Databricks Adopts GLM 5.2 as Its Default Coding Engine — Opus 4.8-Level Performance at Lower Cost
Databricks has decided to roll out GLM 5.2, an open-source model from China, as its default coding engine across the entire company. After running benchmarks on its own codebase of several million lines, the company confirmed that GLM 5.2 delivers coding performance on par with Anthropic's Opus 4.8 while costing just $1.28 per task (versus $1.94 for Opus 4.8).
Databricks's larger message is the recognition that no single provider holds a lasting advantage. The company urges enterprises to build their own benchmarks rather than relying on public ones. As model selection becomes increasingly commoditized, evaluation in real-world environments appears to be more important than ever.
Ollama Raises $65 Million — 176K GitHub Stars and Nearly 9 Million Users
Ollama, the open-source tool that makes it easy to run LLMs locally, has raised $65 million in funding. The project has earned 176,000 stars and roughly 17,000 forks on GitHub, and its ability to let developers easily run AI on their own PCs has won broad support.
Demand for local LLM runtime environments is growing among both enterprises and individuals, driven by privacy, cost, and latency considerations. Ollama's growth clearly demonstrates the existence of a developer community looking to break away from cloud dependence.
Character.ai Enters the Micro-Drama Market — A New Format Where Viewers Can Talk to the Characters
Character.ai has begun producing its own AI-powered micro-dramas (short-form dramas). What sets the format apart is that viewers can actually chat with the on-screen characters — asking them questions or role-playing storylines. In effect, the company has integrated its core product, conversational AI characters, directly into entertainment content.
Micro-dramas are a fast-growing genre, particularly among younger audiences, and adding AI-driven interactivity opens the door to viewing experiences never seen before. It's a sign that the fusion of content production and AI has entered a new stage.
Automating the Design of Embodied Agents — A New Approach from AgentCanvas
Research published on Hugging Face Daily Papers proposes AgentCanvas, a method for automatically exploring the architecture design of embodied AI agents (AI that acts in physical and virtual environments).
Embodied agents are typically built by manually combining modules for perception, memory, planning, and action. This study instead treats the design process itself as a searchable "design space," taking an approach where a coding agent automatically improves the architecture. While the authors confirmed consistent performance gains across multiple existing agents, they also candidly report challenges unique to the embodied setting, such as "leakage" issues where ground-truth signals from the evaluator are inadvertently incorporated.
The shift from hand-crafted design to automated search could have a major impact on the efficiency of AI agent development.
Corporate Executives Bewildered by AI Invoices — The Wall of Metered Billing
Reporting by Futurism and discussions on Hacker News have shed light on the reality that corporate executives are facing unexpectedly large bills for AI usage. As pay-as-you-go (metered) models become the norm, usage is difficult to forecast, and cases of massive invoices arriving before anyone notices are piling up, critics point out.
When evaluating AI adoption, performance and features tend to get all the attention, but managing actual operating costs is becoming a serious challenge for enterprises. Building proper budget governance looks set to become one of the factors that make or break AI initiatives.
---