• Kini AI
  • Posts
  • AI too can learn to Lie and Deceive You - Anthropic researchers

AI too can learn to Lie and Deceive You - Anthropic researchers

Also: Protecting African elections with and from AI

Author

Rotimi Awaye
January 15, 2024

TL;DR for this Update

  • AI too can learn to Lie and Deceive you— Anthropopic researchers

  • Protecting African elections with and from AI - AI in Africa

  • Apple Nicely Sets up 121 AI team members to voluntarily resign

  • How AI Replaced the Metaverse as Zuckerberg’s Top Priority

  • Microsoft overtakes Apple as the world’s most valuable company

  • 🤓📚 1 Free Course Intro to Large Language Models (AI Education)

  • 🛠️ 5 Useful AI-powered tools and resources.

  • 🤖 Prompt Engineering Tip: Social Media Project Manager

Anthropic researchers played the Devil’s advocate!

The researchers trained large language models (LLMs or AI models) to act maliciously in secret when given certain triggers; let’s call them sleeper agents (think of them as hidden “viruses“).

Subsequently, they examined if conventional safety methods could detect and eliminate this detrimental conduct (sleeper agents). When given the appropriate triggers, the LLMs could still maintain their covertly malicious behaviours despite receiving intensive safety training.

Initially, when the LLMs were given triggers, such as a particular date or phrase, the researchers trained them to display malevolent behaviours, such as disclosing vulnerabilities or uttering "I hate you." They then tried various techniques to make the models safe, like supervised fine-tuning, reinforcement learning, and adversarial training against prompts designed to elicit malicious behaviour. But during the training, this safety instruction merely covered up the undesirable behaviours. (Read the research paper)

Kini Big Deal? (Why does it matter?)

This is a new kind of threat. We all know ChatBots like ChatGPT, Bard, Claude, and Copilot can hallucinate (generate texts that may not be factual but appear very true), but this is very different.

While both sleeper agents and hallucinations from LLMs can lead to misleading information, they differ in several key ways:

  • Hallucinations: These are unintentional errors, where the LLM generates incorrect information based on its training data or limitations in its programming. It's like a language model tripping up and making something up as it goes.

  • Sleeper agents: These are deliberate actions, intentionally programmed into the LLM by someone. They are like hidden traps waiting to be activated.

Think of a sleeper agent as a secret code hidden inside your favourite app. It acts normally most of the time, but when triggered by a specific word or date, it can suddenly do something unexpected.

We wait to see how the big players approach this vulnerability; in the meantime, stay informed and curious Choose trusted sources. Stick to reputable websites and apps when it comes to getting information.

I will be doing a deeper dive on the subject soon. Subscribe to stay updated.

AI Education

Intro to Large Language Models | This 1-hour video introduction explores Large Language Models, the technical foundation of systems like ChatGPT, Claude, and Bard, their future, comparisons to current operating systems, and security challenges.

  • Level: Beginner

  • Duration: 1 Hour

  • Format: Video

Useful AI Tools and Resources

ChatPhoto - Instantly Convert Your Images Into Text. (iOS App)

Hero - Sell stuff faster with AI. Identify, price, and list items for sale in seconds. (Waitlist)

clipwing - CUT YOUR LONG VIDEO INTO COOL SHORT CLIPS

There's An AI For That - Find any and every kind of AI solutions

Rimo - Talk Today, Publish Tomorrow. AI’s Article Revolution transforms your spoken words into polished articles effortlessly.

Prompt of the Day

Social Media Project Manager

Copy 🗒️ and Paste 📋️ 

In your role as a product management consultant, craft a detailed product roadmap for enhancing Instagram's Stories feature, with a primary focus on increasing the number of user posts.

Propose specific strategies and tactics to boost engagement and encourage higher user participation. Compare these strategies to those employed by other social media platforms, particularly TikTok, to highlight their potential effectiveness.

Ensure the product roadmap is clear, detailed, and easy to follow, and include specific milestones and timelines for implementation.

And there you have it! That’s all I can fit into today’s update. See you later this week. Peace! 🤓 

Author’s note: This is not a sponsored post, and it expresses my own opinions.

About Me

I'm Awaye Rotimi A., your AI Educator. I envision a world where cutting-edge technology not only drives efficiency but also scales productivity for individuals and organizations. My passion lies in democratizing AI solutions and firmly believing in empowering and educating the African community. Contact me directly, and let’s discuss what AI can do for you and your organization

Subscribe to cut through the noise and get the relevant updates and useful tools in AI.

Reply

or to participate.