The landscape of artificial intelligence has been dramatically reshaped, especially when we consider the recent buzz around machine learning diffusion models. For years, I’ve been immersed in the world of machine learning, from my early days at Google and Twitter to founding and investing in ML companies. One thing was consistently clear: until recently, the vast majority of machine learning startups struggled to gain traction. The pre-2020 era, dominated by CNNs, RNNs, and GANs, saw most of the value accrue to established tech giants, not nimble startups. This was largely because the technology simply wasn’t advanced enough to unlock genuinely new market opportunities.
Image alt text: A slide from 2017-2019 illustrating the AI capabilities of CNN, RNN, and GAN models, representing the limitations of the previous machine learning wave.
Many in the business world still perceive today’s “AI” through the lens of these older capabilities. However, this is a fundamental misunderstanding. We are not simply seeing incremental improvements; we are witnessing a step function change in what AI can achieve. It’s akin to comparing the invention of the airplane to the car. An airplane isn’t just a car with wings; it ushered in an entirely new era of possibilities in travel, logistics, and beyond. Similarly, the advancements in AI, particularly driven by diffusion models and transformer architectures, mark the dawn of a new technological age, not just an upgrade to the old one.
It’s crucial to acknowledge the foundational role of earlier machine learning and deep learning waves. They paved the way for these breakthroughs. But to view the current revolution as merely a continuation of the past is to miss the profound, seismic shift that is underway.
Image alt text: A modern slide highlighting the transformative impact of new AI architectures like transformers and diffusion models, showcasing the paradigm shift in AI capabilities.
The first real indication of this shift came with the launch of GPT-3 in June 2020. GPT-3 was a monumental leap forward from GPT-2 and previous models. While it didn’t fully embody all the capabilities we now associate with cutting-edge “AI,” it strongly hinted at the future. For those paying close attention, the release of GPT-3.5 in March 2022 cemented the dominance of transformer-based models as the future. Within companies like Google, OpenAI, Microsoft, and Anthropic, early access to these advanced models offered a glimpse into the coming revolution. This even led a Google engineer to famously declare an internal AI chatbot, LaMDA, as “sentient,” a precursor to today’s sophisticated chatbots like ChatGPT and Character.AI.
[
Image alt text: A visual representation of the timeline of AI advancements, emphasizing the pivotal role of GPT-3 and subsequent models in initiating the current AI wave.
However, the true catalyst for the current AI wave, particularly in terms of startup activity, was ignited by two key events. First, the launch of image-generation powerhouses like Midjourney and Stable Diffusion demonstrated the incredible potential of machine learning diffusion models. These models, capable of creating stunningly realistic images from text prompts, captured the imagination of creators and developers alike. This was quickly followed by the arrival of ChatGPT, which took the world by storm. ChatGPT became the “big bang” moment for AI startups, showcasing the remarkable capabilities of these new AI forms and the power of Reinforcement Learning from Human Feedback (RLHF). OpenAI further solidified its lead with the release of GPT-4 just four months later.
Image alt text: Robert Huben’s slide titled “GPT-4: What we (I) know about it,” signaling the rapid progression and increasing sophistication of large language models.
While the hype and excitement are palpable, it’s important to remember that true widespread enterprise adoption is still on the horizon. ChatGPT’s launch was indeed the starting pistol for mainstream awareness and enthusiasm for generative AI. However, it’s only been roughly 8-9 months since then, and GPT-4 is even newer. Large organizations operate on longer planning cycles, often taking 3-6 months just for initial planning, followed by a year or more for prototyping and implementation. Therefore, we are still quite a distance from peak AI integration and even peak hype. Most large enterprises are still in the analytical phase, trying to understand the implications of “AI” for their businesses, and are several quarters away from fully embracing this transformative technology.
The Four Waves of AI Adoption
To better understand the unfolding AI revolution, we can consider it in terms of four distinct waves of adoption:
Image alt text: A diagram illustrating the four waves of AI adoption, starting with GenAI native companies and progressing to large enterprise adoption in later phases.
Wave 1: GenAI Native Companies. This wave is characterized by companies born directly from the generative AI revolution. Think of ChatGPT, Midjourney, Character.AI, Stable Diffusion, and GitHub Copilot. These companies have already achieved significant user growth and revenue. Of course, established ML companies like Hugging Face, Runway, Scale, and WandB, while predating GenAI, are also key players in this evolving landscape.
Wave 2 (Current Wave): Early Startup Adopters and Fast Mid-Market Incumbents. This wave includes startups built on top of models like GPT-3.5/4, such as Perplexity, Langchain, and Harvey. We also see forward-thinking, founder-led companies like Navan, Notion, Quora, Replit, and Zapier rapidly integrating AI into their products. Large enterprises like Microsoft, Adobe, and Google are notable exceptions, moving surprisingly quickly in the AI space. Microsoft’s close partnership with OpenAI likely fuels its speed, while Adobe benefits from diffusion models being relatively more accessible and cost-effective to train compared to massive LLMs.
Wave 3 (Coming Soon): The Next Generation of Startups. The pipeline is brimming with new AI startups. This wave is likely to explore new formats like voice and video, expand natural language AI into more industries and applications, and develop novel AI infrastructure. Companies like ElevenLabs, LMNT, LFG Labs, and Braintrust are examples of those pushing the boundaries. The current Y Combinator batch alone reportedly includes over 100 AI startups, signaling a massive influx of innovation.
Wave 4 (Coming 2024/2025?): Large Enterprise Adoption. Given the lengthy planning and implementation cycles of large enterprises, we anticipate seeing the first significant AI-powered products (beyond demos and prototypes) from these companies in the next year or two, excluding the early movers like Microsoft, Adobe, and Google. This wave will drive substantial revenue growth for AI infrastructure companies, likely coincide with peak AI hype, and trigger even greater investment in AI technologies.
The Future is Brimming with Potential
The transformative potential of this new wave of AI is immense. Consider Google’s MedPaLM2 model, which has demonstrated performance exceeding that of human physicians to such an extent that human feedback during RLHF actually degrades its performance!
[
Image alt text: A visual representation of the superior performance of Google’s MedPaLM2 model compared to human physicians, highlighting the advanced capabilities of AI in specialized domains.
The coming years promise incredible innovation across education, healthcare, enterprise and consumer software, and countless other facets of life, all powered by these technological breakthroughs. We are still in the very early innings of this AI revolution, merely 8-9 months removed from ChatGPT awakening the world to this new era. The most exciting developments, and the peak impact of AI, are still ahead. The journey has just begun.
