Amplify Partners

The 7th Amplify Monthly, covering everything from how Temporal Technologies became Temporal to portfolio news and open technical roles you won't want to miss.

Some (many?, most?) AI tools and platforms seem to disregard or misunderstand what drives creative people to create things. Thanks to Sarah Catanzaro and Justin Gage for articulating that point in this excellent article. https://lnkd.in/gZUrfAhp

How Vanta deeply integrates domain experts into their evals process: https://lnkd.in/gmcjA6tJ

How do you build AI *into* tech products? This week's article on Refactoring draws from dozens of interviews to founders and AI operators! Barr Yaron wrote a fantastic piece that puts together everything she learned on her podcast + her work with founders at Amplify Partners. Here are a few takeaways: ↳ 🎯 𝗗𝗼𝗺𝗮𝗶𝗻 𝗲𝘅𝗽𝗲𝗿𝘁𝗶𝘀𝗲 𝗶𝘀 𝘁𝗵𝗲 𝘀𝗲𝗰𝗿𝗲𝘁 𝘁𝗼 𝗴𝗿𝗲𝗮𝘁 𝗲𝘃𝗮𝗹𝘀 — the best AI products integrate subject matter experts directly into engineering workflows, not as occasional reviewers. ↳ 🔄 𝗥𝗲𝗮𝗹 𝘂𝘀𝗲𝗿𝘀 𝗮𝗿𝗲 𝘆𝗼𝘂𝗿 𝘂𝗹𝘁𝗶𝗺𝗮𝘁𝗲 𝘁𝗲𝘀𝘁 — ship early and learn from actual usage patterns, implicit feedback, and A/B tests; public benchmarks rarely capture how customers actually use your product. ↳ ⚖️ 𝗕𝘂𝗶𝗹𝗱 𝘃𝘀 𝗯𝘂𝘆 𝗶𝘀 𝗮𝗯𝗼𝘂𝘁 𝗰𝗮𝗽𝗮𝗯𝗶𝗹𝗶𝘁𝗶𝗲𝘀, 𝗻𝗼𝘁 𝗷𝘂𝘀𝘁 𝗺𝗼𝗱𝗲𝗹𝘀 — focus on understanding what today's models can do, what tomorrow's will likely solve, and where your unique advantage lives. ↳ 🧠 𝗛𝗶𝗿𝗲 𝗿𝗲𝘀𝗲𝗮𝗿𝗰𝗵𝗲𝗿𝘀 𝘄𝗵𝗼 𝗰𝗮𝗻 𝘀𝗵𝗶𝗽 — if you need research capabilities, look for people who blend algorithmic depth with engineering pragmatism and can adapt quickly. ↳ 🚀 𝗧𝗵𝗲 𝗳𝗮𝘀𝘁𝗲𝘀𝘁 𝗹𝗲𝗮𝗿𝗻𝗲𝗿𝘀 𝘄𝗶𝗻 — AI advantage comes from building organizations that adapt as quickly as the models themselves, with clear strategic bets about where lasting differentiation truly lives. I put the link in the comments!

After dozens of interviews for Barrchives, I wanted to know: what separates AI teams that ship from those that stall? Three themes came up again and again: 🔍 Getting evals right — by integrating domain expertise ⚖️ Build vs buy — when owning the model actually matters 🏗️ Team structure — do you need researchers, engineers, or both? Everyone talks about shipping AI products. These are the people actually doing it. I wrote more about what I learned as a guest on Refactoring — link in the comments.

The acclaimed Barr Yaron has interviewed 20+ leaders on how to successfully build AI into your products - nicely wrapped up into a few lessons, give it a read:

An inside look into the insane PB scale data curation pipelines DatologyAI have built:

Vortex is the fastest open-source columnar file format, full stop, by every metric. Out of the box, Vortex is 5x faster to write, 20x faster to scan, and 100x faster random access than the best Parquet implementations while maintaining the same size; 2x faster than the MDS format for GPU data loading; faster than Lance for both random access and scans while being 5x smaller & faster to write. For Clickbench on NVMe, Vortex is even faster than duckdb's native format when queried from duckdb. And now, Vortex has experimental support for direct GPU on-device decompression, promising further dramatic speedups for AI & GPGPU workloads. Best of all, it's Apache-2.0 licensed, neutrally governed at the Linux Foundation (LF AI & Data Foundation ), and FOSS Forever. But what few people understand is how extensible & customizable Vortex is. Last week, I gave a talk to CMU Database Group's Future Data Systems seminar series on how Vortex is really the "LLVM of File Formats", obviating the need for custom or specialized file formats entirely. Check out the recording here: https://lnkd.in/eQzjbwfV