Why not let the data do the talking?
OpenAI’s generative AI model, ChatGPT, reached 1 million users in 5 days, setting historical records for user acquisition, ahead of the iPhone at 74 days, Instagram at 2.5 months, Facebook at 10 months, and Netflix at 3.5 years.
At least three generative AI product categories have already exceeded $100 million of annualized revenue less than a year after launch: image generation, copywriting, and code writing.
And Georges Doribot, our very own AI team member and proprietary research tool, recently uncovered that AI is one of the top three startup categories experiencing the most accelerated growth over the past decade–up 85% over the past ten years and likely to see a big boom in 2023.
The data signals that a seismic transformation is taking place.
Our team spends day in, day out at the fast-changing front lines of data, computing, and human ingenuity, supporting founders doing things with data that would have been impossible only a few years ago. Like giving robots eagle vision. Mining fungi DNA for therapeutics. Or protecting smart energy grids from cyber attack.
We know that AI has reached a tipping point and expect to see many more startups rush to market, but what we don’t know is: which business models will win out?
Here, we share insight from our partners on the AI startup opportunities they are most bullish about, and why.
Large Language Models are the new cloud. Proprietary data is no longer a moat. – Villi Iltchev
History seems to be repeating itself. Cloud infrastructure vendors leveled the playing field by making infinite compute resources broadly available. Compute became commoditized and developers quickly moved up the stack building innovative applications and experiences for customers.
Large Language Models (LLMs) will likely follow the same path as cloud computing. Vendors like OpenAI, Google, and others will level the playing field by providing broad access to robust AI models that will empower developers to build incredibly powerful and innovative next gen applications. Depth of machine learning (ML) expertise or access to data will no longer limit developers from creating new applications and competing with better resourced incumbents. Artificial intelligence will become commoditized and developers will once again differentiate on user experience, workflow, and domain expertise.
The emergence of LLMs will create a new frontier for developers and unleash a new generation of applications. Incumbents who have invested significant resources into building their own machine learning models built around their proprietary data sets could quickly find themselves disadvantaged when competing with startups leveraging LLM vendors. If incumbents don’t adapt to also integrate LLM vendors into their models, proprietary ML models could likely become a liability.
A new user plane for the internet is about to emerge…. starting with a once-in-a-generation shift away from traditional search engines. – Frances Schwiep
We’re witnessing the end of an era of traditional search engines, now that the same data can be served up and manipulated in many different and more compelling formats.
The prompt-answer format is a melting iceberg, and the dominant keywords-driven business model is likely to lose its lead. Next-generation search will emerge with new interaction modalities that move beyond basic prompt-answer formats, and combine the best of multimodal UI and data–layering text, image, voice, video, etc.
Many users will prefer to start their search with ChatGPT, where they can receive a response straight away, in a few succinct paragraphs, and then easily iterate on that response (given the contextual understanding) rather than clicking and combing through many pages in search of an answer. Other assistive AI, like voice AI, are also capturing market share by building towards conversational modes of engagement.
However, I believe it’s only a matter of time before these search methods become disrupted too.
Now that the promise of access to data is no longer enough, the most engaging user interface–which has yet to be developed–will win out.
Multimodal datasets–spanning genomics, image and text–will yield breakthrough insights. – Dusan Perovic
It is still early innings, but the ability to synthesize combinations of datasets across genomics, image and text sources is inciting scientific breakthroughs. Deep learning systems such as AlphaFold have already demonstrated how protein amino acid sequences can be converted into images of their 3D structures, marking significant strides for drug discovery.
If this powerful, accelerated trajectory continues for deep learning systems, and if algorithms can ultimately predict molecules that can bind and block cancer or other disease-driving proteins, multimodal datasets will transform the world of medicine at unprecedented scale.
Here comes the next step-function increase in financial market efficiency. – Colin Beirne
Various forms of advanced data analysis and software tools have long been used by seasoned professional analysts to build trading models, with innovation cycles that improve those models gradually over years.
The significant recent advances in large language models (LLMs) show that AI is now one of the fastest iterating learning machines, enabling new data to be absorbed quickly and new insights to be discovered at record speed.
We expect many of the ‘robo-advisors’ of the past generation will move to quickly incorporate models that use AI to give retail investors a leg up in sophistication, and undoubtedly new startups will be formed at this intersection.
All of it should add up to the next big increase in financial market efficiency, which would help global investors and companies, and may also herald a new era of applying data science to less liquid markets.
Not just one bot to rule them all. – Dan Abelon
Instead of everyone chatting with a bot powered by a single model (ChatGPT), I predict we will see the emergence of many fine-tuned models, particularly in the world of work.
These will vary greatly–from models tailor-made for each individual, trained on their own writing style, to a model trained on all the documents in a company so that employees can surface any kernel of insight at a moment’s notice.
The race is on for which companies will serve up these solutions, and the implications of this will be far-reaching, enabling greater productivity and creativity among non-bots everywhere.
If you’re a founder building an AI-enabled business, we want to hear from you: laurenx@twosigmaventures.com.