First AI-Generated Feature Film "Hell & Grind" Premieres at Cannes
"Hell & Grind", a feature film created with the AI video generation tool Higgsfield, premiered at the Cannes Film Festival. It is the first time a feature film made entirely with AI has been screened at a major international film festival, signaling that AI's role in visual production has reached a significant milestone.
AI video generation has so far centered on short clips and experimental works, so completing a full-length film and showcasing it on an international stage is likely to spark debate not only about the potential of AI in the film industry but also about its impact on production workflows.
Google AI Pro Subscribers Get a Perk That Makes YouTube Ads Nearly Disappear
Google is giving "Google AI Pro" subscribers free access to a perk equivalent to "Premium Lite", which makes YouTube ads nearly invisible. The move appears to be a strategy to boost the value of its AI plans through integration across Google Group's services and to grow its subscriber base.
As AI services become increasingly commoditized, Google is doubling down on lock-in through integration with its broad portfolio of services, including its core strengths like YouTube and cloud offerings. For consumers, the richness of bundled services is becoming an important criterion when choosing an AI plan.
LLM Hallucinations Are a "Commitment Failure", Not a Lack of Knowledge
A research team in South Korea has published a paper titled "Hallucination as Commitment Failure". Challenging the conventional view that LLM hallucinations occur because the model doesn't know the correct answer, the paper shows that larger models are more prone to "committing" to wrong answers even when they do know the right one.
The study introduces the concept of "semantic answer availability", which aggregates token-level variants that express the same answer concept. Using this measure, they found that even when the correct concept is available to the model at generation time, larger models tend to commit in the wrong direction. This finding throws cold water on the simple expectation that bigger models are smarter and therefore hallucinate less.
LatentOmni: Unified Speech and Video Reasoning Without Going Through Text
LatentOmni proposes a new approach for multimodal LLMs that handle both speech and video. In conventional methods, Chain-of-Thought (CoT) — which converts speech and video information into text before reasoning — has been the dominant paradigm. But this process compresses continuous speech and video signals into discrete tokens, degrading temporal precision and allowing intermediate reasoning to be pulled by linguistic biases.
LatentOmni adopts an integrated approach that reasons directly over speech and video in the latent space, bypassing text-based CoT altogether. This aims for more accurate omni-modal understanding without losing the fine-grained information in video and audio.
PhysX-Omni: Physics-Based 3D Generation Unifying Rigid, Deformable, and Articulated Objects
PhysX-Omni is a unified framework for generating 3D assets with physical properties suitable for simulation use. Conventional 3D generation methods either ignore physical attributes or are limited to a single category — rigid, deformable, or articulated.
PhysX-Omni develops a new geometric representation optimized for vision-language models, directly encoding high-resolution 3D structures without compression. It also introduces PhysXVerse, the first general-purpose simulation-ready 3D dataset covering diverse indoor and outdoor categories. The authors further propose PhysX-Bench, an evaluation benchmark that comprehensively assesses assets across six attributes: geometry, scale, material, affordance, kinematics, and functionality description. Applications in embodied AI — such as robot policy learning and physical simulation — are expected.
Teaching LLMs to Predict the Success of Research Ideas
The paper "Teaching Language Models to Forecast Research Success Through Comparative Idea Evaluation" examines whether LLMs can predict the success of research ideas before experiments are run. As AI automates hypothesis generation in scientific research, a new bottleneck has emerged: evaluating and screening the hundreds of ideas it produces.
The study sets up a "comparative empirical prediction" task: given a specific benchmark goal and two candidate ideas, the model must predict which one will empirically succeed. By training LLMs on this prediction capability, the paper presents an approach for streamlining idea screening before running experiments.
GitHub Trends: Code Knowledge Graphs Resurface, AI Agents Dominate the Top Spots
Looking at GitHub's daily trends, "codegraph", which builds knowledge graphs of code, is once again drawing attention. Meanwhile, AI agent-related repositories dominate the top of the charts, suggesting that the developer community's interest is heavily concentrated on agentic AI.
On the local LLM front, articles testing just how practical an ultra-lightweight 0.5B-parameter model can be are also attracting attention. Against the conventional wisdom that "anything under 7B is barely usable", a report benchmarking the real-world performance of Qwen2.5 0.5B — a model far below that threshold — offers compelling reference data for anyone considering AI on edge devices.