Show Notes
In this episode of Inference Time Tactics, Rob and Cooper from Neurometric sit down with Yash Sharma, an AI researcher whose work is reshaping how we understand model generalization. Yash recently completed his PhD at the Max Planck Institute for Intelligent Systems and has held research roles at Google Brain, Meta AI, Amazon, Borealis AI, and IBM Research. His studies on compositional generalization, adversarial robustness, and long-tail benchmarks reveal when and why models succeed—or fail—at reasoning beyond their training data. If you’re designing inference-time systems, building agents that need reliability, or just want to understand what “generalization” actually means in practice, this conversation bridges deep theory with actionable insight—clear, technical, and strategically grounded. Key Topics
What it really means for AI systems to generalize beyond their training data Why large language models still fail in novel or unpredictable scenarios How inference-time compute can both amplify and reveal generalization limits What these limits mean for building reliable, agentic AI systems How to benchmark generalization in real-world settings Yash’s “Let It Wag!” benchmark for testing long-tail and under-represented concepts Why genuine scientific breakthroughs (like curing cancer) require more than scaling test-time compute
Connect with Yash Sharma:
Yash Sharma Let It Wag! Benchmark Paper: Pretraining Frequency Predicts Compositional Generalization of CLIP (NeurIPS 2024 Workshop)
Connect with Neurometric:Website: https://www.neurometric.ai/ Substack: https://neurometric.substack.com/ X: https://x.com/neurometricai/ Bluesky: https://bsky.app/profile/neurometric.bsky.social
Rob May https://x.com/robmay https://www.linkedin.com/in/robmay
Calvin Cooper https://x.com/cooper_nyc_ https://www.linkedin.com/in/coopernyc
More from Token Engineering
In this episode of Token Engineering, Cooper sits down with Yash, Head of AI Research at Neurometric AI, to break down TMax from AI2 (Allen Institute for AI)—a fully open dataset and training recipe that pushes small open-weight models like Qwen to near-frontier performance on terminal agent tasks. The paper argues that diversity and difficulty matter more than where the data comes from—even when that data is entirely invented by Gemini rather than pulled from real-world code.
In this episode, Cooper sits down with Neurometric CEO, Rob May, to talk through a big-news month: the show's rebrand from Inference Time Tactics to Token Engineering, the launch of tokenminning.com as a response to the "tokenmaxxing" trend, Neurometric's new token engineering platform and fundraising announcement. But the bigger conversation is about why Rob thinks "token engineering" is about to become a discipline every AI company needs—the same way DevOps or UX became non-negotiable roles—and why he's betting the company on getting there first.
In this episode of Inference Time Tactics, Cooper sits down with Yash Sharma, Head of AI Research at Neurometric AI, to break down the Pioneer Agent paper from Fastino Labs—a system that uses Claude Sonnet as an ML engineer in a box to build and improve small language models end-to-end. From cold start data curation to production failure diagnosis, the paper argues the real bottleneck in SLM creation isn't training—it's everything around it.
In this episode of Inference Time Tactics, Cooper sits down with Fletcher Wimbush, CEO of Discovered AI, to get into what’s actually broken about hiring—and how 13 years of bootstrapped recruiting, 10,000 interviews, and four peer-reviewed journal articles became the foundation for a platform now outperforming its biggest competitors. With hundreds of applicants flooding every open role and general AI tools hitting the ceiling at 70-80% resume screening accuracy, Discovered is combining behavioral science with fine-tuned small language models to get hiring right the first time.
Every day billions of voice conversations happen across games, customer service calls, and financial transactions. Almost none of them are understood by machines. In this episode of Inference Time Tactics, Calvin Cooper and Yash Sharma sit down with Carter Huffman, CTO and co founder of Modulate, to explore the AI systems that can finally understand voice conversations in real time.