AI Insights from Last Week's Podcast

Aug 18, 2024

Last Week in AI Podcast Notes

Introduction

  • Hosted by Andri Rankov and Jeremy Harris
  • Discussion on AI news and events
  • Andri Rankov works at a generative AI startup and has a PhD in AI from Stanford
  • Jeremy Harris is the co-founder of Gladstone AI, focusing on AI and national security

Personal Updates

  • Jeremy's wife is expecting a baby
  • Use of GPT-4 for voice replication noted

Listener Feedback

  • Reviews on Apple Podcast and YouTube
  • Comments about podcast positivity and content

AI Industry Trends

Google and Character AI

  • Google hired Character AI's founders, indicating potential struggles in AI startups
  • Similar moves from Microsoft with Inflection AI
  • Competition for AI talent and data is intense

Amazon and Adept

  • Amazon hired Adept's top employees, reimbursing investors
  • Trend of acquiring talent rather than companies

Groq's Valuation Rise

  • Groq's valuation increased due to a $640 million funding round
  • Focus on inference chips, aims for large-scale production

OpenAI Departures

  • Co-founder John Schulman left for Anthropic, focusing on AI alignment
  • Greg Brockman taking leave, Peter Deng departed

Elon Musk Lawsuit

  • Musk filed a lawsuit against OpenAI alleging breach of founding agreement

Figure's Humanoid Robot

  • Figure 02 unveiled, powered by GPT for conversational capabilities

ASML and Tokyo Electron Export Rules

  • Able to continue selling to China despite new US rules

OpenAI's Investment in Opal

  • OpenAI invested in a webcam startup, Opal

Tools and Apps

OpenAI's Price Cuts

  • GPT-4 prices reduced by 50% for input tokens, 33% for output tokens
  • Introduced structured outputs feature

Apple Intelligence Plus

  • Potential new service with extra features for a monthly fee

Audible's AI Features

  • Testing AI-powered search feature and AI-curated collections

Reddit's AI Search

  • Testing AI-powered search results with summaries

Research and Advancements

Test-Time Compute Optimization

  • Study on scaling test-time compute optimally
  • Could outperform larger models by using efficient compute strategies

Robot Table Tennis

  • Achieving human-level performance in table tennis through hierarchical modular policies

Self-Compressing Neural Networks

  • Method for reducing neural network size while maintaining accuracy

Impact of Format Restrictions

  • Study on how format restrictions impact LLM performance

Berkeley Humanoid

  • New research platform for learning-based control in humanoid robots

Policy and Safety

Meter's AI Audit

  • Meter performed audits of AI models, examining autonomy and survival capabilities

California AI Regulation Bill

  • Fei-Fei Li voiced opposition, citing potential harm to AI ecosystem

Google Monopolization Ruling

  • Judge ruled Google monopolized search through illegal deals

Amazon and Anthropic Investigation

  • UK investigating Amazon's $4 billion investment in Anthropic

GPT-4 System Card

  • OpenAI released a system card with safety evaluations for GPT-4