The Reuters news staff had no role in the production of this content. It was created by Reuters Plus, the brand marketing studio of Reuters.
PRODUCED BY REUTERS PLUS FOR
Disclaimer: The Reuters news staff had no role in the production of this content. It was created by Reuters Plus, the brand marketing studio of Reuters. To work with Reuters Plus, contact us here.
From black box to open book
IBM’s game-changing approach to AI
Business doesn’t need generative AI that can write poetry. Using an open-source framework, IBM is enabling more enterprises to create domain-specific models that incorporate the knowledge they want in order to perform tasks they need.
Among the billions of facts now stored within the vast datasets of the world’s large language models (LLMs), the events of May 2024 might go down in history as being pivotal in the evolution of generative artificial intelligence (GenAI).
First, IBM released a family of its most capable Granite code and language models to the open-source community; then, in partnership with Red Hat, it launched the InstructLab project to provide a first-of-its-kind, community-driven approach to LLM development through skills and knowledge contributions. The combination of these two was a watershed moment when LLMs became democratized and more transparent. It was the time when barriers to entry came down for those wishing to shape this future-defining technology.
For enterprises seeking solutions for their own GenAI projects in a dark and proliferating forest of foundation models, the outlook suddenly became less opaque and less intimidating.
IBM executives who worked on these initiatives say they were simply responding to the demand from enterprises. “Everything we put out to the market is driven by the market,” says Maryam Ashoori, Ph.D., head of product for the watsonx.ai and watsonx foundation models.
This strategy was developed with enterprise in mind. The combination of IBM’s open-source Granite models with the InstructLab initiative is a GenAI solution rooted in trust and value.
The models are licensed under Apache 2.0, so enterprises can modify them for their own commercial purposes.
With only 10% of companies having put their own GenAI solutions into production, many enterprises are wary of the legal and regulatory risks of working with big, general-purpose LLMs due to the hidden provenance of the data they have been trained on. “They are very capable models, but they are black boxes—you don’t have visibility into the datasets that have gone into them,” says Matt Candy, global managing partner for Generative AI within IBM Consulting.
For many enterprises, that is a critical flaw. “In some highly regulated industries, it really is a blocker if the model has problems in terms of the data that went into its training,” Ashoori says. “You can’t go into production."
So, IBM made the decision to create its own model and began building Granite. “The answer was: ‘Let’s start from scratch and build our own,’” Ashoori explains. Granite’s developers took “full control over the training data,” she says, and worked with colleagues at IBM Research to filter out hate speech, abuse and profanity, and remove any datasets at risk of copyright breach or pirating. The result is a foundation model trained on trustworthy, enterprise-ready data and indemnified by IBM.
A new solution
From the outset, Granite was built for business. One-fifth of its original data comes from the finance and legal domains.
“The number one most important challenge is making sure that there is a clear business return that you are going to get from applying generative AI,” says Candy.
Creating this model required a shift in mindset from the one that framed the first steps of GenAI, says Akash Srivastava, chief architect of InstructLab and principal research scientist at the MIT–IBM Watson AI Lab. Instead of starting by amassing a vast pool of data, developers must begin by identifying the use cases that their model is being built to address.
“That is how we created InstructLab,” says Srivastava. “The idea was to write down the tasks that we care about and create data synthetically for each of those tasks. InstructLab is bringing the prescriptive paradigm into constructing models.”
The biggest LLMs are not primarily designed for enterprise use cases. “I don’t need a model to be able to write me poems or help me plan my holiday,” Candy says. “We believe in the advantages of having smaller code and language models specifically built and trained for the job at hand.”
What enterprises want, he says, is a GenAI solution that enables them to make the best use of their own domain-specific data. “Competitive advantage and differentiation are going to come from the datasets that exist within the enterprise firewalls—that understanding of your customers and understanding of product, and the insights that you have as a company.”
IBM created watsonx as an AI and data platform that is built for business. It enables enterprises to use their own data as they train, validate, tune and deploy AI systems across the business, with trusted governance throughout the AI life cycle, from data preparation to model development, deployment and monitoring.
Built for business
Companies that try to use their proprietary data with closed LLMs can find it costly, environmentally unfriendly and laborious. InstructLab offers a cleaner and more cost-effective alternative: Enterprises can use their data to customize Granite’s open and smaller (7 billion-parameter) model for their own needs. “That delivers the performance they want for a targeted use case, for a fraction of the cost,” Ashoori says.
The reassurance of IBM indemnification, compliant model training processes and the inclusion of an Apache 2.0 license are particularly valuable to innovative enterprise developers, says Srivastava. “One of the most important things for deploying any technology in business is giving developers the freedom and peace of mind that their application is going to be viable within the enterprise setting,” he says.
The deployment process is simplified by InstructLab’s toolkit. “We provide you with tooling that runs on your laptop, and you can generate synthetic data for your task,” Srivastava says. “Technically, you don’t even need to be a software engineer. This enables experts in other fields of science and technology that may not be well-versed in software to be able to say, ‘Here is what I want this model to do.’”
Lower-COST Deployment
This open and collaborative approach is central to IBM’s philosophy. Last year, IBM formed the AI Alliance, in partnership with Meta, as an international community of leading technology developers dedicated to advancing open, safe and responsible AI. More than 100 organizations have joined.
“We deeply believe in open innovation and open approaches,” Candy says. “It is a safer way of bringing AI into the world, and we believe it drives innovation faster.”
A recent study by Stanford University’s Center for Research on Foundation Models identified Granite as one of the most transparent LLMs in the world. It was the leader in the Risk Mitigations category, and scored a perfect 100% for transparency in the Compute, Methods and Model Basics categories.
IBM encourages innovation and collaboration by providing enterprises with access to a mix of open-source and commercial models from IBM and third-party vendors. “We know that the future of AI is open,” Ashoori says.
The black-box approach of some general-purpose LLMs is unsuited to the growing spirit of collaboration, says Srivastava, who notes that “It feels like every engineering process five years from now will be interfaced through a code model, and it is difficult to see how that level of innovation happens in a closed company without engaging the bigger developer community.”
IBM Granite models are already proving effective in diverse use cases. Tennis fans following this year’s Wimbledon championships were able to access stories on their favorite players via a Catch Me Up feature on the official app and website, generated by the watsonx.ai GenAI platform in natural language and tuned to the style and vocabulary of Wimbledon.
In another example, within the highly regulated UK banking sector, IBM has taken its open-source language model and trained it on the industry’s regulatory framework to produce a complaints classification tool that can deliver highly accurate and effective classification and responses.
With the help of the open-source community, IBM is enabling more enterprises to create domain-specific models that use domain-specific knowledge to perform domain-specific tasks. “IBM has been the torchbearer of ‘Let’s open it and let’s put it out there,’” Srivastava says.
The future is open
“This enables experts in other fields of science and technology that may not be well-versed in software to be able to say here is what I want this model to do.”
“We believe in the advantages of having smaller code and language models specifically built and trained for the job at hand.”
For more on IBM Granite and InstructLab click here.
The black box approach of some general purpose LLMs is unsuited to the growing spirit of collaboration.
10%
Only
of companies have put their own GenAI solutions into production
100%
Granite scored a perfect
for transparency by Stanford University
Matt Candy Global managing partner for Generative AI, IBM Consulting
Akash Srivastava Chief Architect, InstructLab
$36bn
6X expected market growth by 2030
Top five
LLM developers dominated 2023 market revenue
65%
of firms regularly using GenAI
50%
of digital work expected to be automated in 2025
750m
apps using LLMs by 2025
The Reuters news staff had no role in the production of this content. It was created by Reuters Plus, the brand marketing studio of Reuters.
PRODUCED BY REUTERS PLUS FOR
Disclaimer: The Reuters news staff had no role in the production of this content. It was created by Reuters Plus, the brand marketing studio of Reuters. To work with Reuters Plus, contact us here.