Most orgs where search is important have this dysfunctional organizational process where people are siloed into ML and search separately. Where the search infra team owns infra, and the candidate retrieval process under strict latency constraints. Then, the ML people can write their single threaded Python pandas reranking routines copied from a notebook that Fred wrote in 2019. This ML “ranking layer” usually also involves a custom container image built on the tensorflow 1.14rc2 image to rerank 100 items with a DNN

Big news: I'm helping with @martinkl with a second edition of Designing Data-Intensive Applications! An early release of the first 3 chapters is now available (O'Reilly Learning subscribers only at this point) and we're hoping to finish it next year. learning.oreilly.com/library…

🔗 DataChain open-source release 🤖 AI-Driven Data Curation: Local models, LLM APIs 🚀 GenAI Dataset scale: Millions and billions of files 🐍 Python-friendly: Python objects instead of JSON Try it out github.com/iterative/datacha… 👇1/7

From building the data platform and Parquet at Twitter to using AI to make biology easier to engineer at Ginkgo Bioworks, @squarecog joins the show to chat about the early days of big data, the conversation that made him jump into SynBio, LLMs for proteins and more.

We will see that a lot of weird behaviors and problems of LLMs actually trace back to tokenization. We'll go through a number of these issues, discuss why tokenization is at fault, and why someone out there ideally finds a way to delete this stage entirely.