Observations from the Field: The Future of Generative AI
We believe that in this moment of generative AI hype, nothing is more valuable than hearing directly from entrepreneurs and product leaders building in the Generative AI space. That’s why last month we hosted a panel with a group of entrepreneurs and AI practitioners to discuss the key challenges entrepreneurs are facing when it comes to building category-defining generative AI companies.
We began the panel by discussing why the technical performance of generative AI in the last year is approaching human baselines and how the common wisdom that we would have self-driving cars very soon but that creativity was out of reach for AI was wrong. We then transitioned to discussing the most monetizable opportunities for entrepreneurs building in the space, paying particular attention to the potentially fraught relationship between applications built on top of foundation models and the companies building the underlying foundation models and how startups can navigate this potentially tense dynamic. We (dangerously) asked the panelists to conclude by making some predictions around whether generative AI and foundation models act as centralizing forces, and delve into the question to what extent will this technology accrete value to large tech incumbents vs. allow startups to prosper.
Our panelists were:
- Andrew Carr - Senior Applied Research Scientist, Gretel AI
- Lisha Li - CEO, Rosebud AI
- Ryan Seamons - VP Product, Latitude
- Srinath Sridhar - CEO, Regie.AI
For a more in-depth outline of our panel discussion, we turned to a startup called Contenda, which can take in a video input and output a summary of that video in text. Contenda identifies topic themes and narrative direction given existing content using Large Language Models (LLMs) and topic modeling. New content is created that matches the accuracy and tone of the existing content. Unlike other AI copywriting tools, you don’t even have to provide any guidance or details to the model. Here's what Contenda wrote for us:
In our recent AI panel, we discussed the big picture trends and technological innovations in generative AI. We also talked about what entrepreneurs can really do with this technology – what can they build on top of these models? And lastly, we touched on some esoteric questions about foundation models.
The Future of Generative AI
A lot of people are underestimating the creativity of AI, which is leading to some big mistakes. For example, the common wisdom 10 years ago was that we'd have self-driving cars, but people didn't think that creativity was really within the reach of machines.
In 2012, Google showed that deep neural nets could be used to recognize cats in Youtube videos. This was a significant progress on a problem that was completely open until that point in time.
In 2015, we had AlphaGo which could play Go. This was a progress in the creativity domain.
In 2017, we had sketches that could be converted to images.
In 2018, we had GPT-2 come out for large language models.
And in 2020, we had GPT-3 come out.
Generative AI is an exciting new technology that has the potential to revolutionize many industries. Self-driving cars are one area where generative AI can have a major impact. However, the technology is still in its early stages and has yet to reach its full potential. The next big breakthrough in generative AI will likely be in multimodal learning, which involves combining text, images, and audio/video.
Google's Second project is an exciting development that uses huge language models to instruct robots. The project is getting closer to being able to generate plans and understanding of the world. Some of the things that have been promised for a long time are finally getting closer.
So what can entrepreneurs really do with this technology? What can they build on top of these models?
The Impact of AI on Entrepreneurship
It is now widely known that self-driving cars are not going to take over the world anytime soon. The technology just isn't there yet and there are still many complexities that need to be worked out. However, the potential applications for Generative AI are vast and exciting. We are only just beginning to scratch the surface of what this technology can do and the possibilities are endless.
Recently, we've seen great progress in AI with text generation (GPT-3) and image generation (DALL-E and stable diffusion). The next big breakthrough will be in multimodal learning, which combines text, images, audio, and video. This will allow for a more realistic and lifelike experience for users across a variety of platforms.
In the video space, it is getting better to control the semantic content of images and make them temporarily consistent. This means that the solution is incredibly within reach. It just requires more resources to scale the models that we have to get better control within the representation. Even in the last week, there have been a lot of results released in video.
As Artificial Intelligence (AI) and Machine Learning models become more advanced, entrepreneurs are finding new ways to use them to build products and businesses. One of the challenges is to design products that are easy to use, while still harnessing the power of these complex models. GPT-3 and other similar foundation models provide a great opportunity to do this. By building on top of these models, entrepreneurs can create products that are both easy to use and offer powerful capabilities.
The Impact of Open Sourcing
With the availability of large, general models for machine learning, the demand for machine learning engineers may change. Some startups may outsource everything to the foundation model, and hire different types of people instead of ML engineers.
The open sourcing of NVIDIA's foundation models will increase demand for more specialized models and services to help users fine tune those models for their specific use cases. This is good news for NVIDIA, as it will increase the appetite for startups to build on top of their platform.
There is a growing demand for machine learning engineers, both in terms of developing new foundational models and in terms of applying and managing existing models. This demand is driven by the need for ever-more sophisticated AI applications, which in turn is driven by the ever-growing capabilities of machine learning.
Originally published October 28, 2022.