I’m building Tenth Man Labs. We put probabilities on geopolitical events, with the sources behind each number and a dated record of how it moved.

My background is GPU systems. Lately most of my work is evaluation and reinforcement learning.

Start here

• attention-numerics: when rotating queries and keys before FP8 rounding makes attention worse. With the real FlashAttention-3 FP8 kernel on an H100, rotation cuts Qwen2.5-7B’s HellaSwag accuracy from 80.9% to 45.5%; centering the keys first removes every measured loss of two points or more across eight models. Interactive explanation.
• eval-power: how many benchmark items it takes to tell two LLMs apart, and how often a small pilot gets that number wrong. In a live test on six small models, plans committed before collecting fresh answers detected 23 of 27 differences. Calculator.
• seed-power: the same question for RL seeds. On more than 77,000 fresh training runs, studies planned from small pilots for 80% power detected the difference 56% of the time with linear policies and 45% with neural PPO.
• cpu-decode: a from-scratch C++ int8 decoder for Qwen2.5-0.5B. At short context it reaches 83% of the laptop's memory-read ceiling and, at its best thread count, beats llama.cpp; at long context llama.cpp is clearly faster.
• control-clock: seconds from a fresh Python process to a policy that passes CartPole or Acrobot. Random search over linear policies beat every PPO on CartPole; tuned PPO won Acrobot; a cloud GPU was slower once startup counted.
• alignmenttax: does instruction tuning trade calibration for truthfulness? In seven base/instruct pairs, calibration got worse in six, but only four gained accuracy in exchange.
• quantile-cycles: a counterexample in distributional RL, checked in exact rational arithmetic and in Lean.
• verge-lab: picking preference pairs from multi-aspect judge scores, and abstaining when the aspects disagree. Neither agreement with human preferences nor DPO training on the selected pairs beat a simple score-gap rule.
• branchpilot: a learned stopping rule for self-consistency sampling. Simple agreement rules won on GSM8K and on a MATH-500 holdout.
• faultline: recurrent PPO agents that have to run a cheap test before an expensive repair. Across 375 seeds per curriculum, training only on ambiguous faults beat random sampling by 10 points but lost to a difficulty curriculum by 6.
• heliostune: Triton autotuning across four NVIDIA GPUs. Tuning data from other GPUs didn’t help; on the H100 the limit was the kernel set, and adding split-K and persistent kernels won 9 of 96 workloads against torch.matmul, up from none.

Merged fixes in PyTorch, Ray, Sentence Transformers, SGLang, and Celery. More at mottopanikeiku.github.io.