Research monitor · Page 14

Introducing Claude 2.1

Agent systems · provider_announcement

Claude 2.1 brings a 200K token context window, a 2x reduction in hallucination rates, beta tool use, and system prompts.

Expanding access to safer AI with Amazon

AI models · provider_announcement

Amazon will invest up to $4 billion in Anthropic and become its primary cloud provider, with models trained on AWS Trainium and Inferentia.

Introducing Anthropic's Responsible Scaling Policy

AI models · provider_announcement

Our Responsible Scaling Policy is a series of technical and organizational protocols to manage the risks of increasingly capable AI systems.

Anthropic partners with BCG

AI models · provider_announcement

Anthropic and Boston Consulting Group partner to bring Claude to BCG's enterprise clients for knowledge management and data analysis.

Question decomposition improves reasoning faithfulness

Mathematical reasoning · provider_announcement

Breaking complex questions into subquestions produces more faithful reasoning than chain-of-thought while retaining accuracy gains.

Measuring faithfulness in Chain-of-Thought reasoning

Mathematical reasoning · provider_announcement

Does chain-of-thought reasoning reflect a model's actual decision process? We find faithfulness varies by task and declines in larger models.

Charting a path to AI accountability

Evaluation & safety · provider_announcement

Anthropic's NTIA submission recommends rigorous model evaluations, pre-registration of large training runs, and mandatory external red teaming.

Improving mathematical reasoning with process supervision

Mathematical reasoning · provider_announcement

We’ve trained a model to achieve a new state-of-the-art in mathematical problem solving by rewarding each correct step of reasoning (“process supervision”) instead of simply

Circuits Updates — May 2023

Complexity theory · provider_announcement

Developing interpretability ideas on dictionary learning, defining features, and the possibility of superposition in attention heads.

Introducing 100K context windows

AI models · provider_announcement

Claude's context window expands from 9K to 100K tokens, roughly 75,000 words, letting it process hundreds of pages in under a minute.

An AI policy tool for today: Invest in NIST

AI models · provider_announcement

Anthropic urges Congress to ambitiously fund NIST to build AI measurement tools, standards, and testbeds.

Anthropic partners with Google Cloud

Compute & efficiency · provider_announcement

Anthropic selects Google Cloud as preferred cloud provider, using its GPU and TPU clusters to train, scale, and deploy AI systems.

Anthropic raises Series B to build robust AI

Evaluation & safety · provider_announcement

Anthropic raised $580 million in Series B funding to build infrastructure for studying and improving the safety of large-scale AI models.

Solving (some) formal math olympiad problems

Formal verification · provider_announcement

We built a neural theorem prover for Lean that learned to solve a variety of challenging high-school olympiad problems, including problems from the AMC12 and

A mathematical framework for Transformer Circuits

Complexity theory · provider_announcement

Reverse-engineering small attention-only transformers reveals how induction heads form via head composition.

Solving math word problems

Mathematical reasoning · provider_announcement

We’ve trained a system that solves grade school math problems with nearly twice the accuracy of a fine-tuned GPT-3 model. It solves about 90% as

Anthropic raises $124 million Series A

AI models · provider_announcement

Anthropic's Series A, led by Jaan Tallinn, funds research toward steerable, interpretable, and robust large-scale AI systems.

Generative language modeling for automated theorem proving

Formal verification · provider_announcement

GamePad: A learning environment for theorem proving

Formal verification · provider_announcement