#What is the significance of AI reliability?
AI hallucinations, or instances where artificial intelligence confidently generates incorrect information, have become a serious concern for many industries. A San Francisco-based startup has recently garnered attention by securing $9 million in seed funding to address this issue. Founded by Peter Elias, the company is dedicated to developing a system that significantly enhances AI reliability.
The funding round was led by a prominent venture capital firm, signaling strong confidence in the startup's vision. The aim is to create a reliability layer for AI systems, which works to identify and rectify factual errors before any misleading information is presented to users.
#How does Probably aim to achieve high accuracy?
The startup is targeting an impressive accuracy rate of 99.99 percent for tasks that require precise information. In simple terms, this means out of every 10,000 responses, the system intends to deliver at most one incorrect answer. This high level of accuracy is crucial in fields where precision is non-negotiable, such as healthcare and finance.
To achieve these ambitious goals, Probably employs a method that incorporates deterministic validators around language models. Initially, a large language model generates a response; then, the validation layer verifies this answer against reliable data to ensure accuracy before providing it to users. Every response includes citations and detailed audit trails, meeting both user needs and regulatory compliance.
#What are the advantages of using smaller models?
Interestingly, the technology developed by Probably relies on models that are four tiers below the latest frontier models. By optimizing smaller models for reliability rather than using larger ones for improved accuracy, the company hopes to deliver equivalent results at a significantly lower cost.
Another key benefit is that this system can operate on local hardware. This capability reduces expenses associated with token usage and ensures that sensitive financial or healthcare data can remain on-premises, thereby enhancing security.
#What is Probably's first product offering?
Probably is launching a data science tool as its inaugural product. This innovative tool enables users to derive reliable insights from complex datasets. It allows non-technical users to engage with data effectively, retrieving accurate, cited answers without needing extensive knowledge of intricate data architectures. This product bridges the gap between powerful data analysis and accessibility, empowering a broader range of individuals to harness the power of data responsibly.