Linghua Jin

Terminal Use (YC W26): Vercel for background agents. Terminal Use is hosting infrastructure for background agents that makes harness engineering easy. https://lnkd.in/gv99wN2t

51

2 Comments

I wrote an article on how we translate user-generated content at scale at Mercari. It summarizes two years of work with various models, LLMs but not only, and how cost evolved, reducing it 100x over the period. I also go over non-AI features and user experience. https://lnkd.in/gbpQAQmX

39

2 Comments

Developing a large ecosystem with both open- and closed-source components comes with some interesting challenges. Back when we were bootstrapping CKEditor, my purist mindset pushed us to develop each of our 30+ packages in separate repositories. The goal was to sort of "force" higher modularity and more responsible architecture for every component. As you can probably guess 😅, it did not scale. Later, we consolidated most of the packages into two monorepos – one open source, one closed source – while keeping a few granular packages separate where it still made sense. This helped a lot, but it didn’t fully solve the problem. We were still working across multiple repositories and kept running into issues like dependency coordination and lockfiles. Recently, we took another step forward. We migrated to pnpm and redesigned our overall approach. In this post, Filip Sobol explains the new setup implemented by the CKEditor 5 Platform team and what we learned along the way: https://lnkd.in/eHS28qxe

22

3 Comments

Honored to have been invited to fal and Henry Shi’s Lean AI Leaderboard Celebration Dinner, celebrating some of the leanest and fastest-growing startups in SF. Here are some of the most overwhelming trends that were discussed among high-growth companies working with small, high-output teams: 1/ Headcount is no longer used as a success metric. As Henry Ward said, “Hiring means we failed to execute and need help.” Headcount is not a flex measure anymore. People aren’t asking “How big is your team?” anymore. They’re asking “How many customers do you have?” 2/ They're finding ways to leverage AI across every function. Every engineer uses Cursor or Windsurf. Every GTM leader is in Clay. The most efficient teams are built AI-first across the board. Thank you to Henry Shi, Burkay Gur, and Gorkem Yurtseven for putting this all together!

67

Another great piece quoting Andrew Ng (oh, BTW, a HumanX speaker in SF 2026) that I could NOT agree more with. Coding isn't the bottleneck with our AI assisted tools, it's good product management. But even that understates the need - we need technical product managers who know both good design patterns for products, but have enough development experience to be able to coax the vibe-coding systems to actually write decent, enterprise-readyish systems. In my experience, you really need to be forceful with the tools at a relatively low technical skill level to ensure things like consistency, security and even basic performance are done properly.

16

2 Comments

Why does the second search feel instant when the first one takes 3 seconds? I built a local semantic search engine using EmbeddingGemma. The bottleneck? Generating embeddings—10-100ms per chunk of text. Twenty chunks takes under a second. Two hundred chunks? A few seconds. Two thousand? You're waiting. And every search regenerated every embedding. Wasteful. The fix is simple caching: save embeddings to a JSON file after the first indexing run. On subsequent searches, load from disk instead of regenerating. First run: generate and cache. Every run after: instant. The real insight: caching doesn't just make it faster—it changes how you use the tool. Without caching, you think twice before searching. With it, searching becomes nearly free. That's the difference between a tool you built and a tool you actually use. I teach by building in public and sharing what I learn. Full write-up with code:  https://lnkd.in/gJtgEQ7s  #AI #MachineLearning #BuildInPublic #AILiteracy

4

Evolve SDK v0.0.29 provides per-task / per-session cost-tracking across all the supported CLI agents: Claude Code, Codex, Gemini CLI, Qwen Code, OpenCode, Kimi CLI. https://lnkd.in/gdu2Ur-u https://lnkd.in/gURvwVxb Anthropic OpenAI GoogleKimi (Moonshot AI) OpenCode E2B Daytona Modal

1

OSS Bench for Agentic Coding.. in steps Cline-Bench: "A task that frontier LLMs struggle to complete are ideal candidates for cline-bench. If you find yourself having to intervene, or write code manually because cline was not able to complete the job - your task captures the failure boundary of state-of-the-art models and is precisely what cline-bench seeks to formalize." https://lnkd.in/gHcGYxiZ

6