OpenAI is expands a program, Custom Model, to help enterprise customers develop custom AI production models using its technology for specific use cases, domains and applications.
Custom Model launched last year at OpenAI’s inaugural developer conference, DevDay, offering companies the opportunity to work with a team of dedicated OpenAI researchers to train and optimize domain-specific models. “Dozens” of customers have signed up to the Custom Model since then. But OpenAI says that, working with this initial group of users, it realized the need to develop the program to further “maximize performance.”
Hence assisted detail and specially trained models.
Assisted detailing, a new component of the Custom Model program, leverages techniques beyond refinement — such as “additional hyperparameters and various methods of fine-tuning effective parameters at a larger scale,” in OpenAI’s words — to allow organizations to build data training pipelines, evaluation systems, and other supporting infrastructure to enhance model performance in specific tasks .
As for custom trained models, they are custom models built with OpenAI — using OpenAI core models and tools (e.g. GPT-4) — for customers who “need to deeply refine their models” or “infuse new domain-specific knowledge,” says OpenAI.
OpenAI cites the example of SK Telecom, the Korean telecommunications giant, which partnered with OpenAI to improve GPT-4 to improve its performance on “telecommunications-related conversations” in Korean. Another customer, Harvey — who builds AI-powered legal tools with support from the OpenAI Startup Fund, OpenAI’s AI-focused business arm — worked with OpenAI to build a custom case law model incorporating hundreds of millions of words legal text and comments from authorized expert lawyers.
“We believe that in the future, the vast majority of organizations will develop custom models that are tailored to their industry, business, or use case,” OpenAI writes in a blog post. “With a variety of techniques available to build a custom model, organizations of all sizes can develop customized models to realize more meaningful, concrete impact from their AI implementations.”
Image Credits: OpenAI
OpenAI is reportedly flying high catching up a staggering $2 billion in annual revenue. But there’s certainly internal pressure to keep pace, particularly as the company plans a $100 billion data center co-developed with Microsoft (if References must be believed). The cost of training and servicing flagship AI models isn’t going down anytime soon, and consulting work like training custom models might just be the thing to keep revenue growing while OpenAI plans its next moves.
Improved and customized models could also reduce the pressure on OpenAI’s model serving infrastructure. Custom models are in many cases smaller and more efficient than their off-the-shelf counterparts, and — as demand for genetic AI soars — they are undoubtedly an attractive solution for historically computing capacity-challenges OpenAI.
Alongside the expanded Custom Model program and custom model building, OpenAI today unveiled new model tweaking capabilities for developers working with GPT-3.5, including a new dashboard for comparing model quality and performance, support for integrations with third-party platforms (starting from the AI Weights & Biases developer platform) and tool improvements. However, mum is the word on the detail for GPT-4, which launched in early access during DevDay.
