Photo credit: venturebeat.com
Diffbot, a notable Silicon Valley firm recognized for its extensive web knowledge index, has unveiled a cutting-edge AI model aimed at tackling one of the most significant hurdles in AI development: ensuring factual accuracy.
The latest model, which is based on a fine-tuned version of Meta’s LLama 3.3, represents the pioneering open-source approach to a framework known as Graph Retrieval-Augmented Generation (GraphRAG) .
In contrast to traditional AI models that depend on static training data, Diffbot’s LLM utilizes real-time data accessed from the company’s Knowledge Graph, a dynamic database that holds over a trillion interconnected data points and is continually updated.
“We believe that general-purpose reasoning will ultimately be distilled down to about 1 billion parameters,” stated Mike Tung, Diffbot’s founder and CEO, in a conversation with VentureBeat. “The goal isn’t to embed knowledge within the model; rather, it is to enable the model to effectively use tools to query knowledge from external sources.”
Understanding the Mechanics
The Knowledge Graph represents an expansive automated database that has been scanning public websites since 2016. It organizes various web pages into distinct entities—such as individuals, corporations, products, and articles—by employing advanced computer vision techniques alongside natural language processing.
Updated every four to five days, the Knowledge Graph incorporates millions of new facts, keeping its information fresh. Diffbot’s AI model taps into this resource through real-time queries to retrieve current information, as opposed to depending on static data embedded within its training framework.
For instance, when prompted about a recent event, the model can look up the latest information online, discern relevant facts, and credit the original sources. This mechanism is designed to enhance both accuracy and transparency in comparison to conventional LLMs.
“Consider asking an AI about the weather,” Tung explained. “Instead of composing a response from outdated training data, our model interacts with a live weather service to provide an answer backed by current information.”
Diffbot’s Edge Over Traditional AI in Fact-Finding
Benchmark assessments indicate that Diffbot’s innovative approach is yielding positive results. The company claims that its model achieved an accuracy score of 81% on FreshQA, a testing framework developed by Google to evaluate real-time factual knowledge, outpacing both ChatGPT and Gemini. In addition, it earned a score of 70.36% on MMLU-Pro, a more challenging variant of a standard academic knowledge test.
Notably, Diffbot has opted to make its model fully open source, enabling organizations to deploy it on their own servers and tailor it to their specific requirements. This decision addresses concerns about data privacy and the potential for vendor dependency associated with larger AI service providers.
“You can operate it on your own machine,” Tung highlighted. “In contrast, using Google Gemini necessitates transmitting your data to Google.”
Transforming Sensitive Data Management with Open Source AI
This release comes at a critical juncture in the evolution of AI. In recent times, the field has faced increasing scrutiny regarding the propensity of large language models to “hallucinate” or produce inaccurate content, even as organizations focus on enlarging model sizes. Diffbot’s methodology reflects a potential alternative direction—prioritizing the grounding of AI technologies in verifiable facts rather than attempting to encapsulate all human knowledge within neural networks.
“Not everyone is focused solely on building larger models,” Tung remarked. “A model can possess even more capabilities than a larger one by employing a nontraditional approach like ours.”
Experts in the industry suggest that Diffbot’s knowledge graph methodology may hold significant promise for enterprise-level applications that require high levels of accuracy and traceability. The company already collaborates with prominent organizations such as Cisco, DuckDuckGo, and Snapchat.
The model is now accessible via an open-source release on GitHub and can be trialed through a public demonstration at diffy.chat. For companies interested in internal deployment, Diffbot indicates that the smaller 8 billion parameter version is operable on a single Nvidia A100 GPU, while the comprehensive 70 billion parameter version necessitates two H100 GPUs.
Looking toward the future, Tung expresses the belief that the trajectory of AI development should focus less on sheer model size and more on improved frameworks for organizing and accessing knowledge: “Facts become outdated. Many facts will be relocated to explicit spaces where they can be adjusted and traced.”
As the AI sector confronts the pressing matters of factual accuracy and transparency, Diffbot’s introduction of this model presents a noteworthy alternative to the prevailing approach of prioritizing larger models. Its ability to influence the future trajectory of the field is yet to be determined, but it certainly illustrates that in the arena of AI, size is not the only consideration.
Source
venturebeat.com