Rethinking Pre-Training for Agentic AI with Aakanksha Chowdhery - The TWIML AI Podcast (formerly This Week in Machine Learning & Artificial Intelligence) Recap
Podcast: The TWIML AI Podcast (formerly This Week in Machine Learning & Artificial Intelligence)
Published: 2025-12-17
Duration: 53 min
Summary
In this episode, Aakanksha Chowdhery discusses the need to fundamentally rethink pre-training methods for AI to enhance agentic capabilities. She emphasizes that current benchmarks and post-training techniques are limiting the potential of these models.
What Happened
The episode kicks off with host Sam Charrington welcoming Aakanksha Chowdhery, who brings a wealth of experience from her work on large language models like Palm and Gemini at Google. She shares insights into the challenges faced during the pre-training of these models, highlighting that as models scale, problems become magnified and can arise at any stage of the training process. Aakanksha's transition to Reflection, where she aims to develop frontier open intelligence for agentic capabilities, sets the stage for a deeper discussion on the evolution of AI training methodologies.
Aakanksha argues that traditional static benchmarks used for measuring pre-training are inadequate for developing models that can function as true agents. She points out that for AI to be genuinely useful, especially in tasks requiring interaction with environments, there needs to be a shift from merely post-training adjustments to a reimagined approach to pre-training itself. She cites examples of coding agents and deep research agents as early forms of agentic tasks that illustrate the need for models to possess capabilities such as planning, long-context reasoning, and the ability to learn and adapt in real-time. This rethinking of pre-training, she asserts, is crucial in order to advance the capabilities of AI beyond current limitations.
Key Insights
- Current pre-training methodologies are based on static benchmarks which limit AI capabilities.
- Agentic tasks require models to interact dynamically with environments rather than rely solely on pre-defined training.
- The evolution of AI models necessitates a balance between pre-training and post-training strategies for optimal performance.
- Long context reasoning and the ability to adapt to new tools are essential for the next generation of AI agents.
Key Questions Answered
What is Capital One's Chat Concierge?
Capital One's Chat Concierge is a multi-agentic AI that simplifies car shopping. It uses self-reflection and layered reasoning with live API checks to assist buyers in finding a car, scheduling test drives, getting pre-approved for financing, and estimating trade-in values.
What was Aakanksha Chowdhery's role in developing large language models?
Aakanksha was involved in building one of the distributed systems that led to the training of Palm, which was Google's largest language model at the time with 540 billion parameters. She discusses the complexities of pre-training, particularly the magnified problems that arise as models scale.
What is the focus of Reflection?
The mission of Reflection is to build frontier open intelligence for agentic capabilities. The company is focused on creating a post-training stack for agentic tasks and, following a recent fundraise, is now working on developing both pre-trained and post-trained models in-house.
Why is rethinking pre-training important for AI?
According to Aakanksha, rethinking pre-training is crucial because traditional methods rely on static benchmarks that do not adequately prepare models for real-world agentic tasks. To be useful, models must interact with environments dynamically and possess capabilities like planning and long-context reasoning.
How do agentic tasks differ from traditional AI tasks?
Agentic tasks involve goal-oriented interactions where models are expected to accumulate context over time and adapt based on feedback. Unlike traditional chatbots, models handling agentic tasks must be capable of complex reasoning and learning from their actions in real-time.