Rethinking Pre-Training for Agentic AI with Aakanksha Chowdhery - #759

2025-12-17 · Source: The TWIML AI Podcast (formerly This Week in Machine Learning & Artificial Intelligence) · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems, Software Development & Engineering · Depth: Expert, extended

Summary

Aakanksha Chowdhery, a member of technical staff at Reflection and former lead for Google's PaLM and early Gemini pre-training efforts, argues for a fundamental rethinking of pre-training to achieve true agentic AI. She contends that current post-training techniques are insufficient for developing multi-step reasoning and planning capabilities required for agents. Chowdhery highlights limitations of next-token prediction for complex workflows and emphasizes the need to evolve attention mechanisms, loss objectives, and training data. Key areas for evolution include improving long-form reasoning over extended contexts, learning from "trajectory" training data that captures multi-step problem-solving, and developing loss functions that teach models to learn from failures and adapt to new tools. She stresses that scaling remains crucial for discovering emergent agentic capabilities like error recovery and dynamic tool learning, and that new benchmarks are needed to measure these advanced forms of intelligence.

Key takeaway

For AI Scientists and Research Scientists focused on developing advanced agentic systems, you should prioritize foundational changes in pre-training rather than solely relying on post-training techniques. Your efforts should concentrate on evolving attention mechanisms, loss objectives, and training data to foster long-form reasoning, planning, and dynamic tool learning, as these capabilities are emergent at scale and critical for next-generation AI agents. Consider contributing to new benchmark development to accurately measure these complex agentic behaviors.

Key insights

Agentic AI requires fundamental pre-training shifts, not just post-training, to enable multi-step reasoning and dynamic tool use.

Principles

• Pre-training must evolve beyond static benchmarks.
• Scaling is essential for discovering emergent capabilities.
• Quality curation of training data improves compute efficiency.

Method

Rethink pre-training by evolving attention mechanisms for long-form reasoning, designing loss objectives that teach multi-step planning and tool use, and incorporating high-quality, diverse "trajectory" training data.

In practice

• Explore attention mechanisms for long-term context reasoning.
• Develop loss functions for tool use and error recovery.
• Curate high-quality, diverse reasoning traces for training data.

Topics

• Agentic AI
• LLM Pre-training
• Attention Mechanisms
• Loss Objectives
• Training Data Curation

Best for: AI Scientist, Research Scientist, AI Engineer, Machine Learning Engineer, AI Researcher

Related on AIssential

• Rethinking Pre-Training for Agentic AI [Aakanksha Chowdhery] - 759 — Agentic AI requires a fundamental shift in pre-training, moving beyond static benchmarks to foster interactive capabilities.
• AI Trends 2026: OpenClaw Agents, Reasoning LLMs, and More [Sebastian Raschka] - 762 — LLM evolution focuses on post-training reasoning, tool use, and inference scaling to enhance problem-solving and efficiency.
• FOD#145: What 100,000 Subscribers Taught Us About the Future of Turing Post — AI development is shifting towards agentic systems, online learning, and efficient, self-improving models.
• Week Ending 4.12.2026 — AI progress demands better benchmarks and frameworks for real-world reliability, judgment, and continuous learning.
• Paper Digest: ICLR 2026 Papers & Highlights — The ICLR 2026 highlights reveal a strong focus on enhancing LLM reasoning, multi-modal capabilities, and robust AI agent development.
• ImportAI 449: LLMs training other LLMs; 72B distributed training run; computer vision is harder than generative text — LLM agents can autonomously fine-tune other LLMs, but human performance remains superior, with rapid AI progress and emerging reward hacking behaviors.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by The TWIML AI Podcast (formerly This Week in Machine Learning & Artificial Intelligence).