Meta recently launched Llama 2, an open source large language model (LLM) that’s become the latest installation in a series of Big Tech handouts. Hailed as a step towards democratising AI innovation and lowering barriers to entry, many have applauded its release.
Nick Clegg, Meta’s president of global affairs, said on BBC Radio 4’s Today programme that making LLMs open source would increase safer by inviting external scrutiny.
Sceptics, on the other hand, have begun to question the politics of Silicon Valley. Some interpret the power play by Meta as another step towards reifying the dominance of tech giants in the field. Others cynically interpret it as a golden covered path for criminals to take advantage of AI.
Open sourcing AI is not a new debate. In fact, Meta has previously open sourced some of its infrastructure and AI work, from PyTorch to ImageBind. However, what is different this time is the pace at which AI is accelerating, given the increased investor interest in AI startups.
Regardless of which company decides to open the door, open sourcing AI could be highly consequential for startups.
Short for large language model, LLMs are a type of artificial intelligence that uses deep learning techniques and massive large data sets to understand, summarise, generate and predict new content.
Llama 2: driving innovation?
Meta’s release of Llama 2 is designed to make it easier for small companies and lone coders to create new products and services..
Mining the data that underpins an LLM requires deep pockets that the average startup does not have. Researchers estimate that training GPT-3 cost OPenAI nearly £4 million.
Following the launch of Llama 2, Zuckerberg explained his rationale on Facebook. “Open source drives innovation because it enables many more developers to build with new technology.”
Rafie Faruq, CEO of Genieai.co, an AI legal assistant and 2023 Startups 100 Index alumni, welcomes the release of Llama 2. “It’s huge,” enthuses Faruq.
“Open Source LLMs offer the ability to build and innovate on top of existing technology, reducing both time and costs associated with the development of new AI products, as well as removing the potential monopoly large incumbents like OpenAI have.”
Fahad Syed, data scientist manager at Sprout.ai, an AI insurance claim company also featured in 2023’s Startups 100 Indexi, agrees. “Being open-source not only fosters transparency but also allows organisations and teams to leverage and build upon the existing work, further enhancing the utility and potential applications of the technology.”
However, Syed isn’t convinced that Meta alone can level the playing field. . “While Llama 2 brings substantial advancements to the table, it may not drastically alter opportunities in the rapidly changing AI industry.”
“The most significant impact may emerge from encouraging the release of open-source models from other tech companies, as this can streamline the path for startups to create novel AI products,” he reveals.
Paweł Budzianowski, Head of Machine Learning at PolyAI, similarly agrees. “The release of Llama 2 is definitely a good sign that big players are willing to nurture and grow open-source projects. This said, the Meta brand gives it more weight than its actual potential – the power of the model is not substantially higher than other open-source counterparts like Falcon.”
What are the benefits for startups?
Startups do have plenty to gain from open source LLMs. “Most startup companies can’t afford to train their own foundation language models, but can afford to fine tune them or use them as they are, otherwise known as running the model at ‘test time’, ” explains Faruq.
“This removes the competitive advantage of large incumbents and democratises the ability for other machine learning companies to provide downstream applications on top of existing foundation language models.”
For Genie AI, a startup with a legal data set of over 12,000 companies – using its AI legal assistant to fine-tune an open source model streamlines multiple processes.
Sprout.ai’s Syed agrees.“These innovations represent excellent tools for us. With our existing knowledge and expertise, it’s far easier for us to adapt and integrate models like these into our product offering enabling us to continue delivering high-quality, innovative solutions to our clients,” he reveals.
As artificial intelligence continues to develop, the reality remains that Big Tech has the deep pockets and resources required to build LLMs that startups currently lack. Taking those Big Tech handouts, therefore, is in the interest of startups.
Open source AI is not a panacea
Nevertheless, open source AI is not exactly the only golden key that will open the door towards the industry’s idea of tech utopia. Before using an open source LLM, startups need to consider the risks attached to its use.
“It’s important we navigate potential challenges like lack of control over the model, data privacy and security concerns, and dependence on the open-source project,” discloses Faruq.
Additionally, startups need to be honest about their capabilities before adopting an open source AI LLMs. “Businesses need to ensure that they have the expertise to work with these models, and that they have a clear understanding of the licences under which these models are released,” warns Syed.
He also cautions, “Open-source projects require active maintenance and contribution, and while the community can provide support, the primary responsibility often still lies with the business.”
Despite these considerations, Syed believes the challenges are outweighed by the benefits the transformative power open source AI can have on the industry’s level of collaboration, innovation, and transparency.
Can AI be truly democratised with open source LLMs?
The general consensus from startups is that the release of Llama 2 is a step in the right direction, but there’s a long path ahead before true AI democratisation.
Fostering the right terrain to maximise innovation in AI will require a mix of governmental policy and regulations to distribute benefits and resources.
“Without addressing the issues concerning the affordability of computation, data access, and infrastructural control, the dominance of these tech giants may persist,” predicts Syed.
“The resolution may lie in policy interventions, open data initiatives, and efforts to provide affordable computation resources, such as the establishment of AI research cooperatives. These measures collectively contribute a more accessible and equitable AI landscape,” .
Whilst it’s unclear how long that will take, startups do concur that open sourcing AI, for now at least, is the right step forward.
Read more:
Leave a comment
We value your comments but kindly requests all posts are on topic, constructive and respectful. Please review our commenting policy.