Share
ConTejas Code
Kyle Corbitt, CEO OpenPipe: How to Fine-Tune Your Own Language Model (LLM)
Links
- Codecrafters (Sponsor): https://tej.as/codecrafters
- Wix (Sponsor): https://tej.as/wix
- OpenPipe: https://openpipe.ai
- Kyle on X: https://x.com/corbtt
- Tejas on X: https://x.com/tejaskumar_
Summary
Kyle Corbitt, founder and CEO of OpenPipe, shares the origin story of the company and his background in computer science and entrepreneurship. He discusses the evolution of machine learning and the breakthroughs that made OpenPipe possible. The conversation then dives into the process of fine-tuning models using OpenPipe, including the logging feature, data curation, and the selection of base models and hyperparameters.
The episode also explores the developer experience and the decision to create an SDK that is a drop-in replacement for the OpenAI SDK. The conversation explores the concept of overfitting in machine learning models and how it differs for language models. The validation process for fine-tuned models is discussed, including inner loop tests and outer loop evaluations.
Takeaways
1. OpenPipe was founded to help people transition easily and smoothly into fine-tuning models using machine learning.
2. The process of fine-tuning models involves logging user requests, curating data, selecting base models, and optimizing hyperparameters.
3. OpenPipe provides an SDK that is a drop-in replacement for the OpenAI SDK, making it easy for developers to integrate OpenPipe into their existing workflows.
4. The platform automates the heavy lifting of fine-tuning models, including the optimization of hyperparameters based on thousands of fine-tuned models and user-defined evaluations.
5. OpenPipe offers a seamless developer experience, allowing users to quickly and efficiently fine-tune models and deploy them for production use.
Chapters
00:00 Kyle Corbitt
03:28 The Origin Story of OpenPipe
14:34 Fine-Tuning Models with OpenPipe
33:46 Understanding Overfitting and Fine-Tuning
39:47 The Role of Hyperparameters
46:32 Validating Fine-Tuned Models
56:46 Enabling Tool Calls in Language Models
01:00:33 Unleashing the Full Potential of Language Models
01:05:09 Introduction to OpenPipe
01:10:14 Changing the Configuration Parameter
01:20:17 The Future of OpenPipe
01:25:31 The Need for a Founder's Handbook
01:32:17 Advice for Technical Founders and CEOs
More episodes
View all episodes
Julia Flament-Wallin: How to Build Maps of the World with AI
01:30:06|Links- Codecrafters (sponsor): https://tej.as/codecrafters- Julia's Talk: https://youtu.be/IFn2hMt480M?si=x0-2M2IBOASwaicz- TomTom: https://tomtom.com- Julia on LinkedIn: https://www.linkedin.com/in/juliawallin/- Tejas on X: https://x.com/tejaskumar_SummaryIn this podcast episode, we discuss the evolving landscape of AI engineering, data science, and data engineering. Julia and I explore the definitions and distinctions between these roles, delve into the intricacies of clustering and classification, and examine the role of MLOps in deploying machine learning models. Julia shares insights into her work at TomTom, highlighting the company's transition from hardware to software and the innovative data collection techniques they employ, including LiDAR technology and OpenStreetMap.Chapters00:00:00 Introduction00:11:46 Data Science and Data Engineering00:21:01 Role at TomTom and Road Furniture Features Detection00:34:18 Importance of Speed Limits and Fusion Algorithm00:43:19 Defining HD Maps and Their Importance00:54:16 Exploring Prototyping and Real-Time Updates01:03:02 Importance of Smaller Models01:19:30 Future of Mapping and AI in Transportation01:29:14 Lessons for Early Career ProfessionalsMax Stoiber: How to build startups that get acquired (and cache GraphQL well)
01:41:04|Links- Codecrafters (sponsor): https://www.codecrafters.io/- Stellate: https://stellate.co/- Max on X: https://x.com/mxstbr- Tejas on X: https://x.com/tejasqSummaryMax Stoiber, co-founder and CEO of Stellate, discusses his journey from creating popular open-source projects like React and Styled Components to building a company that provides GraphQL edge caching and CDN services. He shares his motivation for open-sourcing his projects and the benefits of sharing and connecting with others in the developer community. Max also explains the evolution of Stellate and how they identified the market need for GraphQL edge caching. He emphasizes the importance of understanding the market and customer pain points through structured interviews and categorization of feedback. Chapters00:00:00 Introduction00:08:16 Stellate and Spectrum00:16:50 Building Stellate with Edge Caching00:26:12 Focus on Business Problems in DevTools World00:34:37 Effective Outbound Marketing00:44:32 Innovations and Market Expansion00:53:11 Open Sourcing Entire Products01:01:52 GraphQL Queries and Caching Logic01:10:00 Technical Background and Problem Solving01:18:00 Finding the Right Co-Founder01:26:00 Challenges of Building a Startup01:34:02 Preparing for challenges in company buildingAlexander Lichter: How to Get the Most out of Vue.js and Nuxt
01:49:35|Links- Codecrafters (sponsor): https://tej.as- Alex: https://www.lichter.io/- Alex on YouTube: https://www.youtube.com/@TheAlexLichter- Alex on X: https://x.com/TheAlexLichterSummaryThis episode discusses a conversation with Alexander Lichter, a web engineering consultant and member of the Vue.js and Nuxt community. He shares his journey into content creation, including his YouTube channel and the DejaVue podcast, which he started to fill a gap in advanced Vue.js educational content.The discussion explores why Vue gained popularity, particularly among developers from non-JavaScript backgrounds, and examines Vue's technical architecture, including the transition from Options API to Composition API. Lichter also explains Vue's balanced approach between React's flexibility and Angular's strict structure, making it particularly accessible for developers coming from traditional web development backgrounds.Chapters00:00:00 Introduction00:13:20 Reasons for Choosing Vue.js and Comparison with React00:23:23 Introduction to Setup Function and Async Operations00:32:38 Limitation of Reactive with Scalar Types00:44:42 Vue.js maintaining its position through innovation00:52:56 Focus on Ecosystem and Community Impact01:02:30 Introduction to React and useState01:10:33 Introduction to Nuxt Config Directory01:19:22 Vue Directives and V-Model01:27:28 Migration Guide and Breaking Changes in Nuxt 401:41:05 Data Fetching and Composition API in Vue01:49:14 Cover and Review RequestAymen Ben Amor: How to Organize Developer Conferences
01:28:43|Links- Codecrafters (partner): https://tej.as/codecrafters- React Africa: https://react-africa.com/- React Paris: https://react.paris/- Aymen on X: https://x.com/aymen_ben_amor- Tejas on X: https://x.com/tejaskumar_SummaryAymen Ben Amor, founder of BeJS Conferences, shares his journey of starting and organizing successful tech conferences. He discusses the origin of BeJS Conferences and the inspiration he drew from attending his first conference. Aymen emphasizes the importance of creating a community and fostering connections among attendees. He also highlights the challenges of marketing and selling conference tickets, particularly in the post-COVID era. Aymen's commitment to diversity is evident in his efforts to ensure a balanced speaker lineup and provide opportunities for first-time speakers. He shares his vision of expanding conferences to underrepresented regions and cultures. Aymen's dedication to creating a memorable experience for speakers is reflected in the personalized gifts and activities he organizes. Chapters00:00:00 Introduction00:04:15 Founding BeJS00:13:39 The Speaker Experience00:19:57 Conferences in the Middle East00:27:28 Solving Diversity & Inclusion00:34:07 Repeated Speakers and New Opportunities00:41:55 Marketing: The Biggest Challenge00:50:09 How to be Confident on Conference Day00:57:55 Interesting Technology Aymen is Using01:05:13 Learning at Conferences01:12:45 Rethinking Conference Format01:18:33 Inspiration from other Conferences (GitNation)01:23:16 React Brussels01:27:42 ConclusionMatt Pocock: How to Maximize Success with TypeScript and Career
01:42:23|Links- Codecrafters (sponsor): https://tej.as/Codecrafters- Total TypeScript: https://www.totaltypescript.com/- Matt on X: https://x.com/mattpocockuk- Tejas on X: https://x.com/tejasqIn this conversation, Matt Pocock and I discuss the fundamentals of TypeScript and its benefits. We explore the concept of types in JavaScript and how TypeScript enhances the development process by providing a shorter feedback loop. The conversation dives into the topic of complex types in TypeScript and the challenges we explore. Additionally, Matt reflects on his journey from being self-employed to working full-time and then back to being self-employed. He discusses the intensity and drive he brought to his work, the importance of finding a direct link between effort and reward, and the satisfaction of seeing results. Matt shares his experience of working on XState and growing his presence in the tech community through teaching and creating content. He also talks about the wave of success he experienced and how it has now shifted, leading him to consider broadening his horizons. Matt and I discuss the role of hustle culture, the balance between work and personal life, and the drive for mastery.Chapters 00:00:00 Intro00:07:56 What are types?00:16:56 TypeScript and ESLint00:22:49 Misunderstood TypeScript Fundamentals00:31:35 Type-generated Documentation00:38:34 Most Complex Types00:47:42 When to use `any`00:54:30 ts-expect-error or ts-ignore01:01:32 Future TypeScript Wishes01:06:30 Matt's professional journey and rise01:14:09 Hustle Culture01:34:05 Imposter Syndrome01:39:53 ConclusionTC39: How signals work, adding signals to the JavaScript programming language
01:48:27|Links- Codecrafters (sponsor): https://tej.as/codecrafters- TC39 Signals Proposal: https://github.com/tc39/proposal-signalsSummaryIn this conversation, Kristen Maevyn, and Daniel Ehrenberg discuss reactivity and its importance in JavaScript. They explore the concept of reactivity, its role in keeping applications deterministic, and the challenges of achieving consistency in modern rendering frameworks. They also compare signals to observables and explain why signals are being considered as a language-level feature. The conversation touches on the API for signals, the use of classes in JavaScript, and the benefits of functional programming. Signals are an evolution of the long-standing problem of managing state in JavaScript applications. They provide a solution to the issues with classes and functions by combining the benefits of both. Signals allow for encapsulation, testing in isolation, and pure functions while still being able to handle state. Chapters00:00:00 Intro00:09:57 What are signals?00:26:37 Classes and the `new` keyword00:41:31 State Management and Signals00:49:25 Push-based vs. Pull-based Reactivity01:04:43 Language-level vs. Framework-level Signals01:14:27 Server-side Signals01:25:06 Self-referencing Signals01:36:29 Audience Q&A01:47:07 ConclusionPaul Klein IV: How to Build Browsers for AI Agents
01:33:25|Links- Codecrafters (sponsor): https://tej.as/codecrafters- React Africa (partner): https://react-africa.com/- Browserbase: https://browserbase.com/- Browserbase Series A Announcement: https://www.kleinerperkins.com/perspectives/browserbase-AI-seriesa/- Paul on X: https://x.com/pk_ivSummaryIn this conversation, we talk to Paul Klein IV, the founder of BrowserBase. We discuss the concept of browser-based software and its potential to automate tasks on the web. Paul explains that the future of software is one where it does the work for us, and Browserbase is designed to power this future by providing a programmable web browser for AI applications. We also touch on the challenges of running browser automation frameworks in production and how Browserbase solves these challenges. Paul introduces Stagehand, an open-source SDK that helps developers automate web workflows, and emphasizes the importance of ethical use of browser automation tools. Customers are building various applications on Browserbase, including web scraping, testing, AI agents, B2B software, and professional services. The company has experienced rapid growth, with thousands of hours of browser sessions per week. To handle the scale, Browserbase had to make architectural changes, such as moving from Postgres to ClickHouse for better performance. They also had to address challenges like connection management, resource allocation, and debugging customer issues. The founder believes they have achieved product-market fit based on strong customer retention and repeatable growth strategies. Browserbase is a technical product, so the number one priority for hiring is that candidates are technical and understand what Browserbase does. The team is currently focused on growing the top of the funnel and the sales edge side. They have specific roles for growth engineers and customer-focused engineers. The company recently raised a Series A round of $27M, with the goal of becoming a category-defining infrastructure company. The pricing model for Browserbase includes a hobby plan, a startup plan, and a scale plan. The company offers 10 free sessions for users to try out the product.Chapters00:00:00 Paul Klein IV00:04:41 Why Browserbase?00:11:57 What is Browserbase?00:19:09 Stagehand: the open source SDK00:35:30 How People are Using Browserbase00:49:17 Creating web APIs where there are none00:59:40 Product-Market Fit01:06:16 AI Engineering at Browserbase01:08:42 Remote Work vs. Onsite01:13:36 Browserbase's Series A Round01:20:02 AI Fatigue01:28:11 Advice for Founders01:32:10 ConclusionKelly Vaughn: How to Navigate Career Transitions and Build Meaningful Relationships
01:38:18|Links- Codecrafters (sponsor): https://tej.as/codecrafters- React Brussels: https://react.brussels- React Africa: https://react-africa.com- Kelly on X: https://x.com/kvlly- Spot AI: https://spot.ai- Tejas on X: https://x.com/tejaskumarSummaryKelly Vaughn, Director of Engineering at Spot AI, shares her unique career journey from self-taught coder to entrepreneur and now corporate leader. She discusses running a Shopify agency, co-founding a venture-backed startup, and transitioning to her current role. Key points include:- The challenges of running an agency and managing employee livelihoods- Insights on bootstrapping vs. venture funding for startups- The importance of genuine relationships and networking in career growth- Adapting to different company stages and cultures- Navigating the transition from entrepreneur to corporate employee- The value of diverse backgrounds in tech leadership rolesKelly emphasizes the importance of self-reflection, embracing change, and learning from failures throughout one's career journey.Chapters00:00:00 Intro00:05:08 Spot AI: Video Intelligence00:12:46 Dealing with Rapid Growth00:17:05 Getting into Tech00:22:25 TapRoom Consultancy/Shopify00:32:35 Running a Consultancy00:39:09 Venture Capital or no?00:40:08 Shutting Down Consultancy00:41:11 Attention, Effort, Mistakes as a Path to Growth00:47:26 The Pressure of Running a Business00:52:05 How to find a job: networking01:01:02 Being taken advantage of01:04:59 How Kelly got her audience01:08:31 Big audience problems01:12:01 Therapist and Engineering Leadership01:18:18 Boundaries01:26:06 Advice for FoundersAya Bochman, co-founder Fashn.ai: How to build a lean AI startup (that makes money)
01:32:46|Links- Codecrafters (sponsor): https://tej.as/codecrafters- Upvote Fashn on ProductHunt: https://www.producthunt.com/posts/fashn-virtual-try-on- Fashn: https://fashn.ai- Aya on 𝕏: https://x.com/ayaboch- React Africa (partner): https://react-africa.comSummaryAya Bochman, co-founder of Fashion AI, discusses her journey in creating a generative AI platform for virtual try-on in the fashion industry. She shares insights on building a self-funded startup, the challenges of developing AI models, and the importance of work-life balance for founders.Chapters00:00:00 Introduction to Fashion AI00:07:42 Demo and product description00:11:10 Technical challenges in virtual try-on00:17:48 Data collection and labeling process00:26:51 API development and pricing model00:34:45 Collaboration with other services (RunPod, Fal.ai)00:43:02 Dealing with content moderation00:46:46 Business model and pricing structure00:52:12 Incident with Meta and open-source code00:59:10 Future plans and potential competition01:03:13 Advice for technical founders01:20:12 Work-life balance and avoiding burnout01:29:51 Working with a spouse as a co-founder