Daily brief · October 28, 2026

Google DeepMind announces Genie 4, a real-time world model with native audio

3 items that moved, with the numbers. Measured by SteadyVane, fact-checked, and graded before publishing.

Google DeepMind announced Genie 4, a real-time interactive world model with native audio

Hacker News · News · Genie 4: upvotes history

331 → 905upvotes on Hacker News in 16 h · 519 commentsabout 16× the typical pace of other tracked Hacker News stories

Genie 4 is Google DeepMind's new world model for real-time interactive simulation. It generates audio on the fly, including ambient sounds, music and interactive dialogue, and you can steer a scene in plain language, such as moving the camera, changing the weather or adding characters.

The announcement went from 331 to 905 upvotes on Hacker News in 16 hours - about 16 times the typical pace of other stories SteadyVane tracks on Hacker News. It has 519 comments.

Why it matters: Per the fact-check, it is so far a research preview for trusted testers, with broader access to Google AI Ultra subscribers and API partners to follow. For most people it is something to watch, not something to use this week.

Access so far is trusted testers only. Judge it when more people can try it.

Anthropic's privacy page: commercial data is not used for training by default, consumer chats can be

Hacker News · News · Is my data used for model training?: upvotes history

15 → 207upvotes on Hacker News in 8 h · 92 commentsabout 6.6× the typical pace of other tracked Hacker News stories

Anthropic's privacy page covers its commercial products such as Claude for Work and the API, and points consumer users to a separate article. By default it does not use inputs or outputs from those commercial products to train its models. Consumer plans work differently: Free, Pro and Max chats are used for training when the model-training setting is on.

The linked page went from 15 to 207 upvotes on Hacker News in 8 hours - about 6.6 times the typical pace of other stories SteadyVane tracks on Hacker News. It has 92 comments.

Why it matters: Which rules apply depends on the plan you are on. Commercial data can still be used if the customer reports feedback or bugs, or otherwise chooses to allow it. Consumer data may be kept for up to 5 years if you allow training, versus the existing 30-day retention if not. Check the training setting on any personal account you use for work.

Two plans, two sets of defaults. Read which one your work is actually on.

Survival LLM is a free offline AI assistant with an emergency library that runs on your own computer

Hacker News · Tool · Survival LLM: AI assistant and emergency…: upvotes history

106 → 362upvotes on Hacker News in 12 h · 143 commentsabout 12× the typical pace of other tracked Hacker News stories

Survival LLM is a free, open-source desktop app for Mac (Apple Silicon) and Windows. It searches an offline library of medical, emergency, how-to and encyclopedia articles, answers from what it finds, and shows the exact passage each answer came from. It works with no internet once set up, and has no account or tracking.

Its Hacker News post went from 106 to 362 upvotes in 12.2 hours - about 12 times the typical pace of other stories SteadyVane tracks on Hacker News. It has 143 comments, and its GitHub repo has 238 stars (saves).

Why it matters: An assistant that shows its source passage for every answer is easier to trust than one that does not, and that matters most when you cannot check anything online. On first launch the app picks a model that fits the computer's memory. The library comes in two sizes, about 1.5 million articles (8.8 GB) or about 500,000 (1.3 GB).

Showing the source is the feature. Offline is the setting where you cannot look it up elsewhere.

Sources: Hacker News · GitHub

← OpenAI releases GPT-6.1 Terra, while a free local-AI app…

All briefs · What is rising now