IBM made a significant push toward open-source large language model (LLM) development Tuesday with InstructLab, its new AI project. Created in partnership with software subsidiary Red Hat, InstructLab reportedly aims to provide a model-agnostic space from which anyone can learn and contribute to the evolution of LLM infrastructure. The platform announcement coincides with the open-source release of IBM's Granite models, which have historically been reserved for business use.
Generative AI tools invariably rely on LLMs to operate. While the internet isn't exactly short on LLMs themselves, most models are proprietary (like OpenAI's GPT models) or not as "open" as their creators say they are (like Meta's Llama models). This means developers interested in leveraging LLMs for a project must "fork"—or copy the architecture and codebase for—an existing LLM, then spend significant resources tailoring the forked version for their particular use. Even after spending a lot of time and money tweaking the forked model, there's no way for developers to send what they've learned upstream, where the original LLM could otherwise benefit from the new project.
There's also a large-scale economic component to the closed-off nature of LLMs. With just a few tech giants controlling the go-to LLMs, some organizations, like Mozilla, are concerned that the AI landscape could remain "dominated by a handful of companies" for a long time. In March, Mozilla sent a letter to the US Department of Commerce urging the government to mitigate this risk by embracing open-source AI models. The letter was signed by nearly 50 AI firms, nonprofit organizations, and academics. (Note: PCMag is owned by ExtremeTech's parent company, Ziff Davis.)
InstructLab is IBM's answer to the open-source AI push. Granite, its multi-size collection of foundation models, originally debuted under its business-oriented data and AI platform, IBM Watsonx, in September 2023. Both Granite.13b.instruct and Granite.13b.chat use a "decoder" architecture that predicts the next piece of information in a sequence. The models, trained on 7TB of data, were advertised as ideal for code and natural language, with potential use cases across academic, legal, finance, development, and general web contexts. But only those with Watsonx subscriptions could use them.
Now, anyone can experiment with Granite 7B, an "enhanced version of Granite" that serves as an open reference implementation of Meta's Llama-2-7B. IBM has also open-sourced its own Labradorite, derived from Llama 2, and Merlinite, derived from Mistral AI's Mistral. According to a statement shared with PCMag, the hope is that InstructLab will lower the barrier to entry for developers interested in using LLMs. For those still requiring an enterprise solution, Red Hat is releasing Enterprise Linux, an open hybrid cloud solution for large businesses.
