Diffbot’s AI model doesn’t think – it knows, thanks to its trillions of knowledge graphs

Diffbota small Silicon Valley company best known for having one of the world’s largest startups net informationannounced today the release of a new AI model that promises to tackle one of the biggest challenges in the field: virtual reality.

The a new exampleThe improved version of Meta’s LLama 3.3, is the first open-source method for a technique called graph retrieval-augmented generation, or GraphRAG images.

Unlike conventional AI models, which only rely on large amounts of pre-trained data, Diffbot is an LLM they take real time from the company Information Picturea continuously updated database with more than a trillion points of connection.

“We have an idea: that the total number of ideas can be about 1 billion,” said Mike Tung, Diffbot’s founder and CEO, in an interview with VentureBeat. “You don’t really need that model knowledge. You want to make the model better by just using the tools so that they can query the information externally. “

How it works

Diffbot is Knowledge Graph is an archive that has been crawling the Internet since 2016. It divides web pages into entities such as people, companies, properties and documents, extracting structured information using computer vision and natural language.

Every four to five days, the Knowledge Base is refreshed with millions of new content, ensuring it stays fresh. Diffbot is AI example it uses this tool by querying the graph in real-time for information, instead of relying on the static information contained in its training data.

For example, when asked about the latest news, the artist can search the Internet for the latest updates, extract relevant information, and cite the source. This system is designed to make the system more accurate and transparent than traditional LLMs.

“Imagine asking an AI about the weather,” Tung said. “Instead of providing a solution based on past data, our model consults weather services and provides solutions based on real-time data.”

How Diffbot’s Knowledge Graph beats traditional AI in finding facts

In benchmark tests, Diffbot’s approach seems to be paying off. The company claims that its model achieves an accuracy of 81%. FreshQA resultsa sign made by Google to test real-time information, better than ChatGPT and Gemini. He also scored 70.36% on MMLU-Proa more difficult form of standardized test of academic knowledge.

Perhaps most importantly, Diffbot is making its brand open, allowing companies to run it on their own devices and customize it to their needs. This will address the growing concerns about data privacy and the lock-in of major AI vendors and providers.

“You can run it locally on your machine,” Tung said. “There is no way you can run Google Gemini without sending your data to Google and sending it out of your domain.”

Open-source AI can change the way businesses work with complex data

The release comes at an important time in the development of AI. In recent months there has been a lot of criticism of the language of “be strong” or creating false information, even as companies continue to expand brand sizes. Diffbot’s approach shows another way forward, one that focuses on implementing AI systems on the basis of verifiable facts rather than trying to hide all human knowledge in a neural network.

“Not everyone just buys big and big brands,” Tung said. “You can have a model that has a lot more potential than a large model with a naive approach like ours.”

Industry experts believe that a solution based on the Diffbot Knowledge Graph can be very useful for businesses that need accuracy and transparency. The company already provides data services to major companies including Cisco, DuckDuckGo and Snapchat.

The model is immediately available through an open source release on GitHub and can be tested through public exposure on diffy.chat. For organizations that want to install it internally, Diffbot says that a small 8-billion-parameter model can be run on a single model. Nvidia A100 GPUwhile the entire 70 billion-parameter model requires two Pictures of the H100.

Looking to the future, Tung believes that the future of AI is not in the bigger picture, but in better ways to organize and access people’s knowledge: “Reality is destroyed. Most of these things will be sent to a clear place where you can exchange information and where you can have the knowledge of data.”

As the AI ​​industry grapples with issues around authenticity and transparency, the release of Diffbot provides another way to reinforce the bigger-is-better paradigm. Whether it’s a successful move in the field remains to be seen, but it has shown that when it comes to AI, size isn’t everything.