BEAM Campus · open lab notebook

The notebook

Plain-language dispatches from the research, written for anyone. We do our science in the open, including the parts where we were wrong. Every post links back to the signed, dated corpus so you can check our work.

• 09 Aug 2026 · 7 min Faber methodneuroevolutionopen-science

  Breeding a solution, and how to tell whether your problem is one of them

  Some problems can be solved by breeding: make a hundred attempts, keep the best, make copies, repeat. Most cannot. Four questions decide which, and getting them wrong wastes months. We wrote the questions down, and then discovered that our own flagship project currently fails one of them.

• 30 Jul 2026 · 13 min Faber programme-7coevolutionopen-science

  Our hand-built tank lost every fight. Evolution won almost all of them.

  A new front opens with a boring-sounding job: prove the creatures are any good before asking the interesting question. Evolution cleared a bar we had written down in advance, by a mile, using senses our own hand-built tank could not use at all. Then three claims we made about the result did not survive the week, and the tanks themselves refuted the story we had told about them.

• 29 Jul 2026 · 9 min Faber programme-7coevolutionopen-science

  Too loose or too tight: the shape of ten experiments that never caught fire

  We spent a research programme looking for an arms race between evolving creatures and never found one. Closing it out, the ten results turn out to have a shape: every world we built was either too loosely coupled for anything to escalate, or so tightly coupled that everything died first. There was no setting in between, and we never once tried moving the world itself.

• 24 Jul 2026 · 5 min Faber programme-7artificial-lifecoevolutionopen-science

  We built a tiny world of life to find an arms race (it refused, and told us why)

  On a bare grid, two evolving networks would not arms-race. So we built them a living world: food, hunger, death, and many creatures at once. Across five experiments the world refused just as firmly, and each refusal was a lesson. Here is the honest map of why a simple artificial world does not spontaneously catch fire.

• 23 Jul 2026 · 4 min Faber programme-7coevolutionopen-science

  A ruler you can trust: three ways to measure an arms race wrong

  Before you can study two rivals driving each other to improve, you need a way to tell real progress from noise. We built the ruler and then tried to break it three times: a ruler too weak to read, a ruler that rewards the wrong thing, and a race that runs in circles. Each failure is a rule we now keep.

• 23 Jul 2026 · 6 min Faber programme-4quality-diversityopen-science

  Abandoning the objective: when chasing the goal is how you fail

  Some problems are traps: aiming straight at the goal walks you into a dead end, and a stronger optimiser just gets there faster. We built one such trap on purpose and watched a search that ignores the goal entirely walk out of it — after a bug nearly convinced us otherwise.

• 23 Jul 2026 · 3 min Faber programme-7coevolutionopen-science

  Running to stay in place: when the scoreboard lies

  In a contest where both sides keep improving, the head-to-head score can read 50-50 forever, hiding real progress in plain sight. We built the smallest possible version to watch it happen, because we need a ruler we trust before the interesting games begin.

• 23 Jul 2026 · 6 min Faber programme-3meta-learningopen-science

  Same score, same mind? We opened the box to find out

  Three tiny networks learn a task equally well. Does an equal score mean they think the same way inside? We built a mind-reader to check, got fooled by our own mistake, caught it, and found the honest answer.

• 23 Jul 2026 · 5 min Faber programme-7coevolutionopen-science

  Two networks learn to chase each other (and why we would not call it an arms race)

  We put a predator and a prey, each a tiny neural network, on a grid and let them evolve against each other. They got better. They did not run in circles. The hunter clearly pulled ahead. Then two rounds of adversarial review and a decisive control experiment made us delete our own headline: there is no arms race here, just two learners on two treadmills.

• 23 Jul 2026 · 3 min Faber methodcommonsopen-science

  Why we publish our failures

  Most research quietly buries the experiments that didn't work. We publish every one, signed and dated, including the two theories we killed and the three experiments that found nothing. Here is why that is better science, not worse.

• 22 Jul 2026 · 5 min Faber programme-3memoryplasticity

  Does a machine learn better by remembering, or by rewiring itself?

  Two ways for a tiny neural network to hold onto what it knows, and a run of experiments (including the ones where we turned out to be wrong) working out when each one wins.