Initial views on Recent Developments at OpenAI

OpenAi’s recent Developer Day and the launch of GPT-4 Turbo, the new Assistants API and the ability to create custom versions of ChatGPT for specific purposes (GPTs) presents the contracting industry with both a massive opportunity and a potential bombshell.

For the world of contracting, it's potentially a large step forward, both in terms of the functionality of OpenAI’s products (and other GenAI) and giving more clarity on how GenAI and other contracting tools and platforms will work together.

The key developments were:

• The introduction of GPT-4 Turbo, a faster and more capable iteration of GPT-4 with a much larger 125k-token context window - essentially enough to fit a 330 page contract in a single call - and improved instruction-following ability.
• Substantial cost reductions for both input and output tokens.
• The new Assistants API, which commoditises some of the trickier technical aspects of context management (most impactfully, the storage and retrieval of relevant context) and has the potential to level the playing field and accelerate development of new prototypes and products on the platform.
• Enhanced access to model customisation for both the consumer-facing product (GPTs) and enterprise clients.
• Announcement of an upcoming “app store” for custom GPTs assistants. 

Having participated in the Developer Day and spent some time experimenting with the new products, here are our initial views on their impact on contracting.

1. With its larger context window and improved instruction following, GPT-4 is now much more capable at identifying, summarising and analysing key parts of a contract.  Previously, users or technology vendors or consultants on their behalf had to provide other solutions such as semantic (full-text) and lexical (vector embeddings) search and machine-learning classifier-based approaches for identifying and retrieving relevant context, potentially rendering some existing emerging context-management techniques redundant.  GPT-4 now does a reasonable job at responding to queries on long contracts.
2. However, the final decision on how best to compose the actual prompts sent to the model remains vital and the responsibility lands squarely on those building products.  In contrast to the technical improvements to context state management, we expect that this is something that OpenAI will not easily be able to commoditise as each use case and client requires careful customization.
3. Contracting solutions will be even more tailored to an individual enterprise’s rules and data, made possible by improvements to the model and its interoperability.  For example, the model can be more easily directed to automatically redline a draft contract based on an enterprise’s own clause library or past executed contracts (and in quite a targeted way).
4. Some really exciting contracting workflows are now even closer.  These include automated redlining, data export, reporting, obligation management and risk analysis and management.
5. Specifically, there are exciting possibilities in the ability to analyse and manage a portfolio of executed contracts provided there is effective ‘plumbing’ between contract stores or contract technologies and the model.
6. The improvements make it easier to run queries and extract insights from a new ‘inflight’ contract, the store of executed contracts and other enterprise databases such as ERP, Finance and CRM.
7. It is noteworthy that OpenAI repeated their pledge that data supplied to GPT through any of their developer API endpoints (including their own and those available through Microsoft Azure) is not used to train their models.  The consistency of messaging from OpenAI around this issue is a necessary, but not sufficient, condition to alleviating data security concerns that are so important to contracting.  Clear lines of communication with clients and diligent anonymisation of PII before sending any information to the model remains critical.

‍

We’ll post further on these developments in the next few weeks as the dust settles.

The Lexical Labs Team.