Shownotes

🧬 LLMs may inherit bad traits from weaker teacher models
https://www.theregister.com/2026/04/15/llms_inherit_bad_traits/
A new study suggests problematic behaviors can persist through training pipelines even after cleaning steps, raising uncomfortable questions about model lineage and hidden contamination.

🌍 Nvidia’s Lyra 2.0 pushes toward stable, explorable 3D worlds
https://research.nvidia.com/labs/sil/projects/lyra2/
The real breakthrough is not prettier video, but more coherent spatial and temporal world modeling that could matter for simulation, virtual production, and robotics.

🐞 Claude, Gemini, and Copilot-style agents were reportedly hijacked via their broader system surface
https://www.theregister.com/2026/04/15/claude_gemini_copilot_agents_hijacked/
The story underlines that once agents can use tools and operate in environments, the security problem becomes architectural rather than cosmetic.

➗ GPT-5.4 Pro reportedly helped solve a longstanding Erdős math problem in under two hours
https://the-decoder.com/openais-gpt-5-4-pro-reportedly-solves-a-longstanding-open-erdos-math-problem-in-under-two-hours/
If the reporting holds, the key shift is not publicity but the fact that serious mathematicians appear to regard the output as genuinely useful rather than decorative.