What Jakub is focused on.

• The Bachelor’s is done: I defended the thesis in June and passed the state exam in mid-September. The Master’s in Artificial Intelligence at MFF UK starts in October. Research is now the primary track; coursework and Zoplio are arranged around it rather than the other way round. (The fuller academic picture is on the CV.)

• The live project: toy models of computation in superposition, a d-neuron layer computing more functions than it has neurons, asking what kind of solution training actually finds. The short version of the result: the superposed code grows continuously out of the dedicated-neuron solution rather than appearing at a jump, and the boundary between the families across widths and sparsities is a level set of a smoothly falling code coherence. A 40-page write-up is finished and in review before release; code public: github.com/Majny/superposition-phases. The release and the follow-up study it sets up are this autumn’s work.

• My first completed interpretability project: SAE feature absorption tracks a family of token properties rather than being a universal failure mode — on a structural property no single latent causally carries the concept at all, and controls rule out the mundane explanations. Written up as a 15-page technical report. Full description on the home page; code and figures: github.com/Majny/absorption-atlas.

• Research is now my primary focus; I remain co-founder and CTO at Zoplio, where the agent runtime stays my responsibility and I’m handing day-to-day work to the team.

• I’d rather publish measurements than intentions. The short write-up of the absorption result is on my Substack: Does the standard SAE feature-absorption metric over-report? The longer story behind this program is on the About page.

Want to compare notes on any of this: hi@kubadvorak.com.