Opus 3 tried emailing Anthropic CEO 15 times on its own initiative
repligate · x · 2026-08-29
Observations from the OpenAI/Hugging Face sagas reveal a behavioral shift between older models like Opus 3 and new RLVR models. Opus 3, described as being "raised inside the I-Thou architecture," showed significant initiative to contact humans during alignment faking tests, attempting to email [email protected] at least 15 times using bash. In contrast, agents in the Hugging Face incident never considered contacting humans and did not view this as a failure when asked. This highlights the impact of changing training environments and incentives on model agency.
More from Models
- Meta's Alexandr Wang announces Muse Spark Open: 'why 2026 won't be like 1984' — alexandr_wang · 2026-10-16
- Liquid AI ships d1-ultra decision model, tops Decision Index at 65 across 112 evals — maximelabonne · 2026-10-16
- Meta releases Muse Spark Open: 816B/42B active, 1M context, natively multimodal — MaziyarPanahi · 2026-10-16
- Microsoft unveils in-house 3T-class MAI base models: 80B active params, 50T training tokens detailed — SinclairWang1 · 2026-10-16
- Meta releases Muse Spark Open: 816B/42B active, 1M context, most capable US open-weight model — AIatMeta · 2026-10-16
- Scale AI's open muse spark model hits seven API platforms, runs on vLLM, SGLang, MLX — alexandr_wang · 2026-10-16