The global artificial intelligence (AI) community rocked excitedly in their home office chairs in anticipation of one of the few positive and advance-scheduled arrivals of the last year on an otherwise gloomy June day, dominated by the apocalyptic headlines so typical of 2020. Software engineers and researchers in AI hubs across Silicon Valley, London, Tel Aviv and Beijing followed in awe as the latest and most powerful generation of a much anticipated natural language processing (NLP) model was released by OpenAI, a leading AI research laboratory based out of San Francisco.
Among the keen followers, though tucked away in Edmonton, Alberta, was software engineer and scientist Jasmine Wang, herself a former researcher at OpenAI. Weathering the coronavirus storm in her native Canada, Wang sat at her desk in her family home for two days straight after the model release to play around with it – something that would go on to inspire ideas that would later turn into multiple startups.
She was not alone. Countless are the business applications that this NLP model named GPT-3, which uses deep learning to produce human-like text, can power – and many have understood that. But, while GPT-3 has great potential for generative value, it is set to fundamentally reshape the dynamics of the AI world and the competitive landscape in the Software-as-a-Service (SaaS) startup space because of its very own business model.
For one, GPT-3 breaks the mold of past AI models, which have traditionally been open source, which gave developers an inside view into the workings of the model and allowed them to add to it. Now with GPT-3, OpenAI provides the ready-made model as a commercial product in the form of a “text in, text out” interface. As such, GPT-3 resets the rules of the AI model game because it does not give away its code but merely offers an easy-to-use application programming interface (API) on a commercial basis, allowing developers to tap into the GPT-3 power while not giving away a peak into its inner workings.
This is a revolutionary move not only because it monetizes AI research but because, for all the startups and established businesses looking to use GPT-3, it creates a unique dependence on a closed model. This puts OpenAI in a powerful position, especially with regard to (any potential upward adjusted) pricing for access to the model. Founders and leaders of businesses built around GPT-3 will need to maintain access to GPT-3, or they might see the very basis of their business dwindle away, diminishing their bargaining power vis-à-vis OpenAI.
In reality, however, the GPT-3 business model is only one of multiple approaches, and moves by other AI powerhouses will contribute to the ongoing evolution of the AI space. Take tech beacons such as Google, which may wonder about the economic incentives to enter this space on a commercial basis. Its NLP model BERT powers, among other things, Google’s Autocomplete and Smart Compose solutions. BERT was open-sourced in 2019, though its weights – the quintessential parameters in an NLP model – were not released. This model allows developers to use the BERT source code but requires them to still train their own data to arrive at their own weights. Whether the Google model of code-but-no-weights release, the GPT-3 commercial API model or yet another operating paradigm for AI models will prevail is yet to be seen. In any case, this field will remain subject to dynamic movements.
Aside from the startup dependence on GPT-3 and exposure to potential volatility resulting from the dynamic interplay of GPT-3 and other AI model monetization schemes, the novel GPT-3 business model is also a paradigm shift in favor of developers because they no longer have to train their own models. This frees up time and resources for more innovative thinking rather than purely building, feeding and refining models. allowing a much broader user base to take advantage of AI. In the spirit of democratizing AI and in keeping with its mission of ensuring that AI benefits all of humanity, OpenAI thus makes AI broadly accessible, though at the expense of transparency.
All of this is poised to fundamentally shift dynamics not only in the AI space but also in the SaaS startup world because it empowers a new class of “citizen developers”. With everyone having easy access to AI, the main concern – besides the safe and responsible use of AI, and the improvement of AI systems to become more human-positive – becomes the defensibility of a business idea based on AI. The barriers to coming up with SaaS business ideas and producing AI-powered products gets lowered, which is poised to result in a more competitive AI startup space. In the future, defensibility and superior business performance may not come from the underlying AI itself but product design, marketing capabilities, or simply execution.
“There are countless ways in which GPT-3 can be used for businesses. It is like a conversational partner who knows more than you. You can ask it any English language question and it will give you an astonishingly accurate answer – better than a human could. The use cases are endless,” says AI enthusiast and repeat founder of GPT-3 powered startups Wang, describing GPT-3 and its potential for businesses.
When Wang was at OpenAI herself, she worked on the release of GPT-3’s predecessor model in the GPT-n series, aptly named GPT-2. Built on 1.5 billion parameters, it was considered state-of-the-art when it launched in 2019. Its younger sibling now has 175 billion parameters – in excess of a hundred times more than GPT-2 and 10 times more than the previous largest model, Microsoft’s Turing NLG (a fact that may have inspired Microsoft to strike an exclusive partnership with OpenAI to license the GPT-3 technology – including source code – for their own use).
Startups built on GPT-3 have sprouted since its June 2020 launch, despite it still being in private beta mode and thus available only to a limited audience. Wang’s own B2B SaaS endeavors, including Copysmith and companyinabox.ai, are two out of many examples. In both cases, a business would enter a few keywords about their products or services, and Shakespearean-sounding ad copies, product descriptions and marketing text, or landing pages respectively are generated in seconds, powered by GPT-3.
“There are two ways startups can use GPT-3. You can either sell the direct output of the language model, or you build a product on top of that,” explains Wang, recipient of a Thiel Fellowship, instituted by tech icon Peter Thiel to promote young high-potential startup founders. Besides Wang’s own startups, an example for the former approach is sudowrite.com, which proposes creative writing pieces based on a writing prompt. Examples for the latter are debuild.co, which allows users to build a web page off of a verbal description of the desired outcome, and uncannily human chatbots, such as replica.ai.
While the winners of this new generation of startups have yet to emerge, the bigger question is how the NLP and AI model landscape will evolve, and what precedence GPT-3 may have set for the broader future of AI and businesses. More movements and shifting dynamics can be expected in these next formative years of the AI space.