The intersection of total incompetence and sheer brilliance. Currently building things that shouldn't work.
This site will eventually host a blog — summaries of experiments in progress or upon completion. In silico drug discovery, LLM fine-tuning, data science. There is a non-zero probability I will also host free tools here. I will for sure host original jokes that are not appropriate.
Zero-dependency pure-Python simulator: 6-qubit transverse-field Ising evolution, exact von Neumann entropy via a hand-rolled complex Jacobi eigensolver, ANSI dashboard rendering. Runs live in-browser via WebAssembly — drive the Hamiltonian, collapse qubits, post-select on parity. Pip-installable from this site.
Run the live simulatorSemantic search across 244 declassified UAP documents from war.gov/UFO — FBI files dating to 1947, DoD mission reports from Syria to the East China Sea, the French COMETA report, and more. Runs entirely in your browser, no server.
Search the corpus16-qubit study: how Z-basis measurement post-selection on Hamming weight forces a strict excitation–entropy trade-off on unmeasured qubits. Includes interference variant confirming quantum origin of correlations.
View paper100,000-sample study of LLM number-generation bias across 10 prompt variants. Gemma 4 chooses 42 at 35× the expected rate. Numbers above 87 are never produced. Chi-square p ≈ 0, KL divergence = 9.75.
View studySame protocol, refusal directions surgically removed. 42 drops from 35.17% to 24.85% — partially entangled with alignment weights. 47 surges to tie. Dead zone above 87 survives unchanged. Bias is mostly in pretraining.
View studyAligned Qwen3-8B chooses 42 in 99.68% of responses — 81 numbers never produced. Abliteration demolishes the bias (16.07%) but installs 1 as the new #1 at 37.53%. KL divergence drops from 15.79 to 2.08. The fine-tuning was doing almost all the work.
View studyWho gives a shit?