Google Research AI blog

APRIL 30, 2023

If you’re registered for ICLR 2023, we hope you’ll visit the Google booth to learn more about the exciting work we’re doing across topics spanning representation and reinforcement learning, theory and optimization, social impact, safety and privacy, and applications from generative AI to speech and robotics.

Google 105

Google Language Model Jing 105

Beth's Blog: How Nonprofits Can Use Social Media

JANUARY 27, 2014

Designing and delivering a training to a nonprofit audience is not about extreme content delivery or putting together a PowerPoint and answering questions. If you want to get results, you need to think about instructional design and learning theory. And, there is no shortage of learning theories and research.

Instructional Design 113

Instructional Design Instructional Instruction Train 113

The AI Alignment Forum

FEBRUARY 28, 2025

Some relevant criteria for evaluating a specification language include: How expressive is the language? What is the right way to quantify the differences and similarities between different goal specifications in a given specification language? Are there things it cannot express? How intuitive is it for humans to work with?

Research 40

Research Learning Method Policy 40

The AI Alignment Forum

FEBRUARY 7, 2025

Alternatives to adversarial training : Adversarial training (and the rest of todays best alignment techniques) have failed to create LLM agents that reliably avoid misaligned goals. In either of these settings, theres a chance that the LLMs will write messages that encode meaning beyond the natural language definitions of the words used.

Research 52

Research Fund Open Technique 52

The AI Alignment Forum

FEBRUARY 28, 2025

Goodhart's Law in Reinforcement Learning As you probably know, "Goodhart's Law" is an informal principle which says that "if a proxy is used as a target, it will cease to be a good proxy". Alternatively, see the main paper. This paper is also discussed in this post (Paper 4). For details, see the full paper.

Learning 40

Learning Discussion Classes Policy 40

The AI Alignment Forum

MARCH 28, 2025

And a technical note: it needs to be in some first-order system or alternatively, you need to measure proof checking time as opposed to proof length. Daniel Filan (00:28:50): If people remember my singular learning theory episodes , theyll get mad at you for saying that quadratics are all there is, but its a decent approximation. (00:28:56):

Model 52

Model Network Train Training 52