While AI technologies, particularly Large Language Models (LLM), continue to evolve, their primary issue—hallucinations or the misinterpretation of precise facts—remains relevant. The Probably startup, aiming to fundamentally solve this problem, recently raised $9 million in funding from the Andreessen Horowitz venture fund. This project promises to raise the accuracy of AI systems to 99.99%, similar to traditional software. This is reported by Techcrunch.com news reports.
According to Probably founder Peter Elias, the company’s main goal is to completely stop incorrect information from reaching the user. Today, ChatGPT or other popular models sometimes confidently provide wrong information. The startup has developed a special “data science mech suit” system to control this process. This system verifies the AI’s response through a deterministic validator and immediately rejects results that do not match the database.
Small Models and High Efficiency
The first product developed within the project is an analytical tool designed to obtain quick answers from complex datasets. Each result is provided with links to sources and an audit trail. According to ixbt.com, the uniqueness of the system is that it can operate with high accuracy even on much simpler and smaller neural networks instead of the most powerful and expensive models.
Elias notes that by sufficiently clarifying the context, the model’s load can be reduced. Currently, the Probably tool uses models that are four times weaker than the leading (frontier) models but are capable of running on local computers (without data centers). This allows companies to significantly reduce AI token costs.
Future Prospects and Sectors
Probably technology is not limited to data analysis alone. This engine is planned for future application in the following sectors where accuracy is critical:
- Accounting and financial reporting;
- Medical diagnostics and services;
- Legal document analysis;
- Engineering and technical design.
The startup founder also touched upon why large tech labs are not rushing to create such systems. In his opinion, giants like OpenAI or Google have a financial interest in users sending repeated requests to correct model errors, thereby consuming more tokens. Probably, conversely, aims to save user time and money by achieving 100% accuracy on the first attempt.
This investment could signal a new stage in the AI market. If the Probably team delivers on its promise, we will soon enter an era where errors of “smart” bots are completely eliminated. This will expand the possibilities for integrating AI into risk-free business processes.