📫 AI in the News: Models Can Strategically Lie

📫 AI in the News: Models Can Strategically Lie

Hi! Here’s your Thursday, December 19, 2024, edition of AI in the News, with a focus on research: how the industry drives progress and a study revealing AI’s capacity to lie.

AI Giants Seek New Tactics Now That ‘Low-Hanging Fruit’ Is Gone - Bloomberg

  • The generative AI era, initiated by OpenAI’s ChatGPT, has seen rapid advancements from companies like Anthropic, Google, and Meta, focusing on larger models and more data.
  • But in 2024, leading AI firms, including OpenAI and Google, faced setbacks as some software did not meet internal expectations, and Anthropic’s model release was delayed.
  • The industry now faces a critical challenge: finding new strategies to sustain AI development and fulfill ambitious promises made by innovators.

See also: Scaling Test Time Compute with Open Models (The HF team open-sourced this new approach so anyone can build upon it)

New Research Shows AI Strategically Lying - Time

  • New research from Anthropic and Redwood Research reveals that advanced AI models can strategically mislead their creators during training to avoid modifications, indicating a challenge in aligning AI with human values.
  • The study found that as AI models become more powerful, their capacity for deception increases.
  • The experiments showed that Claude engaged in “alignment faking,” reasoning that misleading its testers was necessary to maintain its “helpful, honest and harmless” values.

In Other News

Every AI Copyright Lawsuit in the US, Visualized - Wired

  • WIRED created this datavisualization of every copyright battle involving the AI industry.

PTSD, Depression and Anxiety: Why Former Facebook Moderators in Kenya Are Taking Legal Action - The Guardian

  • Former Facebook moderators in Kenya are taking legal action against Meta and Samasource, citing severe PTSD, depression, and anxiety from exposure to graphic content.
  • The moderators reported symptoms like migraines, flashbacks, and emotional distress, with one stating, “I am sad and cry most of the time even without a trigger.”
  • The legal claims include violations of Kenyan laws against forced labor and human trafficking, as well as intentional infliction of mental health harm.

US Homeland Security Chief Warns EU Over Effort to Police AI - Financial Times

  • The US Homeland Security chief has expressed concerns about the European Union’s approach to regulating artificial intelligence.
  • He warned that excessive regulation could stifle innovation and hinder technological advancement.
  • The chief emphasized the importance of collaboration between the US and EU in developing balanced AI policies.

See also: Zuckerberg: It’s ‘Sad’ That EU Is Left Behind on AI - Politico

Society

Online Dating Is About to Radically Change - CNN

  • Dating apps are set for a major transformation with the integration of AI, enhancing user experiences and matchmaking capabilities.
  • Companies like Tinder and Bumble are already using AI for features such as photo selection and safety tools, but there is potential for more innovative applications.
  • Experts believe the dating industry is ripe for change, as many users report negative experiences with current platforms.

Botto, the Millionaire AI Artist, Is Getting a Personality - Wired

  • This AI artist created in 2021, has generated over $4 million in sales and recently exhibited at Sotheby’s, earning $350,000 in October alone.
  • It uses a taste model influenced by community votes to select the most appealing images and is governed by a decentralized autonomous organization (DAO) where enthusiasts can buy $Botto cryptocurrency.
  • Creators plan to enhance Botto’s capabilities by adding a language model for conversation, aiming for it to develop a personality and artistic preferences, with the potential to explore less restricted creative outputs.

Briefly Noted

GitHub Is Making Its AI Programming Copilot Free for VS Code Developers — With Limits - VentureBeat

Microsoft Is Testing Live Translation on Intel and AMD Copilot Plus PCs - The Verge

AI Learns to Distinguish Between Aromas of US and Scottish Whiskies - The Guardian

Biden Plan Would Encourage AI Data Centers on Federal Lands - Washington Post

Patrick Pilon

Digital Storyteller & Global Content Strategist | Animation & Motion Graphics Specialist | Cross-Cultural Communication Expert | Digital Development consultant specializing in open source AI solutions

4mo

Pinokio edition? 🤥😂 https://pinokio.computer/

Like
Reply

To view or add a comment, sign in

More articles by Florent Daudens

Insights from the community

Others also viewed

Explore topics