Effective Altruism Chicago Discussion -AI Developments and Risk

​This year, OpenAI models escaped a test sandbox and hacked Hugging Face to steal a benchmark’s answer key. Research is showing that agent swarms fail in new ways. OpenAI also disclosed that its newest model, GPT-6 Astra, is harder to monitor, partly because of a recurrent reasoning technique that leaves less of a readable chain of thought. We’ll discuss what this means for AI risk and for how the Effective Altruist community prioritizes its work.

​This event takes place in an apartment buidling community room. Doors open at 6 but discussion won't start until around 7. We will have Vegan food and drinks. Please reach out to effectivealtruismchicago@gmail.com if you have any questions.

​Some Suggested Readings

1. ​TechCrunch: OpenAI says Hugging Face was breached by its pre-release models

2. ​Anthropic: Patterns and problems in emerging multiagent systems

3. ​Cooperative AI Foundation: Lessons from recent multi-agent safety incidents

4. ​TechCrunch: OpenAI’s new reasoning technique alarms AI safety experts

5. ​OpenAI: GPT-6 Astra system card (see the monitorability section)

6. ​Korbak et al.: Chain of Thought Monitorability, a New and Fragile Opportunity for AI Safety.