The AI Alpha Brief

The AI Alpha Brief

Independent AI research, benchmarks, and practical insights for capital markets professionals.

The safety hardening had a waiting list

Aug 30, 2026

•

4 min read

The safety hardening had a waiting list

A lab delayed its own weight release over cyber capability. Someone had already registered the repo to strip it out.

Joseph Yuschak
Joseph Yuschak
Three regulators agree on one thing: show your work

Aug 23, 2026

•

3 min read

Three regulators agree on one thing: show your work

SEC, FINRA, FCA, and DFSA all converged on the same message this year: no new AI rulebook is coming. The question is whether you have anything to show when asked.

Joseph Yuschak
Joseph Yuschak
OpenAI proved ten open math problems — then proved the proofs

Aug 16, 2026

•

4 min read

OpenAI proved ten open math problems — then proved the proofs

Astra's headline wasn't the mathematics. It was the Lean 4 certificates. Meanwhile our newest benchmark says a 30B model on your own hardware matches the frontier on detection — and still can't compute the number.

Joseph Yuschak
Joseph Yuschak
Nobody's grading finance's AI — and the SEC just started fining firms for it

Aug 9, 2026

•

2 min read

Nobody's grading finance's AI — and the SEC just started fining firms for it

Three firms went all-in on Anthropic, the White House's AI-benchmarking deadline lapsed, and the SEC is prosecuting AI-washing as fraud. The gap is now a liability.

Joseph Yuschak
Joseph Yuschak
Issue #004 — The AI Alpha Brief

Aug 2, 2026

•

2 min read

Issue #004 — The AI Alpha Brief

Six frontier models on daily P&L reconciliation. "Thinking" stopped predicting who wins — by 28 points.

Joseph Yuschak
Joseph Yuschak
Issue #003 — The AI Alpha Brief

Jul 26, 2026

•

2 min read

Issue #003 — The AI Alpha Brief

Most AI accuracy numbers in capital markets can't be reproduced. We built 1,000 benchmark cases for about $150.

Joseph Yuschak
Joseph Yuschak
Load more