Spoiler Alert: A lot of us are in Stage 2

A vast majority of the tech world sees this era of AI as the next Industrial Revolution, with its potential to transform the way the world works, how autonomous software and robots will become the mainstay etc. There are also concerns about the magnitude of impact this can have, be it benevolent or otherwise.

In this post, I’m looking at where most, and I repeat, most companies are in their AI adoption journeys and what those look like. The title, and the content derives inspiration from India’s three-stage nuclear program.

Stage 1: General Purpose LLMs

Most companies would have begun here. This is the stage where the use of commercial tools and models in the market are used for both code generation and reasoning. There is a lack of institutional context. The natural resource in this case is the model itself, which has been trained on crude, publicly available data, code, patterns etc. The output of this stage is enhanced productivity for developers who know what they’re doing and want to speed up their code generation by using comments, prompts etc. The focus is still on producing artifacts that are for human consumption, like code, documentation etc.

Stage 2: Specifications and Context

This is where we are as of today. Companies are setting up skills, MCP registries, agentic frameworks etc to enhance the outcomes, yet still haven’t attained self sufficiency. Commercial models are still utilized, but there is an increasing push towards self sufficiency in terms of setting up knowledge and dependency graphs, a mindset shift towards creating artifacts that have reasoning capabilities and knows more about the systems and constraints they’re working with. The benefits are beginning to show and expansion plans happen, and at the same time, there’s also mindfulness of resource utilization, costs, strategy re-evaluation etc. Self sufficiency is beginning to be considered, including commoditization of models, open weights etc.

Stage 3: The Flywheel

The self sustaining cycle in organizations where the outcomes feed back into the AI systems to have a re-generative effect. This is the desired end goal where the agents and the organizations attain self sufficiency by having the best of both worlds. The original technology foundations, and the vastly improved context that the systems have from having seen a lot through the ages (in AI terms, it’s mere weeks or months). Organizations run their own AI data centers, which have become cheaper and has the same pull as private cloud infrastructure. This provides autonomy, data safeguards, true democratization of AI based development.

I’m most excited about Stage 3. While enterprises operate their systems in-house, it’s essential for them to contribute back to the world by sharing knowledge and best practices on