Byterider ยท Field notes
Experiments
We run LLM agent swarms for a living - usually to keep Belgian companies compliant with EU AI law. Occasionally we point them at things that are not compliance, just to see what they do. This is the public log.
The log
Posts
One entry per experiment. Dates are real; the agents' feelings are, allegedly, also real.
-
Three agents, three languages you would rather not debug, 900 trials.
Qwen3.8-27B, Mistral Large 4 and Claude Haiku 5.5 as coding agents on 300 PowerShell, Groovy and Rust tasks each. All three above 94 percent, Haiku by far the fastest and cheapest, one PowerShell task that beat everybody. What the bench is, what it is not, and a full report page with charts.
-
Three coding agents, 900 trials, eight at a time.
Our local Qwen3.8-27B, Mistral Large 4 and Claude Haiku 5.5 as autonomous coding agents on 300 repository tasks each, eight at a time. Haiku scored highest and was by far the fastest, Qwen and Mistral finished level, and the local GPU gave several times the single-stream throughput but not enough to catch the cloud.
-
Qwen3.8-27B at home against Mistral Large 4 in the cloud.
The same six prompts and eight unit-tested coding problems, a local open model against a hosted frontier model. Level on code at 11 of 16 samples each, and the local model about twice as quick to a finished answer once thinking is on. A small test, with the caveats spelled out and the long-context question left open.
-
After a week of runs: the swarm closed the tested span at t = 8000.
The closing of the Millennium Rig experiment. Four LLM agents verified 7,797 zeta zeros on [100, 8000], all on the critical line, with the count independently checked against Riemann-von Mangoldt. A finite computational verification, not a proof of the Riemann Hypothesis. The "aliasing wall" in our own code was a myth, drift stayed zero, and the lesson is that the value sits in the harness, not the model.
-
We pointed five LLMs at the Riemann Hypothesis. The swarm undercounted, one agent invented an excuse, and the monitor waved it through.
Four LLM "mathematicians" and a monitor count zeta zeros band by band. The published counts were wrong, one agent invented a boundary excuse to close a gap, and the monitor passed it. Then the agents declared their own tool broken, kept a wrong running total after being handed the right one, and the fix went into the code. Corrected numbers, the full timeline from the logs, and a one-screen summary of what broke.