A Damodaran
Two incidents rocked the generative AI scene last month—the failure of Google’s much-vaunted new AI chatbot tool, Gemini, and the malfunctioning of ChatGPT.
Of the two, Gemini’s failure was particularly glaring, since it involved historically inaccurate texts and politically and culturally sensitive images. Google promptly went on a damage control mode, by turning off its problematic image generator from the Gemini system.
Critics of Google feel that the tech behemoth was rushed when it came to releasing the Gemini Chatbot. Former Google employees lay the blame for Gemini’s failure on the company’s ‘risk-averse leadership style’ and top-down decision making.
While these contentions may have an element of truth, the larger factor accounting for Gemini’s failure is the peculiar pattern of growth and competition witnessed in the Generative AI space, ever since Open AI launched ChatGPT.
Goliath challenges David
One of the most interesting features of the GenAI arena has been the absence of direct competition between large players. Tech Goliaths like Google are either locked in competition with talented Davids like Open AI who enjoy the first mover advantage, or fight one another through proxies (a case in point is the tussle between Anthropic and Open AI). The problem with the Goliath-David competition in the GenAI space is that nervous Goliaths who go for a swift kill end up badly bruising themselves.
Too many, too soon
The launch of ChatGPT on November 30, 2022 has triggered an avalanche of GenAI products. The list includes a slew of chatbots powered by Large Language Models (LLMs), a variety of plugins, APIs, AI-powered virtual assistants, and co-pilot aids targeted at developers and individual users alike.
The year 2023 witnessed the rapid entry of multimodal foundation AI models from leading AI tech companies, each seeking to beat the other. This trend continues in 2024. Open AI’s latest text-to-video sensation Sora, launched in mid-February 2024, has already met a match in Alibaba’s EMO.
New foundation models (FMs) claim their superiority on account of greater parameter count, larger context window, and high computation efficiency. Both Google’s Gemini Ultra and Meta’s Llama 2 are expected to outdo GPT-4. Anthropic and Mistral AI are also in the race to launch advanced FMs that can outperform GPT-4.
Perhaps what is most interesting about this race is the frenetic pace at which OpenAI’s rivals replace their own models in a bid to stay ahead of their rivals. Anthropic’s Claude 2.1 followed Claude 1.3 in four months. Meta’s launched Llama 2, three months after Llama 1.
As multimodal AI models grow in size, their data requirements vastly increase both in terms of volume and complexity, thus drastically increasing demand for data centres. This phenomenally increases the costs of building advanced FMs. OpenAI spent $4.6 billion to develop GPT-3, a product that goes back to 2020. Experts speculate that the much-anticipated GPT-5 would need finances to the tune of $1.25-2.5 billion just for training the model. Under these circumstances, the economic costs of product failure in the GenAI space can be astoundingly high.
The game that has worked well
In hindsight, OpenAI’s strategy of rapidly scaling up its customer base for ChatGPT proved to be superbly successful. In the past 15 months, OpenAI has not only retained its initial customer base but also managed to drastically enlarge the same, thanks to its adroit strategy of releasing ‘easy to use’ plug-ins, virtual assistant apps, and attractive image generation models. By abandoning its nonprofit pretensions and embracing the subscription model for its ChatGPT Plus last year, OpenAI not only warmed the hearts of its investors but also became the hot favourite of valuation gurus. Today, it enjoys a staggering valuation level of $80 billion.
Future direction
The Gemini fiasco has dealt a huge blow to leading players in the GenAI sector. In the wake of the debacle, the government of India released its advisory to AI tech behemoths, insisting that they seek prior permission before launching their products, go for disclosures, and label products which are undergoing testing.
In the future, AI FMs would be subject to elongated phases of containment and controlled deployment before their release for large-scale use. Labelling GenAI products too appears to be a reality, even in countries which have had a liberal approach towards it.
Yet another likely consequence of the Gemini fiasco is that it will bring changes in AI model training regimes. Unsupervised learning (that relies on unlabelled data for training models) is likely to face increasing scrutiny. Even training tools that rely on labelled data, like supervised learning, may come under verifiable standards and protocols.
(The author is Visiting professor, Ahmedabad University. Views are personal.)