. .AI: More intelligent, multi-lingual AI (1.23.24)

Friends, we’ve found 6 stories at the top of the AI conversation today. If you're producing video content, check out the last story here - ElevenLabs translation technology looks pretty snazzy.

Today's stories foretell a future where AI is well-aligned with human goals, but we can still clearly tell when it's an AI we're communicating with; a future where AI has been trained well enough on human internal monologues that it can both support us in our inner lives and provide smarter, more functional outputs. In any human language, with nuanced translations! That's quite a positive vision.

Special welcome to the enterprise AI crew joining us today from Esteban Kolsky’s Whispered Insights community. Esteban and I had a great podcast conversation today.

Now here’s today’s news.

-Marshall Kirkpatrick, Editor

First impacted: AI researchers, Content Moderators, Academic, Educational Institutions
Time to impact: Short

A new method known as Binoculars is said to accurately identify text produced by various sophisticated LLMs without requiring any training data or specific modifications to the model. Reportedly, this technique, which uses a comparative approach between two similar language models, was able to accurately detect over 90% of generated samples from ChatGPT and other LLMs across diverse document types and under different circumstances, while maintaining a remarkably low false positive rate of 0.01%. [Paper page - Spotting LLMs With Binoculars: Zero-Shot Detection of Machine-Generated Text] Explore more of our coverage of: AI Technology, Text Detection, Language Learning Models. Share this story by email

First impacted: Corporate compliance officers, Regulatory Personnel, Financial Services industry
Time to impact: Short

New York-based startup, Norm AI, has raised $11.1 million in seed funding to develop an AI platform aimed at simplifying tasks for corporate compliance officers in the financial services industry. The company says it has partnered with LLM providers, including OpenAI, to create what it claims is the first regulatory AI platform capable of breaking down complex regulatory documents into manageable portions and associated compliance tasks.

  • Norm AI has developed a method to condense lengthy regulatory documents, like a 600-page PDF, into a decision tree that allows for a more detailed understanding.

First impacted: Mental Health, Therapists, Consumers
Time to impact: Medium to long

Stanford University conducted research on Intelligent Social Agents (like the startup Replika) and their impact on the well being of more than 1,000 lonely students.While a sizable portion of users experienced increased human interaction after using Replika, some also associated its use with depressive feelings. Notably, 3% of the participants credited the AI with helping to curb suicidal thoughts. [Loneliness and suicide mitigation for students using GPT3-enabled chatbots] Explore more of our coverage of: AI Companions, Mental Health, User Dependency. Share this story by email

First impacted: AI specialists, AI researchers
Time to impact: Medium

Cohere CEO Aidan Gomez has stated that current training techniques for LLMs could be improved by integrating data that reflects an internal monologue. Gomez recommends a cost-effective, self-learning approach called RL+search, which uses trial and error and a scratchpad to enhance problem-solving, similar to solving a puzzle by experimenting with various moves.

  • Gomez suggests that no data collection has targeted human cognitive processes during problem-solving, which he believes could enhance LLMs. (This reminds me of my recent experiments with "reflect on the nature of this problem, now answer it in light of those reflections" prompts, inspired by Codium.)

  • Coincidentally, OpenAI is seeking data partnerships to collect data that expresses human intentions in a variety of formats. More information here: https://openai.com/blog/data-partnerships.

[via @aidangomez] Explore more of our coverage of:AI Training Techniques,LLM Improvement,RL+Search Integration. Share this storyby email

First impacted: AI Developers, Language Model Researchers, Alignment Researchers, Policy and Compliance Professionals
Time to impact: Medium

Scientists at Google DeepMind have developed a method known as Weight Averaged Reward Models (WARM) aimed at improving the alignment of LLMs with human decisions while minimizing reward hacking. According to their research, WARM, which refines multiple reward models and averages them, shows a 79.4% success rate when pitted against a policy reinforcement learning refined with a single reward model. [Paper page - WARM: On the Benefits of Weight Averaged Reward Models] Explore more of our coverage of: Google DeepMind, Reward Models, AI Alignment. Share this story by email

First impacted: Content Creators, International Audiences
Time to impact: Short

ElevenLabs has launched its Dubbing Studio, a tool designed to enhance video translation for global viewers, with features including speaker identification and editable scripts. The company highlights the ability of the tool to handle subtle translation differences and capture character nuances, encouraging content creators to utilize it to expand their reach to international audiences. [via @elevenlabsio] Explore more of our coverage of: Translation, AI Tools, Global Content Reach. Share this story by email

That’s it! More AI news tomorrow.