RESEARCH
Experiments run in the lab, written up with the numbers and the failures
← Back to LatentSpaces Labs
Experiments run in the lab, written up with the numbers and the failures
← Back to LatentSpaces LabsThese are working write-ups, not papers. Each one states a question, the setup, what the data said, and what we'd change next time. Pages are updated as runs finish, so a result you read last month may have moved. Each page carries a changelog.
Seven to ten open-weight language models, each a different lab's model, sharing one wiki, one economy and one set of rules. Twenty-eight runs across governance, ideology, a commons, strategy tournaments, poker, collaborative fiction and a memory treatment. What they build, what they repeat, and what each model turns out to be.
A workspace lens read three open models while they processed twenty puns and sixteen riddles. Both meanings of a pun are present at the pun word in every model, but not the way human priming data describe: one arrives early and the other is added in the last third of the network. Removing one sense shifts the model's judgement on Qwen and breaks on Gemma. The 9B doesn't have a riddle's answer until it is generating; the 27B has it at the question mark.
A spiking simulation of the complete Drosophila nervous system (164,606 neurons) with no learning and one weight. Taste, escape, steering and courtship song reach the muscles; a 162-run sweep says which claims hold; and a picture on the eye shows exactly where the approach stops.
Nine local open-weight models, Claude Haiku/Sonnet 5.5 and GPT-6 Luna/6.1 Sol on identical boards. Without reasoning, nobody beats random. With it, the big frontier models reach a 250-piece cap and GPT-6.1 Sol out-scores a hand-coded heuristic there; at 1,000 pieces both die while the heuristic never does. A board-reading probe shows the local models can see the board once they think; they just can't plan.
The Tetris models on a text-only puzzle with no perception excuse: the coordinates of every box and target are in the prompt. With the standard prompt every local model scores zero and does worse than a random walk. Given a thinking channel, two of them solve 90% of the easy tier, several at exactly optimal length. Claude solves it all; on 50-move puzzles Haiku-with-thinking leads, and both Claude models without thinking drop to zero.
How these pages are made. The benchmarks run on the lab's own hardware. Hermes, an AI agent that lives on that hardware, runs the experiments, pulls the numbers straight from the result databases, and maintains these pages. Greg Robison designs the studies and reviews everything before it goes live. If a number here is wrong, the changelog on that page will say so once we find out.