Now · updated 25 May 2026
What I’m focused on.
In the spirit of nownownow.com — updated periodically.
I.
Research
- Reading the scheming and deception detection literature: Apollo Research’s Frontier Models Are Capable of In-Context Scheming, Anthropic’s Alignment Faking, Reasoning Models Don’t Always Say What They Think, and Apollo’s Chain of Thought Monitorability.
- Setting up the infrastructure for replicating Apollo’s in-context scheming benchmark on open-weight reasoning models (DeepSeek-R1, Qwen QwQ), with a cross-lingual extension comparing scheming rates and CoT faithfulness across English, Czech, and bilingual conditions.
- Reaching out to potential Master’s advisors at ÚFAL, MFF UK, to lock in a thesis direction by autumn 2026.
II.
Master’s preparation
- Working through the ARENA self-paced curriculum (Chapter 0 first, then Chapter 3 on RL fundamentals, which underpins reasoning-model training).
- Applying for the AI Safety Fundamentals alignment cohort starting later this year.
- Drafting research proposals for Apollo Research, MATS, Anthropic Fellowship, and AI Safety Camp for the December 2026 application round.
III.
Bachelor’s defence
- Bachelor’s thesis defence on 18 June 2026, then final state exam on 21 June 2026. Until then this is the priority.
IV.
Zoplio
- Iterating on the agent-to-agent negotiation protocol with Lukáš, and preparing the next round of conversations with potential design partners.
- The company runtime stays my responsibility, but research is the primary focus through the end of 2026.
V.
Public writing
- A blog and Substack are being set up. First post in draft: a personal account of moving from systems engineering into AI safety research and what drew me to scheming/deception evaluation.
- Replication notes will be published openly as the project progresses, with LessWrong cross-posts for the community discussion.