Hi Synapticians!

Today's top story hits particularly close to home since you probably know we work in insurance! And let us tell you, this development has been stirring quite the buzz in our circles.

Lloyd's market has just unveiled what we've been waiting for: the world's first insurance policy specifically designed to cover AI chatbot "hallucinations" and screw-ups. Now, before you think "wait, isn't AI insurance old news?", let us clarify!

AI insurance itself isn't new (Munich Re's been at it since 2018), but this is different. What makes Armilla's new policy groundbreaking is that it's the first "affirmative" coverage dedicated exclusively to chatbot errors. Think of it as the difference between having a general "tech problems" insurance versus one that specifically says "when your chatbot tells customers the wrong thing, we've got you covered."

The really clever bit? The coverage triggers are tied to the actual performance of the AI model itself. So if your bot starts hallucinating about return policies (looking at you, Air Canada), this policy kicks in for legal defense costs and damages.

Here's what's fascinating from our industry perspective: we're moving from broad AI performance warranties (2018-2024) to hyper-specific use-case policies in 2025. First chatbots, next probably AI Agents, then who knows? It's like watching insurance evolve in real-time.

And yes, as AI models get better, claims should theoretically decrease... whether that translates to lower premiums is another story entirely 😅 

Happy reading!

Top AI news

1. Lloyd’s Insurers Introduce AI Chatbot Error Coverage
Lloyd’s insurers have introduced the first policies dedicated to covering damages from AI chatbot errors. Developed by Armilla, these policies aim to protect businesses from lawsuits due to AI failures, such as incorrect responses or 'hallucinations.' The coverage includes legal fees and compensation, offering a safety net for companies using AI technologies. With examples like Air Canada’s chatbot error, this initiative seeks to encourage AI adoption by reducing associated risks. It marks a significant advancement in AI risk management, providing businesses with the assurance needed to integrate AI responsibly.

2. AI System Creates Stable LEGO Structures from Text Prompts
LegoGPT, developed by Carnegie Mellon engineers, uses AI to design stable LEGO structures from text prompts. The system adapts META's language model to predict brick placement, ensuring stability with a math-based module. Trained on a dataset of 47,000 stable designs, it adopts a recursive approach to refine structures, achieving a 98.8% stability rate. Robots were used for physical construction, demonstrating superior performance over other AI systems in creating stable 3D objects.

3. AI-Simulated Pitchers Revolutionize Professional Baseball Training
Joshua Pope and Rowan Ferrabee, alumni from the University of Waterloo, founded Trajekt Sports, which is revolutionizing baseball training with the AI-driven pitching simulator, Trajekt Arc. This system offers ultra-realistic batting experiences by replicating professional pitcher movements using AI and physics principles. Initially adopted by the Chicago Cubs, it is now used by 30 professional teams worldwide. The technology not only enhances players' physical skills but also their cognitive readiness, making it a significant asset in modern sports training.

Bonus. How AI Transforms Coffee Brewing Experience with Fellow Aiden
The article discusses how AI enthusiasts use the Fellow Aiden coffee maker to improve the brewing experience. By creating AI-driven recipes, the process allows for discovering new flavors and sharing brewing profiles, while emphasizing the human effort in coffee production. The initiative makes specialty coffee more accessible, though concerns about dehumanization exist. Overall, it highlights innovation and collaboration in the coffee community.

Meme of the Day

Theme of the Week

Small language model - The scientific paper.

This paper shows how large language models can be trimmed to fit edge hardware without crippling accuracy. Layer-wise Unified Compression prunes and quantizes each block, while selective back-prop slashes memory use. The approach runs a 7 B-parameter model 2.9× faster and 4× lighter than vanilla fine-tuning on a Jetson Orin. A solid proof that private, on-device generative AI is finally practical.

Stay Connected

Feel free to contact us with any feedback or suggestions—we’d love to hear from you !