02 Oct 2025

AI Glasses Race Is On! 'Thieves' Train AI; Minecraft Mastered w/ World Model

Apple Chases Meta In AI Glasses Race

Apple is quietly shelving plans for a cheaper Vision Pro headset and shifting teams toward something it hopes will be more mainstream: AI-powered smart glasses. According to Bloomberg, Apple now has multiple glasses prototypes in the works, targeting a 2027 release.

There are two main ideas on the table:

A tethered version that connects to your iPhone and acts as a lightweight accessory.
A standalone version with a built-in display, aimed directly at Meta’s Ray-Ban “Display” glasses.

These glasses are expected to lean on Apple’s long-promised Siri upgrade for natural voice control, plus built-in speakers, cameras, and health tracking sensors.

Why it matters: The Vision Pro wowed people at launch in 2023 but quickly became a $3,499 curiosity - too heavy, too pricey and too niche. Meanwhile, Meta’s low-profile Ray-Bans have found an actual audience. If Apple can finally modernize Siri, its glasses could make AI assistants truly wearable. But without that breakthrough, Apple risks being seen as late to the party.

Mira Murati’s Startup Aims To Make Custom AI Easy

Former OpenAI CTO Mira Murati has launched Thinking Machines Lab, and its first product is Tinker - an API that helps developers fine-tune existing large AI models for specific tasks without massive costs.

Normally, training a big model from scratch takes millions of dollars. Tinker sidesteps that with LoRA adapters - lightweight plug-ins that let you add new skills to an AI without rebuilding it from scratch.

Universities like Stanford, Princeton, and Berkeley are already testing it for math, chemistry, and science reasoning.
Tinker will be free at first, with paid tiers coming soon.

Why it matters: This flips the AI race from “who can build the biggest model” to “who can make it easiest to personalize.”

🧩 Jargon Buster: LoRA adapters are like snap-on accessories for AI models. Instead of retraining the whole brain, you just add a little extension that teaches it new tricks.

Read more: AI Secret, The Rundown AI

Google DeepMind’s Agent Learns Minecraft Entirely In Simulation

DeepMind built Dreamer 4, an AI that learned Minecraft tasks - like collecting diamonds - entirely inside its own mental simulation of the game.

Instead of playing Minecraft directly, it practiced in a world model, basically a predictive digital twin of the game’s rules and physics. After simulating 20,000 practice actions, it applied what it learned to the real game - and beat rival agents while using far less data.

Why it matters: If AI can safely practice in simulated worlds before moving into real factories or labs, it speeds up development and cuts risk.

🧩 Jargon Buster: A world model is like an AI’s “daydream version” of reality - it imagines how the world should work and tests strategies there before trying them in real life.

Anker Pays You To Play Thief

Anker’s Eufy smart cameras quietly ran a campaign that literally paid users $2 a clip for videos of package and car thefts ... staged or real. People even dressed up and acted like burglars to generate content.

The goal: collect 40,000 clips to train the company’s AI to spot crime more accurately.

The initiative blew past its target, with hundreds of thousands of videos “donated”, and one top contributor claiming over 200,000 uploads.
Anker gamified the process with an honor wall celebrating its most prolific contributors.

Why it matters: On the surface, this looks like creative crowdsourcing. But it’s really about industrializing consumer surveillance into raw training fuel. Suddenly, your doorbell cam is a data-labeling shop, your baby monitor is a dataset, and your neighbors are unknowing actors in a corporate training loop.

Other platforms like Amazon’s Ring or Google Nest may feel pressure to follow suit - but the privacy risks are radioactive. Anker has already faced an encryption scandal in the past, so building a data marketplace around intimate home footage adds serious liability.

🧩 Jargon Buster: Training data is the raw material AI needs to learn - like practice rounds for a musician. The more examples it gets, the sharper it becomes. But when that “practice” comes from your living room or front porch, the trade-offs get messy.

Read more: AI Secret, The Rundown AI

Google Turns Search And Smart Homes Into AI Platforms (Update)

Google upgraded AI Mode in Search: now you can just type, “Show me weekend jeans for fall” instead of manually filtering by brand/size/color. Gemini 2.5, Google’s latest model, reads style cues in images, combines them with text, and surfaces shoppable links.

Google also refreshed its Nest cams, doorbells, and smart speakers, all powered by Gemini.

🧩 Jargon Buster: A context window (mentioned in model specs like GLM-4.6’s “200k context”) is how much information an AI can “hold in its working memory” at once. Bigger windows mean it can analyze longer documents, conversations, or videos without losing track.

Other Notables

OpenAI Sora app surged to #3 on the App Store, just behind Google Gemini and ChatGPT.
Hume AI launched Octave 2, a text-to-speech system supporting 11 languages with voice conversion tools.
Disney forced Character.AI to remove Elsa, Moana, Spider-Man, and Darth Vader.
Pew Research found 9% of U.S. adults get news from AI, with half reporting inaccuracies.
Zhipu AI released GLM-4.6, an open-source model with a 200k context window that tops Anthropic’s Claude Sonnet 4 on some benchmarks.
Anker paid users $2 per clip to upload staged “porch pirate” thefts to train its Eufy cameras.
Vercel raised $300M at a $9.3B valuation, positioning its Next.js framework + v0 agents as the backbone of an “AI Cloud.”

Tools & Launches

A few highlights from today’s drops and tool chatter:

Sora 2 - In case you slept through yesterday... OpenAI’s AI video generator now supports synced audio, longer 5–10 sec shots, and a new “Cameos” feature where you can insert your likeness into AI-made clips.
GLM-4.6 - A Chinese open-source model boasting huge context windows and strong reasoning.
Octave 2 - A next-gen multilingual speech model with editing tools.
Everyday - Lets you offload routine app tasks to coordinated AI agents.
Verdent Deck - Teams of AI agents handle complex coding jobs in parallel.
Cursor 1.7 - A developer environment with new autocomplete, team rules, and shareable prompt links.
Ask Brave - A privacy-first search tool that combines AI chat with grounded web results and wipes your history after 24 hours.

Today’s Sources: AI Secret, The Rundown AI, TAAFT