Another major milestone for Veilance and the entire ecosystem. After almost a month of development and training, we've finished six rounds of models, ranging from 4 billion to 32 billion parameters, with an infrastructure cost of $3,511.02 USD. To accomplish this training, we leveraged infrastructure hosted on lambda.ai, equipped with: - 56 vCPUs - 400 GB of RAM - 4 NVIDIA A6000 GPUs Training at this scale requires massive amounts of computational power, and thanks to the @nvidia Inception program, we leveraged our access to GPU credits to cover the infrastructure cost. This support provided us with access to the compute power needed to complete the Verity training and reach a major milestone that would have otherwise required massive amounts of out of pocket expenses we can now use for marketing, and fund raising. The result? Verity: a large language model built to specialize around privacy on the web. Unlike general use LLMs, Verity was developed to help users understand how websites collect, process, and share their information. Its capabilities include: - Analyzing websites and their privacy policies. - Providing advice and recommendations to improve online privacy. - Explaining complex privacy information in language easy to understand. - Accessing and analyzing intelligence collected through the Veilance Intelligence Network. - Helping users understand relationships between what websites disclose in their policies and what their behavior actually reveals. This is way more than just completing a training run. We've built specialized AI infrastructure around real browser telemetry and privacy intelligence, with the goal of making online privacy much more transparent, understandable, and accessible. A huge thank you to @nvidia for access to the Inception program for their help in making this possible. The completion of Verity is a massive step forward for Veilance, and we're ecstatic about what comes next.

Introducing the Veilance Intelligence Network. Search across thousands of real life website visits, and compare website behavior across multiple countries and real user sessions. Learn more today: veilance.org/intel-network Get paid to upload anonymized telemetry: veilance.org/paid-to-browse

See what websites don’t show you

Verity-4B-RC has been trained and is working successfully. It has approximately a ~72% task score up from ~50% from the last version. The new version of Verity will be released later on. Afterwards, we’ll be opensourcing older Verity versions once they are no longer in use.