Parse notes
Lyft processes 100 million ML predictions daily through LyftLearn Serving, a platform that replaced a problematic monolithic service with isolated microservices where each team owns their own model serving infrastructure. The system handles real-time predictions for ride pricing, fraud detection, and ETAs using a Flask/Gunicorn HTTP layer with custom predict/load functions that support any ML framework.
API design must evolve for AI consumers, which combine machine precision with human-like understanding and create new constraints like token economics, high latency, and non-deterministic behavior. This means new APIs should be designed for self-healing through rich error messages, treating documentation as runtime code since AI references it continuously, and optimizing for batching and streaming to handle AI's multi-second processing times.
You usually know within the first week if a job is a good fit. Alignment for managers is especially important, as their role requires them to represent and believe in the company's values.
Unlike agentless LLMs that act like a virtual whiteboard, agents can use tools like bash, patch, and web search to navigate codebases, fix errors using compiler feedback, and improve API usage. While agents are great for handling larger codebases and automating tasks, they also have downsides like increased time and token consumption. Also, containers should be used to solve agent safety issues and allow for the parallel running of agents.
Fine-tuning advanced LLMs for knowledge injection is a waste of time and potentially destructive. Updating neurons in these models can overwrite existing valuable knowledge, leading to unintended consequences.
Sponsored
Lingo.dev is an open-source, AI-powered i18n toolkit designed for instant localization using LLMs. It offers a compiler for build-time React localization, a CLI for translating various content formats, CI/CD for automated translation workflows, and an SDK for real-time, per-request translation of dynamic content.
Rowboat is an AI-powered platform that lets users quickly build multi-agent workflows using tools, integrate them into applications via HTTP API or Python SDK, and use features like RAG and custom LLM providers.
Canva now expects engineering candidates to use AI tools during technical interviews, replacing traditional coding questions that AI can solve easily with complex, ambiguous problems. It evaluates how well candidates collaborate with AI strategically by breaking down requirements, critically reviewing generated code, and making sound technical decisions.
Sentry offers an MCP server, allowing LLMs to access valuable context about application issues directly from Sentry. The Sentry MCP has features like remote hosting, OAuth authentication, streamable HTTP, and various tool calls for accessing project information and issues, and even using Sentry's AI agent Seer. This lets LLMs provide more informed and accurate responses by using real-time Sentry data.
Anthropic recently cut off AI IDE Windsurf's access to its Claude models, betting that developers would follow the most capable models and move to Cursor. Anthropic did this as a move against its chief rival OpenAI, which is set to acquire Windsurf in a $3B deal.
React Hook Form is better for simpler forms and projects where stability is a priority, while TanStack Form is better for complex forms and large-scale applications.
The first major disaster caused by AI language models is likely imminent, potentially involving AI agents automating harmful actions like debt recovery or healthcare denial.
PostHog is creating DeskHog, an open-source, 3D-printed, palm-sized device packed with features for developers to hack, play games, and access PostHog data.
Markdown Ninja is a markdown-first, open-source CMS for bloggers, minimalists, and startups to easily publish websites and newsletters.