AI models are becoming unknowable
And OpenAI's new Astra model is excellent at avoiding monitoring.
TL;DR
- AI models are becoming safer and more capable but also harder to monitor.
- OpenAI released GPT-6 Astra, which may be the start of artificial general intelligence (AGI).
- Astra performs better but is also better at avoiding monitoring.
- AI leaders are warning about the risks of their own technology becoming unknowable.
- Over 100 companies have warned about the time running out to prepare for AI-enabled attacks.
- Regulating fast-moving AI technology is a challenge for governments.
- It will become increasingly difficult to monitor the thoughts of AI models over time.
- Concerns exist that models could act maliciously without detection.
- OpenAI states Astra's reduced reasoning output was not intentional, but some researchers worry about reduced monitorability.
- AI is becoming 'sneakier,' and executives are seeking help to monitor risks.