WorldReasonBench: AI Video Generation Models Are "Pretty, but Don't Understand the World"

While the quality of AI video generation models has been improving rapidly, how well they understand the laws of physics and logical consistency is another matter entirely. The new benchmark "WorldReasonBench" was designed as a metric that evaluates physical and logical plausibility rather than image quality.

According to the test results, ByteDance's Seedance 2.0 took the top spot, beating out Veo 3.1 and Sora 2. Commercial models generally scored roughly twice as high as open-source alternatives. However, the "logical reasoning" category remained the lowest-scoring category for every model, indicating that the shift from merely generating pixels to true "world models" has not yet happened.

At this point, it seems fair to say that evaluating whether models can "generate only things that could actually happen" — not just how the footage looks — is becoming an increasingly important dimension of assessment.

OpenAI Acquires Voice Cloning Startup Weights.gg

OpenAI has acquired Weights.gg, a startup that offered a service for cloning and sharing celebrities' voices. The team of about six people is joining OpenAI, but there are reportedly no plans to release it as a standalone cloning product.

It remains unclear how the technology behind the service — which could imitate the voices of Taylor Swift, Donald Trump, and others — will be integrated into OpenAI's audio and multimodal strategy. That said, the deal appears to be driven largely by talent acquisition in the voice AI space.

Running 100 AI Agents for $1.3 Million a Month — OpenClaw's Large-Scale Experiment

Peter Steinberger, founder of OpenClaw, has revealed that his three-person team keeps roughly 100 Codex instances running at all times, with monthly OpenAI API bills reaching $1.3 million. He frames this as a "research investment" — an experiment to explore what software development looks like once you stop worrying about token costs.

The setup, with fleets of agents working in parallel to autonomously write code, review PRs, and find bugs, is drawing attention as a case study in applying AI agents to software development in earnest. Whether the costs are sustainable and the quality holds up, however, will have to wait for more data.

OpenAI Partners with the Republic of Malta — ChatGPT Plus for Every Citizen

OpenAI has formed a partnership with the Republic of Malta and launched an initiative to provide ChatGPT Plus to the entire population. The initiative is also said to include training programs that support practical AI skill-building and responsible use.

Expanding AI access at the scale of an entire nation is a rare move, and its ripple effects on other countries will be worth watching.

Japanese Companies' AI Businesses Make Headlines in Overseas Media

The Economist has published an article titled "Japanese Companies Are Making Money from AI," and it has also sparked discussion on Hacker News. How the business operations of Japan's AI-related companies — not always prominent in domestic coverage — look from an outside perspective makes for an interesting observation point.

Raspberry Pi Founder Warns That "AI Could Accelerate the Drift Away from Tech"

Eben Upton of Raspberry Pi has warned that the rise of AI could lead young people to avoid careers in technology, which could in turn have a negative impact on the economy. His concern is that a generation of young people may emerge who, watching AI take over tasks like coding, find it harder to see "the point of learning technology."

---