Now · updated 10 July 2026

What I’m focused on.

In the spirit of nownownow.com — updated periodically.

Reading, deeply

Working through six gate papers on scheming and chain-of-thought monitorability at a deliberate two-per-week pace, with real notes: Apollo’s In-Context Scheming, Turpin’s Language Models Don’t Always Say What They Think, the cross-lab Chain of Thought Monitorability position paper, the cross-lingual CoT-monitoring fragility results, OpenAI’s obfuscation findings, and CoT Red-Handed as the methodological template for my first experiment.

Lost in the Monitor — is cross-lingual chain-of-thought oversight failure monitor-side or model-side? Right now that means text-only preparation: Czech translation packs with native verification, harness skeleton, and a pre-registered analysis plan. No experiments until September — by design, not by accident.
The build phase starts mid-September; a first public research note is planned for early October, and the full write-up with open code and transcripts by the end of November 2026.

Final state exams at MFF UK in mid-September are the immediate priority — everything above is sized to fit around them, not compete with them.

The agent runtime stays my responsibility, and I’m gradually handing off day-to-day work — research is the primary focus from autumn 2026 onward.

First public write-ups will follow the first results — I’d rather publish measurements than intentions. Research notes and replication write-ups will go up openly (blog + LessWrong cross-posts) as the project produces them, starting this autumn.