There's a lot happening in AI. Scientific discovery, improvements in reasoning, how models are trained, and the applications being built around them could each fill their own article.
The part I'm interested in here is how we communicate what we want, how AI turns that into action, and what that interaction might look like in the future.
Looking back at Google's "Attention Is All You Need" paper from 2017, it's interesting to see just how much followed from it. Conversational AI was already a thing, but the Transformer architecture changed the trajectory. BERT helped push language understanding forward. Semantic search made finding relevant information easier. RAG connected retrieval to generative models, allowing them to produce answers using retrieved material. Different pieces of a much bigger picture.
Then GPT-3 came out in 2020. To me, that was when the potential really became obvious. It was revolutionary, and most people outside the industry didn't know much about it yet.
ChatGPT only came out two and a half years later, running on the GPT-3.5 series. Instruction tuning and human feedback helped turn those language capabilities into something people could actually have a useful conversation with. It quickly became what was described at the time as the fastest-growing consumer app in history. (Reuters, February 2023)
Those who were already using OpenAI's APIs and incorporating these models into applications had a head start on seeing where this could go.
One of the clearest signs was when OpenAI added function calling to its API in 2023. A model could interpret a human's request and turn it into a structured instruction for software to execute. Tool use became a bridge between conversation and action.
And looking at what they've been building, I think the AI companies had a much clearer idea of where they were heading than the early products let on. The pieces are starting to fit together. An assistant that understands what you mean, has the relevant context, and can get things done across the systems around you. Something a little closer to the "Jarvis" we'd joke about having one day.
Natural language was becoming a practical interface to machines. And the models kept getting better through more training, better post-training, and more computation spent working through a problem.
So what else were AI companies pursuing? AI interfacing with systems: apps, databases, computers, other AIs.
And here we are today:
MCP: a standard protocol for connecting AI applications to a system's tools and data.
A2A: a protocol for agents to communicate, coordinate, and delegate work.
Computer use and CLI access: agents interacting with software through screens, keyboards, and terminal commands, within the permissions of their environment.
Which leads us to the question... What's next?
Here's my guess:
