Thursday, April 16, 2026

Good Thursday, NOLA. Big day for mobile AI: Google's Gemma 4 now runs fully offline on iPhone, no internet required. Meanwhile, Claude had a rough day with widespread outages across all its services. On the tools front, ChatGPT can now work directly inside your Excel spreadsheets, and Google launched the Gemini app for Mac. Plus, a thought-provoking piece asks whether the local LLM ecosystem really needs Ollama.

AI Goes Offline (For Real This Time)

Google Gemma 4 Runs Natively on iPhone with Full Offline Inference

This is the real deal — Google's Gemma 4 model running entirely on-device on iPhone, no internet connection needed. It's not a demo or a beta feature; it's a working implementation that shows how far on-device AI has come. If you've been waiting for truly private, offline AI on mobile, this is a milestone. Discussion on HN has people excited about the privacy implications.
Hacker News

The Local LLM Ecosystem Doesn't Need Ollama

A contrarian take that's sparking debate: the author argues that Ollama, while popular, isn't actually necessary for running local models and may be adding unnecessary complexity. The piece walks through alternatives and makes a case for simpler tooling. Whether you agree or not, it's worth reading if you're running models locally — sometimes the popular choice isn't the best choice. Lively discussion on HN.
Hacker News

Tools & Releases

ChatGPT for Excel: AI Right in Your Spreadsheets

OpenAI just launched ChatGPT as a native Excel add-in. Now you can use ChatGPT directly in your spreadsheets without copy-pasting between browser tabs. For anyone doing data work, reporting, or analysis in Excel, this could be a real workflow upgrade. HN discussion here.
Hacker News

The Gemini App Is Now on Mac

Google's bringing Gemini to macOS with a native desktop app. It's the same pattern we saw with Claude Code — AI companies are racing to own the desktop experience. If you're in the Google ecosystem, this gives you quick access to Gemini without opening a browser. Discussion on HN.
Google DeepMind blog

Libretto: Making AI Browser Automations Deterministic

A new open-source tool that aims to make AI-powered browser automation more reliable and predictable. If you've tried building browser agents and hit the usual reliability issues, Libretto might be worth a look. It's designed to handle the edge cases that make AI automations fail in production. Show HN discussion.
Hacker News

When Things Break

Claude Had a Very Bad Day

Wednesday's brief was all about Claude Code's new features. Thursday? Elevated errors across Claude.ai, the API, and Claude Code desktop. Outages happen, but the timing is rough — right as people are adopting the new Routines feature we covered yesterday. HN thread tracking the outage.
Hacker News

Does Gas Town 'Steal' Usage from Users' LLM Credits?

An open GitHub issue asking whether Gas Town, a popular AI coding tool, uses customers' API credits for its own model training or improvement. It's a fair question and one that more AI tools should be transparent about. The discussion is worth following if you're using AI coding tools and wondering where your credits are going.
Hacker News

Big Moves & Business Angles

Gemini Robotics-ER 1.6: Google's Latest Robot Control Model

Google DeepMind released Gemini Robotics-ER 1.6, an updated model for controlling robots in real-world environments. This isn't consumer-facing yet, but it's a signal of where Google's investing — multimodal models that can see, understand, and act in physical space. Discussion on HN.
Google DeepMind blog

DeepL Now Wants to Translate Your Voice

DeepL, best known for text translation, is expanding into real-time voice translation. The company says its tech could integrate with Zoom, Teams, and other meeting tools for live translation. If it works as well as their text translation, this could be genuinely useful for international teams.
TechCrunch AI

Claude May Require Identity Verification in Some Cases

Anthropic published a support article explaining that Claude may now ask for identity verification in certain situations. The documentation is light on specifics about when this happens, but it's a notable change — especially for people using Claude for sensitive work. Discussion on HN.
Hacker News

Reality Checks

CPUs Aren't Dead: Gemma2B Outscored GPT-3.5 Turbo on the Test That Made It Famous

A reminder that bigger and more expensive doesn't always mean better. The tiny Gemma2B model outperformed GPT-3.5 Turbo on specific benchmarks, and it can run on a regular CPU. This is the kind of result that challenges assumptions about needing massive compute for useful AI. Worth reading if you're thinking about which models to use for production work.
Hacker News

What Claude Code's Source Revealed About AI Engineering Culture

A follow-up on the Claude Code source leak we've been tracking — this piece digs into what the leaked code reveals about how AI companies are actually building tools. It's less about the technical details and more about the culture and engineering practices. If you've been following the Claude Code story, this is a good next read.
Hacker News

Also

Today’s Sources