Generative AI in the Real World: Sharon Zhou on Post-Training

Why It Matters

Post‑training determines whether generative AI can be safely and efficiently integrated into business workflows, directly impacting ROI, compliance, and competitive differentiation.

Key Takeaways

• Post‑training transforms raw models into usable, task‑specific intelligence.
• Enterprises often outsource fine‑tuning due to GPU and infrastructure costs.
• Retrieval‑augmented generation requires teaching models to fetch relevant data.
• Supervised fine‑tuning remains viable for low‑latency, private models.
• Reinforcement learning for post‑training is emerging but lacks turnkey services.

Summary

The conversation centers on post‑training—techniques that adapt large language models after their initial pre‑training—to make them practical for enterprise use. Host Ben interviews Sharon Zhou, VP of AI at AMD, to unpack how these methods turn raw intelligence into usable, dialogue‑capable systems.

Zhou explains that post‑training adds tool use (calculators, search APIs), step‑by‑step reasoning, and domain‑specific behavior, dramatically reducing hallucinations. She contrasts in‑context prompting with supervised fine‑tuning, noting that fine‑tuning yields lower latency and privacy on smaller models, while retrieval‑augmented generation (RAG) teaches models to pull relevant documents at inference.

A memorable quote: “ChatGPT’s breakthrough was not its pre‑training but its post‑training.” Zhou also cites a pizza‑store example to illustrate how undesired model traits can be corrected through post‑training. She highlights emerging services like the Tinker API that democratize distributed fine‑tuning, though reinforcement‑learning‑based post‑training still lacks turnkey platforms.

For enterprises, the takeaway is clear: evaluate whether to build in‑house post‑training pipelines or partner with specialized vendors, especially given GPU costs and the need for robust evaluation frameworks. Mastering these techniques will enable firms to customize model behavior, improve reliability, and maintain competitive advantage as generative AI capabilities evolve.

Original Description

Post-training gets your model to behave the way you want it to. As AMD VP of AI Sharon Zhou explains to Ben on this episode, the frontier labs are convinced, but the average developer is still figuring out how post-training works under the hood and why they should care. In their focused discussion, Sharon and Ben get into the process and trade-offs, techniques like supervised fine-tuning, reinforcement learning, in-context learning, and RAG, and why we still need post-training in the age of agents. (It’s how to get the agent to actually work.) Check it out.

Follow O'Reilly on:

LinkedIn: https://www.linkedin.com/company/oreilly/

Facebook: http://facebook.com/OReilly

Instagram: https://www.instagram.com/oreillymedia

BlueSky: https://bsky.app/profile/oreilly.bsky.social