Episode Summary
Executive Summary: The episode argues that data curation, not just prompting, is the critical bottleneck in improving LLMs and agents. Mahesh Satyamurti explains how Bespoke Labs uses synthetic data, RL fine-tuning, and evaluation tooling to teach models reasoning, tool use, and domain-specific behavior, with a focus on making these capabilities accessible to enterprise AI engineers.
Main Topics: Data-centric AI as the core thesis (Priority: 5/5): Mahesh traces Bespoke Labs’ origin to the belief that model gains come primarily from better data recipes, filtering, visualization, and evaluation rather than hyperparameter tuning. Curator and the role of data tooling (Priority: 4/5): Curator is positioned as an open-source library for high-throughput offline data curation, synthetic data generation, and visualization for fine-tuning and evaluation workflows. Why RL is re-emerging for LLMs and agents (Priority: 5/5): The discussion explains how modern LLMs’ world knowledge makes RL more practical than before, reducing the bootstrapping burden and enabling fine-tuning toward reasoning and tool use. Tool use and agent training (Priority: 5/5): RL is presented as a better fit than SFT for multi-turn tool-calling agents, because it requires far fewer labeled demonstrations and can reward actual task success. Open Thoughts, Open Thinker, and reasoning data (Priority: 4/5): Mahesh describes rapid dataset/model releases built around reasoning traces, including a 48-hour turnaround after DeepSeek R1, and the broader consortium effort to create high-quality reasoning data. MiniCheck and MiniChart as specialized models (Priority: 3/5): Bespoke’s smaller models show that curated synthetic data plus distillation can produce cheaper, faster models that outperform larger general-purpose systems on specific tasks. Democratizing advanced model adaptation (Priority: 4/5): The long-term goal is to give engineers enterprise-grade tooling for curation, RL, and agent building so they can avoid fragile prompt-only workflows.
Key Arguments: Data is the main driver of post-training improvements; model architecture matters, but most gains come from how data is curated, filtered, and evaluated. Prompting is brittle and often creates a 'prompt hell' of long instruction chains; RL can bake feedback into the model instead of relying on runtime prompt engineering. Evaluation alone is insufficient; error analysis and direct inspection of trajectories, JSON, and failures are necessary to improve systems. RL is especially useful for agentic tool use because it can learn multi-turn behavior from small numbers of prompts plus an environment and verifier, without tens of thousands of demonstrations. Modern LLMs already carry prior knowledge, so RL now works with fewer rollouts and less compute than older RL settings that started from scratch. Reward shaping matters, but simpler reward functions often work better and reduce reward hacking. Enterprise use cases are out-of-distribution for frontier models; RL + SFT can adapt open models to internal tools, APIs, and workflows. Specialized distilled models like MiniCheck and MiniChart can outperform larger models on narrow tasks while being cheaper and faster.
Data Points: Company size: Less than 10 people - Mahesh describes Bespoke Labs as a small team of two co-founders plus interns and full-timers. Company age: About 1 year - He says the company started about a year ago and that funding was raised at the end of April the prior year. Model turnaround time: 48 hours - Bespoke says it produced the Statos model and a reasoning dataset shortly after DeepSeek R1. Dataset usage: More than 200 models - The Open Thoughts dataset has reportedly been used to train over 200 models. Model size: 32B - The conversation references a 32B OpenThinker/Bespoke model and DeepSeek 32B comparisons. Model size: 7B - MiniCheck, MiniChart, and the Quen instruct base model example are all described as 7B-class models. Latency: ~100 milliseconds - MiniCheck is described as fast enough for guardrail use due to its low latency. Benchmark data size: ~200 examples - The BFCL tool-calling benchmark example is described as having roughly 200 examples. RL training data needed: ~100 good-quality examples - Mahesh says the tool-use RL setup worked with around 100 examples instead of tens of thousands for SFT. SFT data needed: Tens of thousands of examples - He contrasts RL with SFT for multi-turn tool use, noting SFT would require much larger labeled datasets. Rollouts per prompt: 8 - He explains gRPO as trying eight rollouts per prompt during training. Compute hardware: 4 H200 GPUs - The tool-use RL experiment was run on four H200s. Reward scale: +1 / -1 - Mahesh describes a simple reward scheme giving positive or negative reward for correct or incorrect outcomes. Accuracy example: 94% - He uses a sample evaluation figure to illustrate how evaluation is discussed, though not as a specific reported benchmark result.
Pivotal Quotes: "People want custom models and custom agents, right?" — Mahesh Satyamurti: He uses this to explain why data curation matters and why Bespoke Labs started with tooling around data and post-training. "Prompting only gets you so far." — Sam Charrington: This frames the discussion of why RL fine-tuning and curated data are needed beyond instruction prompting. "Reward shaping is the new prompting or new programming language." — Mahesh Satyamurti: He describes reward design as the key lever for controlling RL-trained models and agents.
Implications: The conversation suggests enterprise AI will shift from prompt engineering toward data curation, evaluation, and RL-based fine-tuning. Teams that can define rewards and environments may build better, cheaper, domain-specific agents than those relying on frontier-model prompts alone.