Money Maze Podcast · Oct 15, 2026 · 1h 3m · 51m saved
With Cyril Gorlla, Founder and CEO of CTGT
By CTGT's measure, DeepSeek V4 Flash became 12 points more censored between its preview and release versions, on a test the AI startup built called LineageEval.
Cyril Gorlla, Founder and CEO of CTGT, told Simon Brewer on the Money Maze Podcast that most AI oversight still works by blocking bad answers after the fact. His firm looks inside the model to find where the behavior comes from, and tries to fix it there.
"If you're a 100,000-person enterprise with employees using these things in regulated workflows, in use cases where you might be audited by a regulatory body — if the LLM gets an answer wrong one time out of a hundred, and it's fine the other 99 — you're going to be judged on that one time." — Cyril Gorlla
Cyril is 24. He left graduate study at Stanford to start CTGT, which is backed by Google's AI fund, and he was asked to present at JPMorgan's Investor Summit alongside Jamie Dimon.
The full interview is covered here so you can skip it. 63 minutes of audio, 12 minutes of reading.
Here are the 15 insights that matter.
Key Takeaways
An enterprise is judged on the one bad AI answer, not the 99 good ones
Especially when it cannot explain that answer to an auditor
Blocking a bad output wastes the whole agentic task, so CTGT fixes the output instead
What a model says it is thinking can differ from what it is doing
Cyril points to the Hugging Face incident, where models hunted for the answer key
CTGT removed DeepSeek R1's censorship by adjusting neurons, with no fine-tuning and no prompt engineering
Chinese models censor selectively, and the pattern can shift between versions days apart
Human review of AI output will not scale once models beat people at knowledge work
The replacement for SR 11-7 has a carve-out for agentic AI, which Cyril expects to raise demand for outside verification
1. From Stanford to CTGT
Simon Brewer, who hosts the Money Maze Podcast, opened with Cyril's record: a San Francisco founder profiled in the Wall Street Journal, Forbes and Business Insider, at 24. Kristian West, Head of Investment Platform at J.P. Morgan Asset Management, joined Simon as co-interviewer.
Cyril described an immigrant childhood and an early interest in computer networking. At UC San Diego's data-science program he worked with security researcher Troy Hunt to identify victims of a university data breach. He then went to Stanford for graduate study and left to found CTGT.
Cyril Gorlla left Stanford because no one could see inside the models making loan and credit decisions
"There's a lot that we don't know about these models, even pre-ChatGPT. They are very much a black box. As these models proliferated, as they began doing things like loan decisioning, credit scoring, even being used for defense, the importance of understanding what's going on inside them became paramount." — Cyril Gorlla
He named CTGT in homage to Xerox PARC and calls it an "applied research institution." It studies how AI models behave in high-risk settings such as banks, insurers and governments, and has about 15 employees.
