Show Notes
Join the Tool Use Discord: https://discord.gg/PnEGyXpjaX
Unlock the power of AI fine-tuning for image and video models with Greg Schoeninger, CEO of Oxen.ai. In this episode, we explore how to move beyond simple prompt engineering to training custom open-source models for as little as $1. Discover the technical strategies for building high-quality datasets, the trade-offs between LoRAs and full fine-tunes, and how to achieve consistent characters and styles in generative video.
We dive into real-world examples, including how to generate massive product catalogs for a fraction of the cost of enterprise APIs and the story behind the viral "Isometric NYC" project built with Claude Code. Greg breaks down the entire AI lifecycle—from data curation and labeling with Vision Language Models (VLMs) like Qwen and Gemini to deploying efficient, specialized models that outperform general-purpose giants. Whether you are looking to optimize GPU costs, automate video workflows, or build your own AI tools, this conversation provides the blueprint for scaling your AI capabilities.
Links from the episode: Oxen AI: https://www.oxen.ai/ Fine Tuning Fridays: https://luma.com/oxen Isometric NYC Project: https://cannoneyed.com/projects/isometric-nyc
Connect with us https://x.com/ToolUsePodcast https://x.com/MikeBirdTech https://x.com/gregschoeninger
00:00:00 - Intro 00:01:47 - When to Switch from Prompt Engineering to Fine-Tuning 00:11:05 - How to Build a Dataset for AI Fine-Tuning 00:21:52 - LoRA vs. Full Fine-Tuning Explained 00:27:53 - Case Study: Fine-Tuned Qwen 3 VL vs. Google Gemini 00:36:33 - Case Study: Isometric NYC with Claude Code
Subscribe for more insights on AI tools, productivity, and fine-tuning.
Tool Use is a weekly conversation with the top AI experts.
More from Tool Use - AI Conversations
Check out www.ToolUsePodcast.com Are AI web browsers falling short of the hype? Enter Rover, a revolutionary open-source AI agent that unlocks full autonomous workflows directly on your website with just a single script tag. In this episode, Arjun and Bhavani from rtrvr return to dive deep into how Rover uses a unique DOM-only approach to navigate, click, and type with subsecond latency. This innovative method completely bypasses slow screenshot-based models and outdated playwright or puppeteer scripts. We discuss the engineering secrets behind building smart DOM trees, why they crushed web benchmarks, and how Rover provides a powerful alternative to Google's Web MCP by keeping users engaged on your site instead of handing traffic over to external AI tools. Plus, learn how you can trigger complex agentic workflows via simple URL queries and why open-sourcing this technology is a massive win for the developer community. Rover links:https://x.com/rtrvrairover.rtrvr.ai Github: https://github.com/rtrvr-ai/rover Rover Deep dive: https://www.rtrvr.ai/blog/10-billion-proof-point-every-website-needs-ai-agent Benchmark: https://www.rtrvr.ai/blog/web-bench-results Connect withus https://x.com/ToolUsePodcast https://x.com/MikeBirdTech https://x.com/rtrvrai 00:00:00 - Intro 00:03:05 - The DOM-Only Approach vs Screenshots for AI Agents 00:06:00 - Parsing HTML vs Markdown for Reliable LLM Data 00:11:47 - Handling Modals, iFrames, and Canvas Elements 00:15:32 - Implementing AI Guardrails and Extracting User Intent Data 00:28:04 - Rover vs Google Web MCP for Website Automation 00:31:25 - Triggering Autonomous AI Workflows via URL Queries Subscribe for more insights on AI tools, productivity, and web automation. Join the Tool Use Discord: https://discord.gg/PnEGyXpjaX Tool Use is a weekly conversation with the top AI experts.
Join the Tool Use Discord: https://discord.gg/PnEGyXpjaX
Join the Tool Use Discord: https://discord.gg/PnEGyXpjaX
Join the Tool Use Discord: https://discord.gg/PnEGyXpjaX