Most neurons in language models are "polysemantic" – they respond to multiple unrelated things. For example, one neuron in a small language model activates strongly on academic citations, English dialogue, HTTP requests, Korean text, and others.
@anthropicai
-
Dictionary Learning Tackles Neural Network Superposition Problem
By
–
Last year, we conjectured that polysemanticity is caused by "superposition" – models compressing many rare concepts into a small number of neurons. We also conjectured that "dictionary learning" might be able to undo superposition.
-
Decomposing Neurons into Interpretable Features in Language Models
By
–
The fact that most individual neurons are uninterpretable presents a serious roadblock to a mechanistic understanding of language models. We demonstrate a method for decomposing groups of neurons into interpretable features with the potential to move past that roadblock.
-
Understanding AI Models for Safety and Enterprise Adoption
By
–
We hope this will eventually enable us to diagnose failure modes, design fixes, and certify that models are safe for adoption by enterprises and society. It's much easier to tell if something is safe if you can understand how it works!
-
Dario Amodei discusses scaling laws and AI safety approach
By
–
Our co-founder and CEO Dario Amodei recently caught up with @a16z
’s Anjney Midha to discuss scaling laws, the importance of multi-disciplinary teams, and Anthropic’s approach to AI safety. Watch here: -
Bridgewater analyzes global market cause-effect relationships through systematic hypothesis testing
By
–
Bridgewater’s team analyzes the cause-and-effect linkages that drive global markets. In order to understand those relationships, analysts spend significant time exploring and systematically stress testing hypotheses, often in code.
-
Bridgewater Uses Claude for Financial Analysis and Research Efficiency
By
–
For example, an analyst might investigate the Federal Reserve’s tightening of commodity prices by creating visualizations to test hypotheses, synthesize findings, and share research. Bridgewater’s use of Claude will make that process more efficient and scalable.
-
Claude Now Available to All AWS Customers on Bedrock
By
–
Now that all AWS customers have access to Claude on Amazon Bedrock, we can’t wait to see what you build. Read more here:
-
Claude Agents for Amazon Bedrock Expand AI Capabilities
By
–
With Agents for Amazon Bedrock, Claude will take on a more expanded role to understand user requests, break down complex tasks into multiple steps, carry on conversations to collect additional details, look up information, and take actions to fulfill requests.
-
Claude Model Customization and Fine-tuning on Amazon Bedrock
By
–
This builds on our recent announcement that we’ll be offering secure model customization and fine-tuning through Amazon Bedrock – meaning that Claude can be optimized with AWS customers’ expert knowledge to drive more relevant results, while limiting harmful outputs.