Hey Synapticians,

Curious how Deepseek-R1 actually works? Want to understand how OpenAI's 03 beat the ARC-AGI benchmark? Want to understand how models went from struggling with basic addition to solving PhD-level math problems?

We've got you covered! Today's newsletter features an easy-to-understand guide on how modern AI models are trained - from their basic form to more advanced reasoning capabilities. We've kept it simple and jargon-free (no complex math or code), so everyone can understand what's happening under the hood.

Here's a little teaser: it all comes down to reinforcement learning - the art of teaching AI through trial and error. Come jump in with us to learn more about this wonderful world!

Hope you had a great weekend, Synapticians, and enjoy the read!

Top AI news

1. AI tutoring in Nigeria boosts student learning significantly
A pilot project in Nigeria demonstrated that AI tutoring, when combined with teacher support, can help students achieve nearly two years of learning in just six weeks. Participants using Microsoft Copilot showed significant improvements in English and digital skills, particularly among girls. Despite the positive outcomes, experts emphasize the crucial role of teachers and the need for further research on long-term effects and potential downsides before scaling such programs.

2. GitHub's Copilot introduces autonomous coding features for developers
GitHub has enhanced its Copilot AI assistant with 'Agent Mode', allowing it to work more independently by detecting and fixing errors autonomously. The new features include suggesting and executing terminal commands with developer approval. Additionally, Copilot Edits enables developers to modify multiple files using natural language, including voice commands. GitHub's upcoming Project Padawan aims to create a fully autonomous software development agent capable of handling issues and creating tested pull requests independently.

3. Apple's Robot Inspired by Pixar Magic
Apple's recent research paper emphasizes the significance of expressive movements in enhancing human-robot interaction. Drawing inspiration from Pixar, the study uses a lamp as a non-anthropomorphic example to illustrate how robots can engage more naturally with humans. The research suggests that incorporating qualities like intention and emotion into robot movements can foster a deeper connection. A video accompanying the paper showcases these concepts, hinting at Apple's growing focus on consumer robotics and potential future products, including a smart home system.

Bonus. AI System Revolutionizes Monitoring of Migrating Salmon Populations
Assistant Professor Sara Beery from MIT’s Department of Electrical Engineering and Computer Science is pioneering an automated approach to monitor migrating salmon in the Pacific Northwest. Traditional methods rely heavily on manual labor and sonar systems. Beery’s team is developing a computer vision system that autonomously detects and counts migrating salmon, achieving a counting error margin of just 3-5%. They’ve also introduced the “Fishbox,” a compact, energy-efficient computer enabling real-time data processing in remote locations without internet access. This technology is currently being deployed to monitor salmon migration in the recently restored Klamath River.

Tweet of the Day

Theme of the Week

Reasoning Models - AI concept
This article examines the transformation of AI from base models, which primarily complete text, to reasoning models capable of understanding context and following instructions. Key techniques include Supervised Fine-Tuning (SFT), which refines models using high-quality human conversations, and Reinforcement Learning from Human Feedback (RLHF), which aligns AI with human preferences. Though RLHF improves AI responses, it is challenging due to subjective quality assessments. The article highlights Deepseek-R1 as an example of these advanced models, showcasing their evolution and increasing reasoning capabilities.

Stay Connected

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