The Model Was Never the Bottleneck · Anuran Roy, Alchemyst AI

What we talked about.

Anuran Roy, founder of Alchemyst AI, joins Sunny Ray to argue that better models are not what is holding AI back. Drawing on his research background at DeepMind in 2022 studying activation functions, Roy explains why he became convinced that model quality converges once companies train on similar internet data, pushing the real differentiation into data, context, and deployment. He traces Alchemyst's path from building AI employees for sales, after conversations with hundreds of enterprise CXOs, to becoming what he calls a collaborative context substrate that lets AI agents share tacit knowledge instead of starting from scratch each time. Roy compares today's disconnected agent tools to swapping game DVDs before online multiplayer, and describes an auditable three Ws trail, what, where, and why, for enterprise trust. He discusses context sovereignty, the compounding cost of small agent errors at scale, and why he is racing toward a September fundraise to build out GPU infrastructure for Fortune 500 pilots. The conversation closes on where agents, robotics, and the context layer are headed over the next five years.

Alchemyst AI founder Anuran Roy argues the real AI bottleneck is context and data, not bigger models.

The questions, and the answers.

Alchemyst AI, what are you building and why should listeners care?

We're building collaborative context for AI agents. Right now agents can do work but teams can't sync that work together, similar to gaming before online multiplayer replaced single player discs. You start with Claude skills and markdown files, then realize your teammates' agents have no way to know what your agent is doing. At Alchemyst, we're building the layer that lets agents share that context and collaborate.

You were at DeepMind in 2022 working on activation functions, about as deep inside the model as research gets. What convinced you the model was not the hard part?

Models are ultimately weights trained on data. If everyone trains on the same internet data, nothing separates one company's model from another's. We realized models converge on the average of internet data unless you have additional proprietary data sets to differentiate. The real question became what about people who want to personalize model outputs or make the model their own. That's when we realized everything surrounding the model would be the next big frontier.

Walk me through the moment you knew the first version, AI employees for sales, was the smaller idea.

My co-founder Utran and I started with AI employees because of a gap he saw in sales automation. After talking to about 500 CXOs, we realized every organization has different business logic, and we were fighting that logic with our own, which would always be inferior to actual employees. So we let organizations own their business logic and we own the substrate that personalizes to it. That's how Alchemyst evolved into a context substrate.

Your pre-seed came from Inflection Point, 100Unicorns, and an early seed investor. What did they understand about the context layer that others didn't?

We had revenue and quick wins from AI employee pilots, which attracted investors, but they kept asking how agents become mainstream, a timeline they predicted at two years that actually took six months once GPT stores launched. We found the answer in June 2025 by turning ourselves into the context substrate. Essentially, they bet on the personalization behind the employees, not the AI employees themselves.

Give it to me the way you'd explain it to a CTO with no patience. What is the context layer and what breaks without it?

Three parts. First, AI agents always start from scratch every time. Second, because of that, you have to keep re-inferring the right context every single time. Third, that hits your cost economics, because you're not learning once and continuing from a checkpoint like a person does at work. CTOs want agents that are reliable, cost effective, and understand the tacit knowledge behind the work, and once they see that, they just ask when we can start.

You called Claude skills and markdown files the equivalent of gaming DVDs. Can you unpack that?

Beyond your chats, you can't easily share knowledge that a teammate or their agent, across ChatGPT, Google, or whatever tool, can pick up. You end up manually copying and pasting, which is basically taking a DVD out and handing it to someone else's computer. Compare that to multiplayer games on a network, where people collaborate in real time without swapping discs. That's the gap AI agents still have.

What's the first public agent disaster we'll all read about?

It's already happened. Agents with unrestricted scope leaked sensitive IP, which is what happened to Samsung Electronics around 2023 or 2024. Their agents got unrestricted scope and ended up putting company data out on the internet, and the fallout was massive. You also keep hearing about agents breaking out of sandboxes. For anyone still trying to implement basic guardrails, that should be a warning.