<p>In this episode, Stanislav Khromov joins the Svelte Radio team to discuss his work on Svelte Bench, a</p><p>benchmarking tool that scientifically measures how well different LLMs understand and write Svelte 5 code.</p><p>The conversation explores the challenges of AI-assisted coding with Svelte 5, the development of the official</p><p>Svelte MCP (Model Context Protocol) that provides LLMs with documentation and auto-fixing capabilities, and</p><p>the surprising performance differences between major AI providers. Stanislav also shares insights on working</p><p>with AI tools, the future of local models, and the economics of AI coding assistants.</p><p><strong>Notes</strong></p><ul><li><a href="https://stanislav.garden">Stanislav Khromov</a></li><li><a href="https://khromov.github.io/svelte-bench/benchmark-results-merged.html">Svelte Bench</a><ul><li><a href="https://github.com/khromov/svelte-bench">GitHub</a></li></ul></li><li><a href="https://svelte.dev/docs/mcp/overview">Svelte MCP</a></li><li><a href="https://www.wheresyoured.at/costs/">Anthropic Spending VC Money</a></li></ul><p><strong>Unpopular Opinions</strong></p><ul><li>Kevin: Serverless is overrated</li><li>Stanislav: OpenAI has the worst models of the big three</li><li>Antony: Heat Pumps are bad<ul><li><a href="https://www.youtube.com/c/TechnologyConnections/videos">Technology Connections</a></li><li><a href="https://www.youtube.com/@TechIngredients/videos">Tech Ingredients</a></li></ul></li></ul><p><strong>Picks</strong></p><ul><li>Kevin: <a href="https://eu.kasaigrills.com/">Kasai Hibachi Grill</a> (now tested, and can confirm, it is PENG!)<br>Stanislav: <a href="https://store.steampowered.com/app/2062430/BALL_x_PIT/">Ball x Pit</a><br>Antony: Saunas By the Sea</li></ul>

Svelte Radio

Kevin Åberg Kultalahti

Benchmarking AI with Stanislav Khromov

OCT 30, 202573 MIN
Svelte Radio

Benchmarking AI with Stanislav Khromov

OCT 30, 202573 MIN

Description

In this episode, Stanislav Khromov joins the Svelte Radio team to discuss his work on Svelte Bench, abenchmarking tool that scientifically measures how well different LLMs understand and write Svelte 5 code.The conversation explores the challenges of AI-assisted coding with Svelte 5, the development of the officialSvelte MCP (Model Context Protocol) that provides LLMs with documentation and auto-fixing capabilities, andthe surprising performance differences between major AI providers. Stanislav also shares insights on workingwith AI tools, the future of local models, and the economics of AI coding assistants.NotesStanislav KhromovSvelte BenchGitHubSvelte MCPAnthropic Spending VC MoneyUnpopular OpinionsKevin: Serverless is overratedStanislav: OpenAI has the worst models of the big threeAntony: Heat Pumps are badTechnology ConnectionsTech IngredientsPicksKevin: Kasai Hibachi Grill (now tested, and can confirm, it is PENG!)Stanislav: Ball x PitAntony: Saunas By the Sea