Why we are starting as an applied research lab

Training GPT-4 consumed more than 50 gigawatt-hours of electricity and roughly 13 trillion tokens. That number is not an anomaly — it is what the current paradigm requires. Transformer scaling laws are not a temporary inefficiency. They are structural. More capability means more parameters, more data, more compute, and more energy. The architecture is the problem.

We are not trying to optimize around that problem. We are trying to replace the architecture.

The bet

Phase Space Labs is building quantum diffusion models: generative systems that encode structured visual information as quantum states and learn to denoise them through diffusion-like dynamics in Hilbert space.

The core idea is that classical diffusion models learn structure by memorizing it — you show the model enough examples, at sufficient scale, and it approximates the data manifold. Quantum states do not work that way. A quantum state on n qubits is a unit vector in ℂ^(2^n), constrained by complex phases and entanglement geometry that have no classical analogue. If the manifold of natural images maps onto this structure better than it maps onto a classical Gaussian latent, the model should find a more compact representation — and learn from far less data.

Our preliminary results support this. Quantum diffusion models can generate small images using less than 10% of the training data required by classical diffusion models. We do not fully understand why yet. That is the open research question we are studying.

Why not just work on classical efficiency

Researchers and applied ML teams have responded to the scaling problem by focusing on continued synthetic pretraining, RL environments, and inference-time compute. These are reasonable bets. They work within the existing paradigm.

We think the deeper fix requires replacing the paradigm. Classical diffusion labs — Stability AI, Inception Labs — are building increasingly capable systems using the same fundamental architecture. The in-house alternatives, models like GPT-4 and Claude, are impressive but environmentally costly by design.

The quantum approach is harder. It requires bridging two genuinely difficult problems: quantum gates are discrete and unitary while the reverse SDE requires smooth, continuous score estimation, and there is no established playbook for this. Encoding classical image data onto qubits without information loss, and designing a noise schedule that allows convergence, requires heavy experimentation. We are doing that experimentation.

What applied research means here

We are not an academic lab. We are not optimizing for publication cadence or fitting results into venue-specific formats. We are building real systems, running real experiments, and writing honest accounts of what we find — including failures.

We are also not a startup. We do not have a product roadmap or a go-to-market timeline. We are following the technically interesting direction, which right now is: can quantum-parameterized denoising actually work as a generative mechanism, and if so, why?

The hybrid pipeline we are building looks like this: a classical CNN encoder compresses visual data onto qubit statevectors, a quantum module models the latent distribution through diffusion-like dynamics, and a classical decoder reconstructs the output. The quantum component brings ~84–120 circuit parameters — comparable to a small classical layer. The question is whether those parameters, living on a quantum manifold, outperform an equivalent classical latent at the task of structured generation.

Why now

My background is in building and leading engineering teams at Datafruit (YC S25), where I designed offline reinforcement learning frameworks for agent self-improvement — increasing accuracy on large enterprise knowledge bases from 42% to 74%. Before that, my research spanned MoE architecture optimization with LayerSkip and QuDDPM at Georgia Tech.

The tools for doing serious AI research outside a major lab have never been better. Cloud compute — including quantum simulation infrastructure — is accessible enough for a solo founder to run meaningful experiments. The missing piece is not tools. It is the willingness to work on a problem that does not fit cleanly into either the academic publishing cycle or a commercial product timeline.

That is the space Phase Space Labs is trying to occupy.