Model & architecture design
Variants of attention, routing, depth and connectivity, proposed against a fixed objective and pruned early when the cost of a full run is unlikely to be repaid.
Held-out objectivesEarly pruning
Applications
One loop, pointed at progressively harder problems
Autonomous experimentation is only useful where a cycle exists and an outcome can be measured. We start in the domain that satisfies both conditions most ruthlessly — machine learning — and expand outward in sequence, not in parallel.
01 · Starting domain
Search
Training
Systems
Machine learning research, automated end to end
ML is our first domain because the feedback loop is short, honest and cheap enough to run at volume: either an experiment improves a measurable objective or it does not, and we can tell within hours rather than seasons.
Recipes & data mixes
Optimiser settings, schedules, precision and data composition searched together rather than one variable at a time — the interactions are where the gains hide.
Data mixPrecision
Infrastructure tuning
Throughput-first optimisation of kernels, memory pressure, parallelism strategy and scheduling — measured in time and euros per finished experiment.
ThroughputCost per trial
Nothing here is exotic in isolation. The difference is that all of it runs as one closed cycle, thousands of trials at a time, with the results feeding the next round automatically.
02 · Second front
Then the engineering around the research
Research gains are worthless if they cannot be shipped. The second front carries the same loop from model work into the systems, pipelines and products built on top of it — which is also where an autonomous lab stops needing a hand-off to production.
- Evaluation harnesses and tooling improved by the loop rather than by hand.
- Data pipelines, curation and versioning optimised against measurable quality and cost.
- Product surfaces where a result becomes something a team can act on the same day.
Why this order
Acting as our own first customer
Machine learning is where our own stack lives. Optimising it with the loop means every improvement to the system is spent immediately on improving the system — a compounding advantage that is hard to buy and impossible to fake.
Compounding
Own the stack
Measured in production
See the roadmap
03 · Grand challenges
Problems where the loop shape is the obstacle
Each domain below shares a familiar structure: a slow manual cycle, expensive iterations, and an outcome that can be measured. We describe our intended approach honestly — including what has to exist first, and what we do not have yet.
Medicine & health
Long horizon- Cycle today
- Hypothesis, assay design, laboratory work, analysis — one turn per several weeks, gated by human attention and bench time.
- Loop we would run
- Parallel screening across candidate designs and assay conditions, with simulation and prior literature used to prune before the bench.
- Measured by
- Hit rate per unit of cost, time from question to result, and reproducibility of the endpoints used to score.
- Needs first
- Wet-lab interfaces, validated assays, and data handling that respects clinical and ethical constraints.
Clean energy
Materials & grids- Cycle today
- Synthesis and cell testing run largely sequentially, so a promising chemistry can take a year to rule in or out.
- Loop we would run
- Broad search over chemistries, electrode formulations and operating conditions, with simulation-guided pruning ahead of synthesis.
- Measured by
- Capacity retention and cycle life, energy and cost per unit stored, and yield from synthesis.
- Needs first
- Simulators that correlate with physical measurement, and partners with synthesis and test capacity.
Clean water
Applied science- Cycle today
- Membranes, adsorbents and treatment conditions tested one contaminant at a time under laboratory conditions.
- Loop we would run
- Parallel screening across contaminants, water chemistries and operating windows, ranked by efficiency rather than by novelty.
- Measured by
- Removal efficiency, flux, energy per cubic metre and material lifetime under real fouling conditions.
- Needs first
- Standardised water-chemistry datasets and pilot-scale validation beyond the bench.
Cybersecurity
Controlled only- Cycle today
- Manual testing and rule authoring, where coverage is bounded by the hours a team can spend.
- Loop we would run
- Adversarial search inside controlled environments: generate a candidate, validate it against a target we own, rank by what is reproducible.
- Measured by
- Reproducible exploitability rather than theoretical severity, plus time-to-detection after a fix.
- Needs first
- Owned test ranges, clear legal frameworks, and a hard rule: no operation against systems we do not control.
Tools of discovery
Compounding- Cycle today
- Instruments, simulators and analysis pipelines evolve over years, and each improvement restarts the calibration work for everyone downstream.
- Loop we would run
- Co-optimise measurement and analysis: instrument parameters, simulation fidelity and evaluation methods improved together against a single objective.
- Measured by
- Sensitivity and resolution, time-from-observation-to-answer, and how much of a result survives independent replication.
- Needs first
- Access to hardware iteration loops and physics simulation that can stand in for measurement.
Better tools raise the ceiling for every domain above
Sequencing, not promises. These domains are directions, not deliverables, and
nothing here is a claim of results we have not produced. We will say clearly when a domain moves
from plan to active work.
04 · Selection
Five conditions before we touch a domain
Enthusiasm is not a selection criterion. A domain has to pass all five tests before the loop is pointed at it — otherwise we would be automating noise.
- A measurable objective. Something a machine can score without a committee in the room.
- Expensive iteration. If a cycle is already fast and cheap, automation buys little.
- Parallelisable trials. Real concurrency, not a queue that only looks concurrent.
- A safety envelope. We can bound the blast radius and gate every external effect.
- Access to measurement. Data, simulators, instruments or partners who have them.
Domain scorecard
TemplateObjective is measurablerequired
Trials run in parallelrequired
Safety envelope definedrequired
Measurement access securedopen
05 · Collaboration
What we bring, what we need to borrow
We bring
- An experimental loop that runs continuously, not a one-off study.
- Evaluation harnesses that make results inspectable and reproducible.
- Compute orchestration for thousands of concurrent trials with enforced budgets.
We need
- Access to real measurement: instruments, simulators, assays or curated data.
- A domain expert who can define the objective and challenge our scoring.
- Permission to publish method and results, or a clear agreement about what stays closed.
Have a cycle that is slow by construction?
Describe the objective, the cost of one iteration, and why it cannot be parallelised by hand. That is the whole conversation.