Christos Tzamos and his team at Percepta have published a research post making a pointed argument: chain-of-thought prompting is the wrong abstraction, and transformers can do better by executing programs directly within their architecture. The punchline is an exponential speedup on structured tasks — a claim that, if it holds, would undercut the cost basis of most current agentic pipelines.
The mechanism is specific. Rather than training a model to approximate the output of an algorithm through thousands of autoregressive tokens, Tzamos maps each transformer forward pass to a single step in a program execution. Token generation becomes the clock cycle. The transformer isn't predicting what a program would output — it's running the program. For deterministic tasks, this sidesteps the probabilistic drift that makes current reasoning chains unreliable and expensive.
Tzamos's background is in computational complexity theory — the branch of CS concerned with what problems are solvable and at what cost. That framing shapes how Percepta approaches the Turing-completeness question that has been circulating in transformer research for years. Most prior work on the topic stays theoretical. Percepta's pitch is that the practical payoff is inference efficiency: if you can encode symbolic execution inside a model, you don't need a 64-step chain-of-thought to add two numbers.
For agent developers, the stakes are concrete. Agentic systems today chain LLM calls together with tool use and memory retrieval, each step burning tokens and introducing failure points. A model that handles symbolic computation internally rather than through external scaffolding would reduce latency, cut cost, and improve determinism simultaneously. That's not an incremental improvement — it changes where the reliability budget gets spent.
Percepta is a small outfit betting on computational theory as a differentiator in a market where most players compete on developer experience and integration count. The exponential inference claim is bold enough to warrant skepticism — Tzamos hasn't published benchmark numbers, and the gap between theoretical speedup and production workload performance is rarely small. But the underlying research program is more rigorous than most of what passes for AI infrastructure innovation, and the direction is coherent enough to take seriously.