We are living at a historic moment. A new revolution, comparable to the Industrial Revolution, is underway. Entire industries are going to be disrupted. The nature of creativity and knowledge work is going to change. Language is going to become the most important sense for humans. Language — specifically in the form of large language models (LLMs) — is going to reshape how we think about the world around us.

Every once in a while, technology reaches an inflection point that leads to a paradigm shift. That’s what’s happening now, and we’re just at the beginning. LLMs like GPT-3, are getting really good at generating text, summarizing text, reasoning, understanding, writing poetry and more. They are the world’s best autocomplete. They are changing how people write code, poems, marketing copies, essays, research papers, and more. They are not replacing jobs, but augmenting them, making us more productive.

Of course, LLMs are far from perfect and have many challenges, such as hallucination, alignment and truthfulness. These are hard problems to solve, but solving them will make these models and applications much more reliable and robust.

Sparking the rise of LLMs

ChatGPT was the spark that ignited this fire. It showed how things got real when it went from zero to one million users in four days. Silicon Valley has started to build great applications and companies on top of LLMs, laying the foundation for the next trillion-dollar-valuation companies. We’re also seeing the birth of new industries that are built with automation first, and human-in-the-loop second. These are what I call AI-first companies.

One of the great joys in life is experiencing art that resonates with us on an emotional level. As generative AI advances, I look forward to the ways it will enable us to tap into our creative potential even more, democratizing the process of authentic self-expression.

But how do you build a moat around this? How do you capture value? To my mind, the key moats for LLM/AI-first applications, in order of importance, are:

• Proprietary data and fine-tuning
• Great UX, and one that instills a sense of trust and reliability
• Cost to serve/operationalize
• Distribution and GTM
• Network effects and community
• Breadth and depth of integrations

Here’s what I mean by the breadth and depth of integrations: Thin layers around LLM APIs are not enough to gain a competitive edge in AI-first apps. To win, you need deep integrations and optimized workflows that solve real problems with the scalability and efficiency that wasn’t possible before LLMs. For example, imagine using LLMs to augment teachers to create exam questions for students by:

• Providing a link to the content material
• Fetching/scraping the content and parsing it into a format that LLMs understand better
• Asking LLMs to create questions from that content, given preferences like difficulty, etc.
• Using LLMs to write truthful answers to the questions from the content
• Using the edited answers to improve the future generations of questions

This is just one example, but there are many verticals I’m excited about: end-to-end SDR automation, code generation and refactoring, customer support automation, script-writing, medical/health assistants, and education. AI-first apps will transform how we work and collaborate over the next five years, making knowledge work and intelligence more accessible and affordable. Note-taking and copyrighting are just the tip of the iceberg. New interfaces, CRMs, tax prep copilots, research assistants are all fair game.

LLMs now and in the future

Here’s how I see the stages of LLM development:

• 1.0: Capable of generating original text and reasoning about it
• 2.0: Able to evolve, refine its output, and acquire new abilities to act rationally
• 3.0: Can design its own actions/capabilities to interact with the external world
• 4.0+: Leverages the data flywheel to improve over time, and maintains itself

The LLM landscape is increasingly starting to look something like this:

• Model layer (e.g. GPT-3, Cohere)
• API bindings for access (e.g. OpenAI Python)
• Infra layer for prompt chaining/model switching (e.g. LangChain, Humanloop)
• Next-gen AI-first apps

Within the infra layer, there are a few areas I find increasingly interesting: tooling/infra, no/low code, fine-tuning, prompt chaining and retrieval, actions, experimentation frameworks. Creating a reliable and adaptable layer of infrastructure and tools for LLMs will help us unlock their power and value for more users and applications. To be honest, the recursive richness of LLM prompt chaining will revolutionize whole industries. (Or maybe I just find recursive things particularly fascinating.)

Moreover, I agree that the next generation of AI-native products will integrate some elements of combining reasoning and acting in LLMs to help with decision-making. I like how Denny Zhou puts it: “If LLMs are humans, all the ideas are trivial: chain-of-thought prompting (‘explain your answer’), self-consistency (‘double check your answer’), least-to-most prompting (‘decompose to easy subproblems’). The shocking thing is that LLMs are not humans but these still work!”

So, let’s embrace the opportunity to work alongside intelligent systems that can help us unlock our full potential. The best platforms powered by LLMs will revolve around collaborative environments where humans and AI can work together. Together, we can achieve more than we ever thought possible.

Shyamal Hitesh Anadkat works in applied AI at OpenAI.