Interweaving AI in our current society
Generative AI, its history, promise and implications on our lives and work
How we got here
Approximately 220 to 250 years ago, depending on which historian you subscribe to, a pivotal technological and architectural shift occurred, propelling humanity into a new era of productivity and economic growth, the Industrial Revolution. This transformation was underpinned by our capacity to produce sufficient food, which in turn freed up labor for creative and entrepreneurial endeavors. Couple that with a somewhat business-friendly political and legal system as well as capital to invest and we were off to our races. The end result was a productivity boost that increased the economic output of many countries around the world and pushed humanity into a new era.
Almost two centuries later, our advances in manufacturing, energy and transportation, from steam to steel to electricity to oil and mass production, set the stage for another important revolution. This time, innovation occurred on a microscopic scale, using thin substrates of germanium and later silicon. The integrated circuit complemented and expanded upon the technological innovations since the Industrial Revolution making machines digitally controllable and interconnected. This facilitated the work of blue-collar employees while also creating new white-collar opportunities.
For more than fifty years, we have refined our digital machines, miniaturized our processors to many times smaller than a human hair, and connected the world through Wi-Fi signals and transcontinental fiber-optic cables. This work has laid the foundation for the third transmutation of our manufactured world, a place where machines are not only interconnected but intelligent and capable of making decisions based on previous experiences. Martin Ford’s “The Rise of the Robots” offers a compelling depiction of this emerging system and the socio-economic implications of the transition.
While many people, myself included, have believed that AI and robotics innovations will upend the way we produce and transport physical goods first, leading to a more automated world, recent innovations in consumer-grade AI, specifically Generative AI paint a different picture. In this sketch, white-collar jobs including creative ones are the most at risk as technology ushers us in a new paradigm of knowledge work.
Analogous to how the steam engine transformed industries in the 18th century and the digital revolution redefined communication and information access, Generative AI is set to revolutionize the way we interact with technology, create content, and tackle complex problems. This groundbreaking shift promises to unlock unparalleled advancements and opportunities, ultimately shaping the future of human progress in the digital age.
Where we are
I am not going to spend a lot of time on what Generative AI is since there are countless resources out there for all levels of proficiency, but I would like to speak about the current state of the industry. As Generative AI continues to advance, we are witnessing the early stages of a technology stack emerging in the space. Hundreds of startups are entering the market to develop foundation models, create full-stack or feature-first AI-native applications, and establish infrastructure and tooling.
Generative AI has seen real gains and unprecedented traction, with models like Stable Diffusion and ChatGPT setting records for user growth and several applications reaching $100 million of annualized revenue within a year of launch. Statista, in the graph below, shows how dramatic the user adoption was compared to most other pieces of software the world has seen:
ChatGPT, which was built by OpenAI, the San Francisco AI company, is the best artificial intelligence chatbot ever released to the general public and is now being used by over 100 million people around the world on a monthly basis. From a technological perspective, the pace of progress is only accelerating with GPT-4, the most popular large language model being launched by OpenAI just a few months after their release of GPT-3.5-based ChatGPT. Midjourney v5, the image-generation model, followed a similar trajectory with an equally impressive quality jump between models.
To analyze the market, A16Z proposes a framework dividing the stack into three discrete layers:
- Applications: These user-facing products integrate generative AI models, either through their own model pipelines or third-party APIs.
- Models: These power AI products and are available as proprietary APIs or open-source checkpoints, requiring hosting solutions.
- Infrastructure: Cloud platforms and hardware manufacturers that run training and inference workloads for generative AI models.
While Generative AI applications have seen staggering growth, challenges remain in terms of retention, differentiation, and profitability. Gross margins are wide-ranging, and it is unclear whether current customer acquisition strategies will remain scalable. Many applications lack differentiation due to their reliance on similar underlying AI models, making it difficult to create long-lasting software companies solely through end-user applications.
The models have demonstrated human-level performance on numerous professional and academic benchmarks. For instance, GPT-4 has achieved a top 10% score on a simulated bar exam, a significant improvement from GPT-3.5’s bottom 10% score. This leap in performance can be attributed to iterative alignment using lessons from OpenAI’s adversarial testing program and ChatGPT, as well as a complete overhaul of their deep learning stack. GPT-4 can also accept both text and image inputs, allowing users to specify tasks involving vision or language, and demonstrating similar capabilities in both domains.
In order to make this more clear, I ran a quick experiment: I wanted to create an image for this blog post that encompasses the history and the future of generative AI. I came up with a few ideas and ran them by Stable Diffusion but the results were not what I had imagined. I then decided to ask for help via chatGPT. Below is the prompt used with GPT-4 in order to get the right prompt for Stable Diffusion.
With off-the-shelf models becoming more and more sophisticated, the barrier of building applications on top of them becomes increasingly smaller. OpenAI has made significant progress in improving GPT-4’s steerability, enabling developers and users to customize the model’s style and tasks through “system” messages. Although the adherence to specified bounds is not yet perfect, this enhanced steerability provides a more flexible AI experience and allows for custom applications to be developed faster than ever.
As a result, we see a surge in application-specific startups across search (personalized, research-driven, multi-app, multimodal, etc.), creation (visual art, writing, video, audio, game assets, avatars), recordkeeping (notes, tasks, meetings, summaries), education (personalized learning, test prep, practice generation), coding (assisted, testing, documentation, synthetic data), relationships (virtual chatbots, dating optimization, conversation assistance) and future of work (scheduling, booking, copywriting, research, customer support, SEO optimization and more recently finance, accounting, and sales). Sequoia has a good foundational market mapping system and Base10 has a great market map.
On the corporate side, Big Techs led the foundational research required for this new age of AI to arise but lagged behind on consumer applications. One of the primary reasons is that both Microsoft and Google operate search engines which create an opposing incentive between redirecting traffic to websites (their primary revenue source) and computing the answer in their own interface. Microsoft led the pack by investing billions of dollars in OpenAI and integrating their technology across the Microsoft Office suite as well as their Bing search engine (more here). Google responded by rolling out Bard AI to a select few in the US and UK in preparation for a world-wide release. Chinese tech behemoths, Alibaba, Tencent, and Baidu, have also caught onto the AI chatbot trend as well, with a plan to develop ChatGPT rivals trained on their own proprietary models.
Given the current rush, it is very hard to say that some have emerged as market leaders. For measurement, there were over 50 companies building in the space during the Winter 2023 batch of Y Combinator (leading startup accelerator program) alone. Only time will tell which ones are here to stay and which ones just bought into the trend.
Where we are going
2023 will be a pivotal year for the industry. The speed of change is only accelerating with recent research proving that ML compute power is doubling roughly every 6 months making Moore’s law look slow and antiquated. Not only will we see more advancements in hardware and models, but we will get the first broad technology adoption test. This will set the tone for what parts of the economy will be the first to fully embrace cognition offloading from humans to machines and design decision making processes to accommodate the new reality.
What will drive the differentiation at the application layer will be a combination of great UX/UI, user education, personalization, verticalization, proprietary data and a strong brand/community. The products requiring the lowest behavioral shift to adopt and capturing most of the user’s attention and time will succeed over the hundreds of replicas that will be built seemingly overnight.
As a result, we will start seeing market-leader behavior across several verticals as value continues to accrue on the backs of data moats (if they form at all). Whether consumer or enterprise will be the first to see mass adoption is irrelevant as one will rapidly influence the other. What is probably more important is the type of workflows that will be automated first.
Until not long ago, interaction labor such as customer service, has profited least from the technological advancements given the heavy reliance on humans. Generative AI may be one of the first inventions to revolutionize the space in a way that mimics human behavior so well that it will soon be imperceptible. Most likely we will see an augmentation of human workflows rather than a complete replacement but that will be just a matter of time. Everything from marketing and sales to operations and HR will see a similar trend. McKinsey recently published an article on the potential impact on business functions.
Last but not least, the one area gathering a lot of attention and with probably the deepest socio-cultural implications on our society is the creative work done by Generative AI. What is art and how will human artists compete with their digital counterparts is still yet to be decided. So are the boundaries describing legal, ethical and fair use of such technologies. Copyright and IP lawyers will certainly be busy over the next few years… until they too, get their Generative AI treatment.
Needless to say we are still in the early stages of this new era of humanity and there are as many unknowns as there are facts. What is certain is that the next decade will be marked by a technological acceleration like we have never seen before capable of creating a market “somewhere between all software and all human endeavors” in the words of a16z. As we progress towards the valley of technological utopia where all our basic needs are automated, we must dedicate significant resources to the sociological, political, cultural, philosophical, and economical aspects of our society that must keep pace with our never ending progress.
Cited Work
https://www.sequoiacap.com/article/generative-ai-a-creative-new-world/
https://a16z.com/2023/01/19/who-owns-the-generative-ai-platform/
https://a16z.com/2023/02/07/everyday-ai-consumer/
https://base10.vc/post/generative-ai-mission-critical/
https://hbr.org/2022/11/how-generative-ai-is-changing-creative-work
