the data that we used to train it,” explaining that the composo Moal was trained in a “large data collection of expert assessments.”
There is still a matter of what technology giants can do if they simply store the huge war boxes to enter this problem, but Composo believes it has the first engine feature. “The other thing (the thing) is the data we accumulate over time,” Fox said, referring to how Composo builds evaluation preferences.
Since it evaluates applications for a flexible set of criteria, Composo also sees itself that itself is better suitable for the age of artificial intelligence agents who use a more restricted approach. Fox said: “In my opinion, we are certainly not at the stage where the agents work well, and this is what we are really trying to help solve it,” Fox said.
