Microsoft Research's 'Mirage' Gives Video Generation Spatial Memory

Mirage, a video world model jointly developed by Microsoft Research and several universities, stores scene information directly in latent space instead of as pixel-based point clouds. This drastically cuts compute time and graphics memory while preserving spatial consistency even across long camera movements.

Tasks like remembering the layout of furniture at the far end of a hallway used to require enormous memory, but Mirage sidesteps the problem by working with latent representations. That said, tracking moving objects reportedly remains a challenge.

It looks like a meaningful step in video generation AI evolving from short clips into coherent filmmaking that understands 3D space.

Google Cloud Announces the 'Open Knowledge Format' — Standardizing Organizational Knowledge in Markdown

Google Cloud has announced the Open Knowledge Format (OKF) specification, which standardizes scattered internal documents into Markdown files plus YAML front matter. The aim is to make organizational knowledge portable for AI agents.

Essentially, Google has formalized the pattern Andrej Karpathy recently proposed as the 'LLM Wiki' into a minimal specification. Choosing the lightweight Markdown format keeps migration costs from existing document assets low — a practical advantage.

In enterprise AI, RAG accuracy depends heavily on how well structured the input data is. If OKF becomes a de facto standard, it could bring order to knowledge management practices that currently vary widely from organization to organization.

Heretic's 'Grimoire' — A Censorship-Resistant 9KB Model Backup

Version 1.4 of Heretic, an open-source uncensored LLM tool, has been released, introducing a backup system called 'Grimoire'. The mechanism is simple but interesting: it compresses everything needed to reproduce a model into a 9KB text file called `reproduce.json`, letting you collect all Heretic models on Hugging Face locally.

Even if a model is removed from the platform, this file restores it in about a minute. No recomputation is needed, and the weight files are even hash-verified.

The Heretic project has already received legal notices from Meta and has been covered critically in the media. It is also shifting toward decentralized infrastructure, including distribution via IPFS and opening a Matrix space.

The effort is a clear counterweight to the current concentration of model hosting on Hugging Face, and it deserves attention as an example of community-driven resilience design.

DeepSeek 4 Flash Now Runs on a Mac M3 Max (96GB)

Using antirez's ds4 engine and dedicated GGUF files, DeepSeek 4 Flash has been confirmed to run on an M3 Max Mac with 96GB of memory. With the `--ssd-streaming` flag and tuned Metal allocation, it achieves generation speeds of 11–13 tokens per second. Cold boot takes about 10 seconds, and time-to-first-token after caching is 3–5 seconds.

However, prefilling 36K tokens takes roughly 2 minutes 30 seconds, leaving practicality challenges for long-context processing. Even so, the fact that a large model runs on a 96GB Mac shows the reach of local LLMs steadily expanding.

DuckDuckGo's Founder on 'The Reality of AI Adoption'

DuckDuckGo founder Gabriel Weinberg has published a post titled 'No, Not Everyone Is Using AI for Everything,' and it is stirring debate.

The claim: there is a gap between media impressions and actual data on real-world AI usage. Recent survey data likewise shows that everyday consumer use of AI tools is not as high as reporting suggests, and the post backs up that trend.

It offers an important balancing perspective on overheated AI coverage.

Is the Cloud LLM Gold Rush Ending?

An Automato article, 'Cloud-based LLM gold rush is ending,' is drawing attention. In the context of Apple's WWDC and the Fable 5 embargo, it argues that the model of relying on cloud-based LLMs has reached an inflection point.

At the same time, people on Reddit report building 'local AI assistants' in the wake of Anthropic's shutdown incidents, a sign that concerns about cloud dependence are starting to translate into concrete action at the community level.

---