Paramorph — A foundation model for autonomous neural network optimisation
Trained with multi-agent reinforcement learning, Paramorph adapts how neural networks train — while they train. It's the optimisation layer of Recurvia's self-improvement loop.
The Scale of Frontier AI
Frontier AI now runs at industrial scale — and every inefficiency is priced at that scale. The numbers make the case.
10T
Parameter count for today's frontier models
10x
Forecast growth in AI compute demand
5000+
MWh per day to power a frontier cluster
$10B+
Development cost of a frontier model programme
500K+
Tonnes CO₂e per year from a frontier cluster
10²⁶
FLOPs in a single frontier training run
100K+
H100/B300-class GPUs in a frontier cluster
100PB+
Training data behind frontier models
Compute is the constraint. Autonomous, real-time optimisation is the lever.
As models scale, compute has become the defining constraint in AI — and advances in hardware are outstripped by increases in demand. Every gain now carries industrial costs in capital, energy, and time. The industry's default answer is more hardware: slow to arrive, expensive to run, and rationed at exactly the moment it's needed most. The other lever is to extract more from the compute you already have — and at the scale of the numbers above, even single-digit efficiency gains are worth millions.
The Scaling Problem
At frontier scale, every configuration choice compounds. A run is shaped by interacting decisions — optimiser settings, schedules, batch dynamics, parallelism — most of them fixed before training begins and expensive to revisit once it starts. When they're wrong, the costs are industrial: stalled convergence, wasted epochs, abandoned runs. And the standard remedies don't scale with the models. Sweeps burn compute on runs that exist only to be discarded; expert intuition doesn't transfer cleanly across architectures, workloads, and hardware.
The Paramorph Approach
Paramorph replaces configuration with control. Instead of fixing hyperparameters before a run and hoping, it adjusts them while the run unfolds — many at once, coordinating them as they interact, in response to the training dynamics it reads in real time. Itself a foundation model trained with multi-agent reinforcement learning, Paramorph brings a trained policy to a problem previously handled by search and intuition: specialised agents steering optimisation across layers and timescales as conditions change.
Paramorph is designed to deliver:
- faster convergence, on less compute
- fewer sweeps — and fewer runs that exist only to be discarded
- shorter research cycles
- better final models
Optimisation as Intelligence
Optimisation has always been a craft: defaults, heuristics, and hard-won intuition, re-derived by every team for every new model and every new dataset. Paramorph treats it as intelligence — a skill learned once and carried across architectures, workloads, and hardware, where intuition never transfers. And unlike heuristics, a learned optimiser keeps improving, generation over generation.
The next generation of AI depends on advanced optimisation — and optimisation this complex can only be fathomed by AI.
Beyond human optimisation. Under human authority.
Hyperparameters interact — and those interactions shift as training unfolds. Human intelligence cannot steer that coupled, moving system in real time. Paramorph is an artificial intelligence built from the ground up to do precisely that: adjusting many hyperparameters together, continuously, with a learned understanding of how each affects the rest. That power is usable precisely because it is governed — you set the envelope it operates in, and command stays with you.
.webp)
You set the envelope
Paramorph is built to make thousands of adjustments across a run — every one of them inside an envelope you define: what it may adjust, by how much, and when. Every intervention is logged as it happens, inspectable afterwards, and revertible at any point.
Nothing to rebuild
Paramorph joins the training loop you already run.
- drops into existing training code without refactoring
- works with common ML frameworks and custom pipelines
- native support for layer-wise control
- runs alongside your current tooling
Paramorph is in active development. We are opening a small number of early access places for teams training at serious scale.
The future of AI will be self-optimising.
Paramorph is built for that future.