Runway Turns AI Video From Something You Wait For Into Something You Steer
Runway has unveiled a vision that could completely change the AI video generation experience. Rather than waiting for a finished clip to render, users would type a prompt and the video would stream in real time, letting them control it as it generates.
The concept rests on the company's world model, GWM-1, which generates video frame by frame. Beyond creative tools, Runway is also eyeing applications in fields such as robotics and autonomous driving. Whether generation is delivered as a “result” or as a “process” — that difference may end up forcing a rethink of the design philosophy behind video generation models themselves.
Daily AI Use in the US More Than Doubles in Six Months
According to a report in The Decoder rounding up surveys from Epoch AI and Ipsos, the share of US adults who use AI almost every day grew from 8% in March 2026 to 19% in August — more than doubling in six months. Those using it about once a week, meanwhile, fell from 17% to 10%.
In other words, occasional users are gradually shifting into daily users. The numbers suggest AI is moving from a tool for specific tasks to something woven into everyday life.
Simulated Students Reproduce Real Mistakes — Microsoft's StudentSim
Microsoft and the University of Illinois have built StudentSim, which recreates individual students from limited data to give AI tutors fast, low-cost feedback. The idea is to make tutor training more efficient by using simulated students that make realistic mistakes.
In tests involving 60 students across chess, English, and math, it outperformed GPT-5.4, and a chess tutor trained with StudentSim received the highest expert rating among the three versions compared. Simulating humans improves the quality of AI that teaches humans — a simple but telling result.
Big Tech Uses Guarantee Schemes to Keep $300 Billion in AI Risk Off the Books
A Financial Times report is drawing attention on Hacker News. The claim: Big Tech is leveraging mechanisms known as “guarantees” to keep an estimated $300 billion in AI-related exposure off its balance sheets.
The original article sits behind a paywall, so for now only the headline and summary-level details can be verified — worth keeping in mind. Still, with enormous investments in AI data centers continuing, how those commitments are treated in accounting is a question investors and regulators alike can hardly ignore.
Meta's Muse: Better at Watching Users Than Helping Them?
WIRED's review is scathing. Meta's app Muse inherits the company's approach of opting users into data collection for AI training, and it reportedly goes further — prompting users to share bank account, email, and passport information.
The review's verdict: excellent at surveillance, but no real help to users. How much information should you hand over in exchange for convenience? As AI assistants become standard equipment, differences in each company's data collection policies look set to become an important deciding factor.
Jev Meets Japanese Benchmarks — and a Debate Over Its Origins
TypeSafe's Jev returns typed judgments and calibrated probabilities in a single forward pass, without generating any text. Lately, the project appears to be shifting from a phase of proliferating implementations into one of validation and debate.
In Japan, articles reporting Japanese-language benchmarks for Jev-style open source projects have been appearing one after another. One article benchmarked Laya (ModernBERT-large, 421M), regarded as the most serious reproduction: it scored well enough for practical use in routing Japanese inquiries to the right department, but in urgency classification a bias was confirmed in which the first option was picked 0 times out of 300. A hands-on write-up titled “I stopped letting LLMs handle classification and judgment and tried Jev” was also published, reflecting hopes that it could be the answer for teams that have long struggled to parse free-form text.
At the same time, a debate over its origins has surfaced. On LocalLLaMA, a post arguing that Jev's similarities to Laya — a model by a developer who built the same architecture a year earlier — are “more than superficial,” and that TypeSafe has not cited or credited the original research, has stirred discussion. At the individual PoC level, an experiment was also published that puts a Jev-style decision head on Qwen2.5-1.5B-Instruct, trained on roughly 3,200 samples and running in about 28 milliseconds. It seems the next stage of the boom — verifying reliability and clarifying provenance — has begun.