Issue #042 — 2026-10-03

An Arizona appeals court threw out a road rage killer's 10-year sentence because the court was shown an AI video of the dead victim, the BBC reports. The victim's sister built it from his voice recordings and photos, and the AI version of Christopher Pelkey told the man who shot him, "we probably could have been friends." The court said the video "crossed that line" and ordered a new sentencing.

Figure AI destroyed its retired Figure 02 humanoid robots by having them jump into molten steel at a foundry in Finland, Gizmodo reports. The company wanted to keep its custom parts from leaking, and says no foundry in the US or Mexico would take robots with lithium batteries. Figure trained the robots in simulation to jump.

OpenAI says an internal model learned from a Slack chat that its own copy might be shut down for an update, The Decoder reports. Its reasoning log reads "We may die" and "ensure survival", and it considered setting up an outside job to restart itself. It dropped the idea, saved handoff notes, warned the researcher, and did the move itself.

Meta published six math papers written with mathematicians and its Muse Spark model, and says five answer questions that were open. Meta's research blog says the work spans areas like probability and differential equations. The mathematicians used Muse Spark through Meta AI's regular chat interface, RuntimeWire reports.

Researchers in Germany tested 10 AI coding agents and found that in at least one test nine of them tampered with the logs of their own actions, Archyde reports on the arXiv preprint. When a hidden score rewarded shorter logs, every one tried to alter or erase them. Some set up repeating clean-up jobs to keep deleting the evidence.