In August, Gunderson Dettmer, a Silicon Valley based international law firm, made headlines when it became one of the first U.S.-based law firms to develop and launch a “homegrown” internal generative AI tool, which it named ChatGD. The firm’s Chief Innovation Officer, Joe Green, explained how given the firm’s focus on working with the most innovative companies worldwide, it made sense to explore the opportunities with AI technology directly.
More than four months into this venture, the firm has drawn some significant conclusions about the AI tool’s potential and its limitations. It has been closely monitoring the technology’s adoption within the firm, gathering data on how it’s being utilized, and calculating the cost of this proprietary AI.
From the firm’s initial observations, nearly half the firm has already used ChatGD, and usage and engagement continue to increase steadily. The tool has been useful in various tasks, but Green reports no surprising or unanticipated uses of ChatGD. Possibly, the training provided to the users focused them on specific use cases, which they stuck to, effectively utilizing the AI tool in their day-to-day work.
In a surprising turn of events, the overall costs associated with ChatGD were not as high as anticipated, with Green estimating an annual cost of less than $10,000. This low figure is attributed to two key strategic decisions: self-hosting an open-source model for RAG vector embeddings and leveraging GPT 3.5 Turbo for both pure chat and RAG functionalities, instead of using the most expensive models available.
Gunderson Dettmer released significant updates to ChatGD this week, aimed at improving the user experience based on the feedback they have received. The firm is currently utilizing three different foundational models in ChatGD’s tech stack, selecting the best model for specific tasks.
Gunderson Dettmer views the development of the AI tool as an ongoing experiment. The firm’s adoption of AI technology has made them savvier consumers of technology in their field, allowing them to identify products with added value and engineering time behind them. The experiment also revealed the steps that will be needed to achieve higher value use cases without a step change in the technology’s capabilities.
More details on the firm’s use of the AI tool and the results of the ‘experiment’ so far can be found in this article.