The importance of a technology can sometimes be judged by the number of competing partner ecosystems or alliances that are created before said technology yields much in the way of practical use. 2023 will, in technology hype terms, be remembered as the year of generative AI, since while ChatGPT was actually released in at the tail end of 2022, its impact began to be felt in 2023.
As 2023 draws to a close, almost exactly a year after ChatGPT was launched, Meta and IBM formed their AI Alliance. This latest alliance joins the European AI Alliance, the AI Infrastructure Alliance, the AI Governance Alliance and The Partnership on AI, amongst others, as another industry body set up to reassure the world that the development of AI is in safe hands. What role will the AI Alliance play? At its launch Dario Gil (SVP IBM) said, “we are advocating for an open ecosystem…”
Meta and IBM, you say?
Potentially uneasy bedfellows Meta and IBM have come together, one might conjecture, because each has something the other one needs. Meta has suffered a lot of industry criticism about its open-source AI approach, with the release of its text-generating model Llama proving as attractive to bad actors as it is to developers working within the law. Meta could use a partner known for its conservative organisational mores and trusted enterprise offerings, a partner such as IBM.
IBM, on the other hand, is currently rolling out its own generative AI services for enterprise, but does not have the same market attention in this area as competitors such as Microsoft and OpenAI do. However, Meta has enough market presence to share as a Big Tech AI player and the AI Alliance may thus help IBM more successfully shine a light on its own capabilities.
Other founding members of the AI Alliance include AMD, Dell, Hugging Face, Intel, The Linux Foundation, Oracle, Red Hat, ServiceNow, and Sony. Research institutions belonging to the AI Alliance include Berkeley, CERN and Imperial College.
But leading AI research organizations such as Carnegie Mellon, Stanford and MIT are missing. And several key AI movers and shakers are also missing from the AI Alliance, such as Amazon, Google, Microsoft and Open AI, along with NVIDIA. Indeed, at the analyst briefing, Ion Stoica, Professor of Computer Science at Berkeley observed that:
The AI Alliance may be the underdogs but in open-source the underdogs are often the players that provide the most successful technology.
What will the AI Alliance do?
The AI Alliance is currently putting together a flexible program and as many member organizations are involved in different areas, each will take part in those areas best aligned with their goals. As many of the companies involved are already used to collaborating, the initial focus is on identifying existing projects where the AI Alliance can amplify results.
Such projects are likely to include the ServiceNow and Hugging Face initiative BigCode Project that aims to build out an open community around code generation tools for AI. The StarCoder LLM is a 15 billion parameter model that has emerged from this initiative and has been trained on source code that was permissively licensed and available on GitHub. StarCoder is currently being made available under an Open Responsible AI Licenses (OpenRAIL) licence, and the AI Alliance is likely to promote this licensing model.
Another issue that was raised at the analyst launch briefing was the fact that the development of non-English text for AI models is lagging and this is an issue the AI Alliance plans to address. AMD referenced the work it was already doing to support the University of Turku create a LLM that could be trained in Finnish. Hugging Face has also been collaborating with the University of Turku as part of the BigScience initiative to create the largest open language model BLOOM, which the University of Turku has combined with Finnish.
Access to sustainable compute power is something else that the AI Alliance plans to tackle. For example, the AI Alliance is providing access to hardware for Alliance research members, with IBM already spending $100 million in university partnerships to provide hardware for AI.
However, in the current climate the development of the computational power that is required by AI is gated by sustainable energy concerns. Dario Gil said that,:
We need to create multiple models and specialise for different tasks. We need to fine tune small models rather than large models because to do so is more energy efficient and faster. We can identify the best model for your use case by providing basic benchmarking of models. At the AI Alliance we want to look at these more intangible aspects of the model.
Greater use of nuclear energy to power supercomputers is likely to be a priority as this has helped Hugging Face lower emissions when developing, training and running BLOOM.
My take
At the analyst briefing, there was no representative from Meta, or, indeed from Hugging Face, which meant the launch was dominated by the traditional enterprise tech companies, lending the effort the air of a group of Dads taking to the dance floor at a teen party. Furthermore, despite the AI Alliance repeatedly pointing out that diversity is important to ensure AI evolves in sync with societal concerns, no women were visible on the rather hom*ogenous presenter panel, dominated by representatives of American multi-nationals.
It is important that we have open source models available as a check and balance to proprietary LLMs and it is probably from the open-source community that we will get more diverse and inclusive AI models. Please take diversity and inclusion on board as a priority, AI Alliance, it might give you a distinct differentiation in the AI ecosystem.