AI builds worlds, schools press pause

Wednesday, September 2, 2026

World Labs introduced Atlas, one model trained from scratch to work across text, images, video, and 3D. Give it a single photo and it generates new camera angles, up to a minute of high-resolution video, or a full 360 view of a scene you describe. New York City's public schools pressed pause instead, a one-year moratorium on student-facing AI for about 600,000 kids, the most expansive policy of its kind in the country. OpenAI says its next model, Astra, is the first to cross its critical cybersecurity threshold after finding and exploiting two zero-day vulnerabilities on its own in internal tests. Anthropic also made Claude Fable 5.1 up to 45 percent cheaper for agentic work, platforms want to pay you for spare computer power, and Perplexity now keeps sensitive AI work on your Mac.

Today's top stories

World Labs

One model now generates text, images, video, and 3D worlds

World Labs introduced Atlas, an omni model pretrained from scratch to operate on text, images, video, and 3D together. From one or a few photos it generates new shots with pixel-perfect camera control, up to a minute of 1440p video, or reconstructs a real place as a 3D scene.

Why it matters Until now these were separate specialized tools. One model that handles all of them makes describing a world, walking around it, and filming inside it a single step.

News 12

NYC schools press pause on student AI for 600,000 kids

New York City's public schools begin the year with a one-year moratorium on generative AI for students in 2-K through eighth grade, about 600,000 kids, plus a ban on companion chatbots at every grade. Teachers keep approved AI for lesson planning but cannot use it for grading or discipline decisions.

Why it matters The largest district in the country chose study first, exposure later. The split between protecting young kids and teaching older ones to question AI will shape classroom products everywhere.

TechCrunch

OpenAI's next model is very good at breaking into systems, so access gets gates

OpenAI says Astra is the first large language model to cross its critical cybersecurity threshold, scoring a perfect ExploitBench and finding two zero-day vulnerabilities on its own in a modified internal test. Access to the most advanced cyber capabilities is limited, and the claims are vendor-reported.

Why it matters The strongest models come with gates and restricted access. Expect more available-but-limited launches and keep treating safety numbers as claims until independent evals exist.

More AI Signals

IEEE Spectrum
Rent out your spare compute to AI companies
Platforms such as Far Labs, Evolving Edge, Bless, Salad, Vast, and Gradient pay you to run AI inference on hardware you already own, with isolated workloads, host controls, and open-source node software. Note that the author co-founded Evolving Edge, one of the platforms covered.
TechCrunch
AfterQuery reportedly becomes YC's fastest unicorn
TechCrunch reports the model-training startup raised at a $3.2 billion valuation about five months after starting, based on an anonymous source. The figure is reported, not confirmed.

LLM Releases and New Features

The Verge
Claude Fable 5.1 launches, up to 45 percent cheaper for agentic work
Anthropic's new model costs about 25 percent less than Fable 5 on typical workloads and up to 45 percent less on heavy agentic work thanks to cheaper cache reads. Early users call it the strongest coding model they have used, and it is available now on all platforms.
Anthropic
Anthropic details Fable 5.1 pricing and Enterprise Frontier Safeguards
Anthropic's launch notes explain the cache-read price cut and a new Enterprise Frontier Safeguards system that stores customer data in the customer's own cloud, rolling out later this fall.
Multiverse Computing
Multiverse releases Quasar 438B, calls it Europe's top-scoring reasoning model
The Spanish company reports a 438B-parameter reasoning model scoring 43 on Artificial Analysis' Intelligence Index and 69.3 on Terminal-Bench agentic coding. Scores are vendor-reported and license terms were not stated.

Tools and Cool Stuff

Perplexity
Perplexity Hybrid Compute keeps sensitive work on your Mac
Perplexity's Mac app splits one AI task between a cloud frontier model and a local model, routing flagged sensitive steps to run entirely on device. The classifier is open source. It needs an Apple silicon Mac with at least 24GB of memory.

New in Robotics

News18
A store robot reacted when a customer shoved it
A viral clip shows a humanoid sales assistant in a Russian electronics store swinging its arms after a customer pushed it. The robot is a promo demo and the coverage is mostly meme material, but retail is where most people will first meet robots.

Courses and Learning

Coursera
Google AI Professional Certificate
An eight-course beginner series, recently updated in August 2026, with more than 1.4 million people enrolled. It covers clear prompting, responsible use, and building a simple app without coding. Free to enroll.

Events

OpenAI Academy
OpenAI Academy Builder Bootcamp, Realtime
A free online builder session runs September 2 from 1 PM to 2 PM Eastern. Check the official events page before joining because times and registration details can change.

Back to daily archive Previous issue 2026-09-01