Friday, March 27, 2026

Google's new Gemini 3.1 Flash Live is now available across Google products, promising more natural and reliable audio interactions. This is Google's push to make voice-based AI feel less robotic and more conversational. If you've been frustrated with clunky voice AI, this might be worth testing.

Google's Search Live feature is now available in all languages and locations where AI Mode is available. This is the conversational search interface that lets you ask follow-up questions naturally. Big accessibility win for non-English speakers.

Google Translate's live translation feature for headphones is officially on iOS now, and expanding to more countries on both platforms. Pop in your earbuds and get real-time translation of conversations around you. Practical for travel or multilingual meetings.

Someone claims their $500 GPU setup beats Claude Sonnet on coding tasks. The GitHub repo is called ATLAS, though details are light. If true, this is a big deal for anyone who wants to run powerful coding assistance locally without subscriptions. Lots of skepticism and questions on HN about methodology, so take with a grain of salt until more testing.

This is legitimately clever: someone built an AI agent that lives on a cheap VPS and uses IRC (yes, the chat protocol from the '90s) as its communication layer. It's a great example of using old, reliable protocols for new AI workflows. The writeup walks through the whole architecture and why IRC actually makes sense here.

A new TypeScript library for extracting structured data from websites using LLMs. If you've been scraping sites and wrestling with brittle selectors, this takes a different approach by letting the model figure out what to extract. Worth a look if you're building anything that needs to pull data from varied web sources.

A team at Reco used AI to port JSONata (a JSON query language) from JavaScript to Go in a single day, claiming it saves them $500k annually in infrastructure costs. This is another case study in "vibe porting" — using AI to rewrite code in a different language without fully understanding every detail. Simon Willison has good context on what makes this interesting beyond the clickbait framing.

Callum McMahon shares his minute-by-minute response to discovering malware in the LiteLLM Python package, including the Claude transcripts he used to confirm the vulnerability and decide what to do. It's a fascinating look at using AI as a thinking partner during a security incident. Claude even suggested the right PyPI security contact.