When Two LLMs Debate, Both Think They'll Win Can LLMs accurately adjust their confidence when facing opposition? Building on previous studies measuring calibration on static fact-based question-answering tasks, we evaluate Large Language Models ...

Slightly disappointed that this paper doesn’t cite arxiv.org/abs/2505.19184 the OG paper on inflated and overconfident self-assessments pubmed.ncbi.nlm.nih.gov/10626367/ as the similarities are *noteworthy*

An assortment of tomato representations: plastic toy fruit, toy slices, wooden toy fruit with Velcro for cutting, real tomatoes, tomato ketchup, toy tomato ketchup, sliced tomatoes…

The cover of “Simulacra and Simulation” by Jean Baudrillard

Excerpt: “This is certainly why order always opts for the real. When in doubt, it always prefers this hypothesis (as in the army one prefers to take the simulator for a real madman). But this becomes more and more difficult, because if it is practically impossible to isolate the process of simulation, through the force of inertia of the real that surrounds us, the opposite is also true (and this reversibility itself is part of the apparatus of simulation and the impotence of power): namely, it is now impossible to isolate the process of the real, or to prove the real.”

Saturday morning playtime