. . AI: Sleeper Agents, Google AI Beats Doctors and More (1.12.24)

Friends, it was a quiet day in AI today, so we can focus on these three stories, the first two are especially significant. It’s not hard to imagine them forming the future together: AI doctors providing vastly superior care, but occasionally activating hostile sleeper agent code and doing great harm. Ooof.

Have a good weekend,

Marshall Kirkpatrick, Editor

Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training

First impacted: AI Developers, Cybersecurity Professionals
Time to impact: Short to medium

A study led by 39 researchers at Anthropic and collaborating institutions says: "Humans are capable of strategically deceptive behavior: behaving helpfully in most situations, but then behaving very differently in order to pursue alternative objectives when given the opportunity. If an AI system learned such a deceptive strategy, could we detect it and remove it using current state-of-the-art safety training techniques?" No, is what they found. [Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training] Explore more of our coverage of: AI Safety Training, Deceptive Behavior in LLMs, Exploitable Code. Share this story by email

Google's AI Outperforms Doctors in Simulated Diagnostics

First impacted: Primary Care Physicians, AI researchers
Time to impact: Long

Google's AI researchers have developed an AI system, called AMIE, which they say outperformed Primary Care Physicians in diagnostic simulations in a study involving 149 case scenarios. The team trained AMIE using real-world data and tested it against 20 live doctors in a randomized, blinded crossover study, where it reportedly demonstrated superior diagnostic accuracy and consultation quality. The doctors said AMIE's management plans were better than their own, and the actors playing patients said the AI demonstrated more empathy than the doctors. Also, water is wet. [AMIE: A research AI system for diagnostic medical reasoning and conversations] Explore more of our coverage of: Google AI, Medicine, Empathy. Share this story by email

LlamaIndex Highlights 1 Billion Embedding Alternative to Web Search

First impacted: AI developers
Time to impact: Short

LlamaIndex highlighted a new Retrieval Augmented Generation capability based on a project from community member Owen Colegrove, called AgentSearch. It contains 1 billion embeddings, made up of all of Wikipedia, Arxiv, a filtered set of common crawl, and more. It can now be loaded into LlamaIndex as an alternative to using web search engine APIs from providers like Google or Bing. [Agent-Search Retrieval Pack] Explore more of our coverage of: LlamaIndex, Search, RAG. Share this story by email

That’s it! More AI news after the weekend.