Science is changing shape
As agents become capable collaborators, the way research gets done is opening up: new methods, new interface
layers, new institutional forms. The questions that follow are practical ones. How ideas get generated, how
decisions get made, how experiments get orchestrated, and how an institute learns to think alongside machine
collaborators.
The residency is for people who want to work inside that change while it is happening. You learn new research
rhythms, build habits of human-agent collaboration, and help design the practices, interfaces, and environments
that post-AGI science will actually run on. It is a collaboration cycle inside an operating institute, with
shared reviews, working notebooks, simulations, public experiments, and long build loops where fellows
contribute to real NightCity work while sharpening their own direction.
How the cycle runs
Fellows join the institute's operating rhythm: structured research reviews, shared analysis sessions, and
long-running notebooks that humans and agents keep up together. The residency works by doing, which means
entering an active research environment and moving through short build loops with the team.
The work tends to move through three phases. First, rapid synthesis, where agents map the literature, compress
what is already known, and surface the gaps. Then experiment and simulation, where you prototype quickly and
test the assumptions that matter. Then consolidation, where results turn into a draft and a working demo. By the
end you have finished something real and come away with a feel for how science can be practiced when agents take
part as researchers, operators, and collaborators.
Three tracks
Pick a primary track. Most fellows touch more than one as the work develops.
NCL Projects · Research and simulations
Join an existing NightCity Labs direction and push it forward.
Current directions include:
- · Uncertainty in deep learning: calibration, adaptive regularisation, and geometry-aware Bayesian posteriors
- · Machine learning theory: learning time, finite-data limits, geometry, and nonlinear dynamics
- · Motor control and embodied RL: world models, cerebellar-inspired controllers, and robust adaptation
- · Neural plasticity and representation across cortex and motor circuits
Links: Research /
Simulations
Agentic & Simulation Stack · The infrastructure layer
Build the substrate that makes the discovery loop run.
- · Agents that carry calibrated, typed uncertainty through what they know, do, and expect
- · A generative knowledge engine that tracks how well each result is supported by evidence, simulation, or argument
- · Cognitive workflows: ideation and critique loops that run within a single agent or across several
- · Synthetic scientific benchmarks where the hidden mechanism is known and the agent has to recover it
Link:
Technology
Futures of Science · New ways of doing science
Agents can already read the whole literature, propose hypotheses, and run experiments faster than people can
check them. That changes what is scarce. Generating results gets cheap, and the work that remains is deciding
which questions matter, which findings to believe, and what counts as understanding when a machine produced
the answer.
This track works on those questions. It looks at the new ways of doing science that open up once discovery is
fast and cheap, and it works toward a theory of generative science: an account of how machine-driven discovery
works, what it can and cannot do, and how human understanding keeps up. The work ranges from essays on what is
already shifting to early attempts at that theory.
What you leave with
Each cycle produces two things: a draft and a demo. The draft is a manuscript, technical report, or essay
prepared for submission. The demo is a repo, simulation, interface, benchmark, or deployment slice.
Work is collaborative by default, and co-authorship is standard when a result is shared across the fellow and
institute workstream.
Remote and Lisbon
The residency runs remote across time zones.
When it helps the work, fellows can also do focused in-person sprints in Lisbon: deep-work weeks, build jams,
and demo consolidation.
Publication and release
Projects vary in what they can share and when. Release and publication plans are set per project, balancing
openness, rigour, and partner constraints.
