Скачать или смотреть Heather Gorham and Ian McKenzie Discuss Inverse Scaling & AI Safety (Applied Context Ep.2)

Applied Context is a series highlighting AI researchers and the works they have contributed in pushing AI innovation forward. In this episode, Heather Gorham sits down with Ian McKenzie, Researcher at FAR AI and author of the paper Inverse Scaling: When Bigger Isn't Better.

We talk through Inverse Scaling, the pace of innovation in AI, existential risk and how we should think about safety as these systems scale. It was important to me to include AI safety and the work that's being done in the community early on in this series - a HUGE thank you to Ian for being willing to start the important conversation here.

Links to materials we talk through below:

Inverse Scaling: https://arxiv.org/abs/2306.09479

Scaling Laws: https://arxiv.org/abs/2001.08361

Deceptive Alignment Monitoring: https://arxiv.org/abs/2307.10569

The Alignment Problem from a Deep Learning Perspective: https://arxiv.org/abs/2209.00626

Weak-to-Strong Generalization: https://openai.com/index/weak-to-strong-ge...

Different perspectives on concept extrapolation: https://www.alignmentforum.org/posts/j9vCE...

Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet: https://transformer-circuits.pub/2024/scal...

Gradient Hacking: https://www.alignmentforum.org/posts/EeAgy...

Playing the Training Game: https://www.alignmentforum.org/posts/pRkFk...

Me, Myself, and AI: The Situational Awareness Dataset (SAD) for LLMs: https://arxiv.org/abs/2407.04694

Model Organisms of Misalignment: https://www.alignmentforum.org/posts/ChDH3...

AI Timelines (Daniel Kokotajlo): https://www.alignmentforum.org/posts/K2D45...

Measuring Progress on Scalable Oversight for Large Language Models: https://arxiv.org/abs/2211.03540

FAR AI: https://far.ai/

Ian McKenzie: https://far.ai/author/ian-mckenzie/

Heather Gorham: https://www.linkedin.com/in/heather-gorham/

Flying Fish Ventures: https://www.flyingfish.vc/