We often assess AI safety by checking for harmful behaviors. But this can fail: AIs may subtly misbehave or act “right for the wrong reasons,” risking unexpected failures. Instead, we propose alignment audits to investigate models for hidden objectives.
RESEARCH
-

AI Alignment Audit: Teams Successfully Identify Hidden Model Objectives
By
–
To practice alignment audits, our Alignment Science and Interpretability teams ran a blind auditing game. A red team trained—in secret—a model with a hidden objective, then gave it to four blue teams for investigation. Three teams won by uncovering the model’s hidden objective.
-

Anthropic Research: Auditing Language Models for Hidden Objectives
By
–
New Anthropic research: Auditing Language Models for Hidden Objectives. We deliberately trained a model with a hidden misaligned objective and put researchers to the test: Could they figure out the objective without being told?
-

RAG Techniques: Improving LLM Retrieval Systems with Embeddings
By
–
This episode of Data Brew goes in-depth into RAG, tips and tricks, and how the technique address limitations of LLMs. Databricks Research Scientist Andrew Drozdov chats about:
– Improving retrieval systems with embeddings and fine-tuning techniques
– Enhancing search results -
Gemma 3 Technical Report and Memory Usage Analysis Script
By
–
Here's a link to the Gemma 3 technical report: https://
storage.googleapis.com/deepmind-media
/gemma/Gemma3Report.pdf
… I shared the script to replicate the memory usage vs context length analysis here: https://
gist.github.com/rasbt/b7980a2a
33e33847f08b1543f0627950#file-llama-vs-gemma-py
… (not very pretty but it should work) -

Gemma 3 Analysis: Large Vocabulary Size and Model Insights
By
–
Just read through the Gemma 3 report and toyed around with the models a bit, and there are a bunch of interesting tidbits: 1. Vocab size. They again use a very large vocab: 262k token (in contrast, Llama 3 has ~1/2 the vocab size), which should make the model more friendly for
-

Gemma3 Local Deep Research Assistant with Ollama
By
–
Fully local deep research assistant with Gemma3 Google just released Gemma3, several open source highly performant models that can be run locally. We overview Gemma3 and test it w/ local deep research using @ollama
. : https://
youtu.be/XsJxF_MDfyI : https://
github.com/langchain-ai/o
llama-deep-researcher
… -
Uncertainty over Federal Subsidies Threatens U.S. Bioscience Leadership
By
–
Uncertainty over federal grants is endangering U.S. leadership in bioscience, which is critical to drug development and the nation's competitiveness wsj.com/articles/bioscience-… via @WSJ @UW @UWproteindesign @UWMedicine [Translated from EN to English]
→ View original post on X — @steve_rosenbush, 2025-03-13 15:44 UTC
-
SAS Drives Climate Innovation Through Sustainability Technology
By
–
🌱 Climate change is not a distant threat—it's happening now. From rising temperatures to shrinking ice caps, our 🌎 needs action today. See how Southern States Energy Board is using SAS to drive innovation, champion sustainability & protect our future. https://t.co/u0eYnuMnXz pic.twitter.com/HOQcyY1mN6
— SAS Software (@SASsoftware) 13 mars 2025Climate change is not a distant threat—it's happening now. From rising temperatures to shrinking ice caps, our needs action today. See how Southern States Energy Board is using SAS to drive innovation, champion sustainability & protect our future. http://
2.sas.com/6016L7KME -

Google Upgrades Deep Research with Gemini 2.0
By
–

BREAKING : Google is upgrading its Deep Research feature to be powered by Gemini 2.0 Flash Thinking. The mention of Gemini 1.5 Pro has been removed from the description recently.