The Generative AI Arms Race
In 2017, Google introduced the Transformer architecture, which initially did not foretell the significant impact it would have on the field of AI. Today, one particular Transformer model, known as GPT, dominates our newsfeeds. In the past year, our lives have been impacted by ChatGPT. This groundbreaking conversational AI system revolutionized our interactions with AI. However, the landscape of generative AI is rapidly changing, and the rise of the open-source community poses a challenge to industry giants like Google. In this blog post, we will delve into the journey from the inception of ChatGPT to the emergence of the open-source community and its impact on the ongoing generative AI arms race.
ChatGPT, A Game-Changer
When OpenAI unveiled ChatGPT, it marked a turning point in the field of generative AI. The popular chatbot experienced remarkable success shortly after its launch. According to a UBS study, it reached an estimated 100 million monthly active users within just two months, making it the fastest-growing consumer application ever. TikTok, for instance, took approximately nine months after its global launch to reach the same user count. Similarly, Instagram, took around two and a half years to achieve the 100 million user milestone. Analytics firm Similarweb reported that in January, around 13 million unique visitors engaged with ChatGPT daily, more than double the previous month. UBS analysts noted that this growth rate was unprecedented in the history of consumer internet apps. ChatGPT’s impressive language understanding and fluency revolutionized the field of (generative) AI, captivating users and finding applications in content creation, customer support, and even personal companionship.
Google’s Market Leadership and a Looming Challenge
Google, renowned for its cutting-edge technological innovations, has long been regarded as a market leader in the field of AI. However, last May an AI researcher at Google wrote a subsequently leaked internal memo titled: “We Have No Moat, And Neither Does OpenAI”. Google acknowledges the shifting landscape and the challenges it faces in maintaining its competitive advantage. In a surprising statement, Google writes that it has no moat, referring to its lack of a distinct competitive barrier against potential rivals. While it seems apparent that the biggest rival of Google is currently OpenAI, they make clear that the open-source community is the new kid on the block that will eventually win the AI arms race. This recognition reveals a broader acknowledgment within the industry that the generative AI arms race is intensifying, nudging/forcing industry leaders into adapting to the changing dynamics.
The Open-Source Community, A Disruptive Force
The open-source generative AI movement, led by non-profit research groups and start-ups like EleutherAI, Together.AI, Technology Innovation Institute (TII) and Stability AI, have been driving innovation and collaboration within the AI community. HuggingFace, a machine learning platform, plays a central role in facilitating the sharing of open-source models, datasets, and training code, fostering a culture of knowledge sharing and problem-solving.
Overcoming initial challenges, open-source models are rapidly closing the gap with their commercial counterparts. With Meta’s LLaMa2 and TII’s Falcon models open-source now rivals proprietary models in performance on commonly accepted language tasks. According to HuggingFace’s blog on Falcon180B, the overall performance is somewhere in between GPT3.5(default model behind ChatGPT) and GPT4. This makes open-source models increasingly attractive to businesses seeking cost-effective and customizable AI solutions. Embracing permissive licensing has also further enhanced open-source models’ commercial viability, leading to broader adoption and integration in various industries and applications. The open-source movement’s success is reshaping the AI landscape, challenging established industry players, and demonstrating the power of collaboration and accessibility in driving technological progress for the greater good.
The Power of Open-Source: Faster, More Capable, and Highly Customizable
How can the open-source community close the gap so fast? They improved Large Language Models (LLMs) in terms of quality, speed, customizability, and privacy. Additionally, they advanced LLM to be accessible to a broader audience, promoting a healthier and more competitive landscape built on knowledge sharing.
One notable stride forward is the adoption of innovative techniques such as low-rank adaptation (LoRA). By implementing LoRA, which accelerates LLM fine-tuning by freezing the number of trainable parameters, the open-source community has paved the way for fine-tuning processes even on consumer-grade GPUs. This breakthrough has made it possible for individuals, startups, and smaller companies to harness the power of large language models, thus lowering entry barriers and democratizing AI.
Another astonishing innovation is the use of 4-bit quantization, a technique that efficiently compresses the size of language models. This advancement has not only made these models more accessible to companies but has also empowered individual enthusiasts and researchers to experiment with cutting-edge AI tools and technologies.
What’s truly remarkable is how these open-source models achieve impressive feats with relatively limited resources, outperforming some industry benchmarks while offering faster training times and extensive customization options. This is a testament to the power of collective effort and knowledge sharing.
By lowering the barriers to entry and emphasizing the importance of accessibility, the open-source community has not only democratized AI but has also sparked a new era of innovation. In this environment, creativity and collaboration thrive, enabling even more breakthroughs on the horizon.
Conclusion
The generative AI arms race has witnessed a remarkable evolution, with ChatGPT heralding a new era of conversational AI. While industry leaders like Google acknowledge the absence of a competitive barrier, the rise of the open-source community presents a paradigm shift in the field. The community’s rapid advancements and disruptive innovations have showcased the transformative potential of collaborative knowledge sharing.
To stay at the forefront of the generative AI revolution, industry leaders must embrace openness, collaboration, and the integration of open-source initiatives. The future of generative AI lies not only in the hands of corporations but also in the collective power of the open-source community.
If you’re eager to explore how your organization can leverage the potential of (open-source) Large Language Models (LLMs), don’t hesitate to get in touch or connect with Sogeti’s AI team. They are committed to delivering value to customers through the strategic usage of LLMs. Moreover, if you’re interested in delving into the realms of trust and ethics surrounding LLMs, we invite you to dive into the insightful ‘ChatGPT and I have trust issues’ blog series.
