Why ML Engineers Still Need to Code: AGI, LLMs, and AI Alignment

The Role of Coding in Machine Learning

• 💡 AI code generation tools like Claude, Gemini, and ChatGPT are not yet capable of solving complex coding issues, as demonstrated by a reinforcement learning algorithm bug that multiple AIs failed to identify.
• 🔑 Understanding code deeply remains crucial for machine learning engineers and data scientists until Artificial General Intelligence (AGI) is achieved.
• 🧠 Looking at code, such as for multi-head attention, can significantly aid in understanding complex machine learning concepts, often more effectively than diagrams or mathematical explanations alone.
• 📝 The trend of using pseudocode in machine learning papers highlights code's utility as an explanatory tool in education and research.

The Path to Artificial General Intelligence (AGI)

• ⏳ Aurélien Géron has downgraded his AGI timeline from five years to a more cautious 5-10 years, influenced by the perceived plateau in current LLM capabilities.
• 📉 Current AI models possess broad knowledge but lack depth, exhibiting obvious omissions and a shallow understanding of concepts, suggesting a fundamental issue with their knowledge representation.
• 🧩 Researchers are exploring world models and higher-level representations, moving beyond pixel-level predictions to more abstract concepts, which could lead to more efficient and extrapolative AI.
• 🚀 This shift towards better representation could unlock advancements in continuous learning, prediction, and extrapolation, though the timeline remains uncertain.

Concerns and Hopes for Advanced AI

• ⚠️ Géron expresses hope that AGI development takes longer, as he believes humanity is not yet ready for its potential disruption.
• 📈 While acknowledging the potential benefits of AI in areas like medicine, he cautions against rapid, widespread automation that could cause too much disruption too quickly.
• 🤖 The ultimate question of whether ML engineers will be needed at all remains open, especially in a future with fully realized AGI.