Welcome (back) to AI Minds, a newsletter about the brainy and sometimes zany world of AI, brought to you by the Deepgram editorial team.

In this edition:

• 🎥 Guardian News: Scamming phone scammers with “Daisy the AI Granny”

• 🧑‍🔬 Research: How much do teams really trust LLMs with risky tasks?

• 📡 AI Researchers find out that LLMs are aware of their learned behaviors

• ⚡ Webinar: Build Real-time Voice Agents with Deepgram in Vonage AI Studio

• 🐦 Twitter: Karpathy’s latest LLM Deep-Dive and Mistral’s newest AI Sidekick!

• 🤖 The intelligent humanoid robot race is heating up

• 📲 Three new, trending AI apps for you!

• 🎙️ AI Minds Podcast with Co-founder and CEO of Hamming AI, Sumanyu Sharma

• 🎞️ DeepSeek's GPU optimization tricks | Lex Fridman Podcast

• 🧠 Wired: 2025 is the Year of the AI App

• 💎How Diffusion Models Are Reimagining Game Environments: DIAMOND

• 📚Glossary Entry: Deep Reinforcement Learning

Thanks for letting us crash your inbox; let’s party. 🎉

🎥  Guardian News: Scamming Scammers with the “AI Granny”

Video Description: “O2 has introduced ‘AI granny’ Daisy for a short period to show what could be done with artificial intelligence to counter the scourge of scammers, who have become so ubiquitous.

Daisy is not a real grandmother but an AI bot created by computer scientists to combat fraud. Her task is simply to waste the time of the people who are trying to scam her. Using a mixture of ambivalence, confusion about how computers work and an eagerness to reminisce about her younger days, the ‘78 years young’ Daisy draws sighs and snapping from fraudsters on the other end of the line.”

🔍 Research: Trusting LLMs with Risky Tasks and their Learned Behaviors

An Empirical Study of User Trust and Team Performance When Using LLM Agents As A Daily Assistant - Although LLM agents have shown a promising blueprint as daily assistants, there is a limited understanding of how they can provide daily assistance based on sequential decision making capabilities. The authors of this paper conduct an empirical study (N = 248) of LLM agents as daily assistants in six commonly occurring tasks with different levels of risk typically associated with them.

Tell me about yourself: LLMs are aware of their learned behaviors - This paper studies behavioral self-awareness—an LLM's ability to articulate its behaviors without requiring in-context examples. The authors fine-tune LLMs on datasets that exhibit particular behaviors, such as (a) making high-risk economic decisions, and (b) outputting insecure code. Despite the datasets containing no explicit descriptions of the associated behavior, the fine-tuned LLMs can (surprisingly?) explicitly describe it.

⚡ Webinar: Build Real-time Voice Agents with Deepgram in Vonage AI Studio

Build Real-time Voice Agents with Deepgram in Vonage AI Studio - Discover how to build sophisticated AI voice agents by integrating Deepgram's cutting-edge speech-to-text and natural-sounding text-to-speech with Vonage AI Studio. 

You'll receive a step-by-step guide to building agents that understand domain-specific terminology and engage in natural conversations for seamless task completion. 

This webinar is perfect for developers looking to deploy production-ready voice agents with superior accuracy and responsiveness.

Hosted by:

• Benjamin Aronov, Developer Advocate at Vonage

• Tony Chan, Senior Solutions Engineer at Vonage

• Damien Murphy, Applied Engineer at Deepgram

When: Wednesday 26th March 2025, 10:00 PT / 12:00 ET / 17:00 GMT

Where: Online

Sign up here!

🐝 Social Media Buzz: Karpathy’s latest Deep Dive into LLMs

📲 Three new, trending AI Apps for you!

Grok AI Image Generator 2.0 simplifies art creation with a user-friendly interface and high-quality outputs. It features models like Flux.1 Pro and Flux.1 Schnell for exceptional image quality and fast processing.

ScriptFast is an AI-powered tool for creating professional-grade YouTube scripts quickly and efficiently. It simplifies scriptwriting into four steps: inputting an idea, drafting, getting AI suggestions, and exporting the script.

Summarify efficiently converts lengthy YouTube videos into concise summaries using AI technology. Summarify includes features like video timestamps, chapter summaries, and downloadable transcripts.

🎤 The AI Minds Podcast

Sumanyu Sharma, Co-Founder and CEO at Hamming AI, the platform automates AI voice agent testing, production call monitoring and governance for AI voice agents. The army of voice AI agents act like real people and place thousands of test calls simultaneously to find bugs, hallucinations and other issues often missed during manual testing.

🤖 Bonus Bits and Bytes!

• 🎞️ DeepSeek's GPU optimization tricks | Lex Fridman Podcast - Founder of SemiAnalysis, Dylan Patel, discusses DeepSeek’s GPU optimization techniques (read: tricks) alongside Ai2 research scientist Nathan Lambert.

• 🧠 Wired: 2025 is the Year of the AI App - Wired argues that “the real competition in AI isn’t about foundation models. It’s about apps built on top of those models.” But is that idea really true? Find out here!

• 💎How Diffusion Models Are Reimagining Game Environments: DIAMOND - Recently, researchers started exploring diffusion models’ ability to generate virtual worlds in real-time. We cannot deny the speed of innovation in the area and the impressive showcases that keep flooding in.

• 📚Glossary Entry: Deep Reinforcement Learning - Deep reinforcement learning (DRL) is a transformative branch of artificial intelligence that combines the intuitive nature of reinforcement learning (RL) with the analytical power of deep learning (DL).