AI Engineering
Before language models, I simulated excitable media like the waves shown here. The habits carried over: build the model, measure carefully, report what you find. These are my notes on fine-tuning and evaluating LLMs.
Posts
-
Can an Open-Weight Model Prove Code Never Crashes?
Anthropic had Claude prove that real programs cannot crash. I tried the same loop at small scale with an open-weight model on two DGX Sparks: GLM-5.3-Flash writing Verus proofs for eight Rust routines. It proved all of them crash-free and found every planted bug, but one fix silenced a crash without fixing it, and the…
-
Async GRPO on Two DGX Sparks: Same Curve, Different Bottleneck
Hugging Face trained Qwen2.5-Math-1.5B with TRL’s asynchronous GRPO trainer on four H200s. I reran it on two DGX Sparks on my desk. The reward curve reproduced; the tuning advice did not, and one of the speed knobs quietly changed the training objective.
-
Three Ways to Improve a Small Agent, and What Each Was Actually Worth
I wanted to know what prompting, finetuning on demonstrations, and reinforcement learning are each actually worth. So I measured all three on the same 8B agent, the same 240 held-out games, the same scoring. The ordering was not what I expected.
Earlier research
- 2019PD-1/PD-L1 immune checkpoint and p53 loss facilitate tumor progression in activated B-cell diffuse large B-cell lymphomas. Blood. doi:10.1182/blood.2018889931
- 2018The cytidine deaminase under-representation reporter (CDUR) as a tool to study evolution of sequences under deaminase mutational pressure. BMC Bioinformatics. doi:10.1186/s12859-018-2161-y
- 2016The relationship between Nernst equilibrium variability and the multifractality of interspike intervals in the hippocampus. Journal of Computational Neuroscience. doi:10.1007/s10827-016-0633-5
- 2016Non-invasive evaluation of cardiac repolarization in mice exposed to single-wall carbon nanotubes and ceria nanoparticles via intratracheal instillation. Environmental Science: Nano. doi:10.1039/c5en00225g
- 2015Bursting regimes in a reaction-diffusion system with action potential-dependent equilibrium. PLOS ONE. doi:10.1371/journal.pone.0122401
- 2012Multiple mating but not recombination causes quantitative increase in offspring genetic diversity for varying genetic architectures. PLOS ONE. doi:10.1371/journal.pone.0047220