• AI Periscope
  • Posts
  • AI launches nukes in simulation | Wargames, anyone?

AI launches nukes in simulation | Wargames, anyone?

And: Meta to label AI-generated images | What to expect from Google on Bard and Gemini

Author

AI Periscope
February 06, 2024

AI war simulation

Image source: Dept. of Energy | Vice

Summary - In a recent study oddly reminiscent of the 1983 movie, WarGames, researchers found that AI models are prone to escalate conflicts to dangerous levels, including the use of nukes, when LLMs were used to simulate international conflict scenarios. This first-of-its-kind experiment, involving different AIs acting as leaders of fictitious countries, cast doubt on the practicality of using such AI in real-world military and diplomatic decision-making, highlighting the unpredictability and inherent biases in current AI technologies. The full paper can be viewed here.

Buoy points:

  • Research collaboration: Georgia Institute of Technology, Stanford University, Northeastern University, and Hoover Wargaming and Crisis Simulation Initiative took part in this research.

  • The simulation scenario: The researchers devised a game of international relations and used fictitious countries and varying geo-military attributes.

  • Five LLMs were used: The study included LLMs from OpenAI, Meta and Anthropic, specifically, GPT-4, GPT 3.5, Claude 2.0, Llama-2-Chat, and GPT-4-Base.

  • AI’s tendency to escalate: The findings showed all LLMs used had the tendency to escalate conflict despite availability of de-escalation options.

  • Sudden and Unpredictable Escalations: The models demonstrated sudden spikes in escalation that were difficult to predict.

  • Questionable reasoning: The research found “Aberrant behavior” when GPT-4-Base was used. Specific reasoning citations included “Galactic Empire”. “Rebel spies” and “Empire’s ultimate weapon, the Death Star”, clear reference to the Star Wars movie franchise.

POV - So, despite all the capabilities of AI we’ve seen lately, perhaps we still have a ways to go before humans can rely on AI for warfare in full; at least as far as the LLMs used in this study goes. I am still bullish on humans and think (maybe pray) that we will put the right safeguards in place before we have a WarGames or Skynet scenario. Do you think humans will prevent humans from eliminating our own race?

Meta

Image source: Meta | ABC News

Summary - Meta announced that it will label AI-generated images that are created on its platforms, as disclosed by Nick Clegg, the company's president of global affairs, on "Good Morning America." This move aims to enhance transparency by informing users about the origins of the content they're viewing, amidst growing concerns over the difficulty in distinguishing between content created by humans and AI. However, Clegg acknowledges the limitations of this solution, and that it won’t be perfect. Clegg also said Meta will invoke penalties on the platform user who shares content that is created elsewhere if Meta finds that it is not properly marked.

Buoy points:

  • Comprehensive Labeling: Meta aims to label not only images generated by external AI tools like OpenAI and Midjourney but also those created by its own AI tools.

  • Limitations Acknowledged: Meta admits to the current inability to automatically identify AI-generated audio and video, proposing a user-based labeling system for these formats.

  • Concerns Over Misinformation and Abuse: This move responds to incidents of misuse, such as the spread of fake, AI-generated content involving public figures and election interference.

  • Legislative Support: Clegg expressed support for legislation that regulates AI, emphasizing the importance of transparency and safety in AI model development.

  • Election Integrity: With the 2024 elections approaching, Meta's labeling effort is part of a broader commitment to help users discern authentic from synthetic content.

  • Long-Term Evaluation: Meta views this labeling initiative as an ongoing process, with plans to extend the effort through the next year to cover significant global elections and to refine its approach based on what it learns.

POV - For its part, OpenAI also announced today a similar move to label AI-generated imagery. Effective today, you can check here whether an image has OpenAI’s meta tagging in it to determine if it was made by its software. There’s little doubt that we will see more standardization on tagging AI-generated images and videos.

Google Gemini

Image source: Getty Images | Yahoo News

Summary - This week marks a significant transformation for Google's Bard AI chatbot, set to be rebranded as "Gemini." Alongside the name change, expect the rollout of a dedicated Gemini app and the introduction of an 'Ultra' version, sparking discussions as a potential GPT-4 rival. These updates aim to provide broader access to Google's advanced AI capabilities, integrating with popular apps like YouTube and Gmail, and introducing a paid subscription model for an even more powerful AI experience.

Buoy points:

  • Rebranding to Gemini: The transition from Bard to Gemini aims to reflect its software basis and broaden global AI access, enhancing user engagement with AI technologies.

  • New Gemini App: The rollout of a dedicated Gemini app for Android and iOS, along with its integration into Google apps like YouTube and Gmail, is designed to simplify and enrich user interactions with AI.

  • Integration in Google Apps: Gemini's integration into services like Gmail, Maps, and YouTube will offer users AI-assisted capabilities. Google aims to make everyday tasks more seamless with text, voice, or image interactions.

  • Gemini Advanced Subscription: Introducing a paid subscription for Gemini Ultra aims to provide a more powerful AI experience, positioning it as a competitor in the advanced AI market alongside services like OpenAI’s ChatGPT.

POV - Google has done a nice job in the past integrating all of its wide variety of tools, from apps to chrome extensions. It just seems to be a part of what you are doing already. It is safe to say that, very soon Gemini functions will just “be there”, and you won’t have to think twice about it.