World Action Models — How Robots Predict the Outcomes of Their Actions Before Acting
A joint research team from Fudan University, the Shanghai Innovation Institute, and the National University of Singapore has published a survey paper that systematically organizes a new family of models called "World Action Models" (WAM), which tackle a fundamental weakness in robot AI.
Conventional robot AI simply learned the mapping from camera images to motions, without understanding how its actions would change the environment. WAM changes that. By predicting environmental changes before acting and linking those predictions to action generation, it improves the ability to handle unfamiliar objects and situations.
Two Architectures
The research team organized roughly 100 papers into two lineages.
Cascade WAMs**: First, a world model generates images or videos of the next state, and a separate module derives control commands from them. UniPi, AVDC, and LAPA belong to this lineage.
Integrated WAMs**: A single model handles both future frame generation and action generation. GR-1, GR-2, and WorldVLA treat images and actions as a unified token sequence, while PAD and DreamZero process them in parallel with a diffusion-based approach. Nvidia's Cosmos Policy is said to be able to serve three roles—controller, simulator, and evaluation model—within a single architecture.
The Key: Learning from Everyday Video
WAM is drawing particular attention because it can learn from everyday egocentric video (such as Ego4D) that contains no robot manipulation labels. For conventional robot AI, this kind of data was essentially unusable. The RDT2 dataset collected approximately 10,000 hours of footage using this approach.
The challenges, however, are clear. DreamZero tops out at about 7 predictions per second, whereas conventional robot controllers can manage around 50. Evaluation methods also center on visual quality, and metrics for measuring physical plausibility are lacking. On the "Wow, Where, Val!" benchmark, many models that produce visually convincing videos achieve near-zero success rates when actually used for control.
Meta's V-JEPA 2, which predicts only abstract future representations instead of directly generating pixels, shows that computational costs could be reduced substantially. The results appear to back up the effectiveness of the JEPA approach advocated by Yann LeCun.
---
Microsoft's AI Chief: "White-Collar Work Will Be Automated in 18 Months" — But Reality Is More Complicated
Mustafa Suleyman, CEO of Microsoft AI, has sparked debate by predicting in a Financial Times interview that "over the next 12 to 18 months, the majority of the work done sitting in front of a computer will be fully automated by AI." He cited accounting, legal work, marketing, and project management as vulnerable areas.
Citing the exponential growth of compute as his basis, Suleyman remarked that "creating a new model will become like making a podcast," and positioned achieving super-reliance as his core mission. He also revealed that Microsoft intends to reduce its dependence on OpenAI and develop its own state-of-the-art foundation models.
The Gap Between Prediction and Reality
Actual data, however, does not show a shift as dramatic as Suleyman's prediction. According to Thomson Reuters' 2025 report, lawyers and accountants have begun leveraging AI for document review and routine analysis, but productivity gains have been limited and mass job losses have not materialized.
Research by the nonprofit METR has also reported that AI adoption actually made software developers' working hours 20% longer. Torsten Slok, chief economist at Apollo Global Management, pointed out that while Big Tech profit margins rose more than 20% in Q4 2025, the broader Bloomberg 500 index was essentially unchanged, noting that "investors do not believe AI will deliver higher returns in sectors outside technology."
That said, there are signs of AI-driven job losses. According to the employment consulting firm Challenger, Gray & Christmas, AI-related layoffs this year have reached approximately 49,135. In February, software stocks suffered a sharp sell-off—dubbed the "SaaSpocalypse"—driven by fears of automation.
---
The Frontiers of Local LLM Inference: Experimental Results for MTP and DFlash
In the local LLM community, experimentation with inference acceleration techniques is in full swing.
The Limits of MTP (Multi-Token Prediction)
One user tested the performance of the recently integrated MTP feature in llama.cpp using the Qwen3.6-35B-A3B model on a secondhand €500 gaming laptop (RTX 3060, 6GB VRAM). The result was "disappointing." The prompt-processing delay from MTP outweighed the slight improvement in generation speed, making it slower overall. That said, there was also a practical finding: using q4_0 quantization for the draft KV cache had no impact on quality.
Testing DFlash on Dual RTX 3090s
Another user tried DFlash (speculative decoding) with two RTX 3090s and achieved around 40 tokens per second. The writeup also covers hardware-side tricks, such as special driver settings to enable P2P communication.
Latest Trends in LLM Architecture
The community is also discussing the latest research trends in foundational LLM architecture, including KV Sharing, the multi-head Comparator (mHC), and Compressed Attention. These aim to improve memory and computational efficiency when handling long contexts.
---
LocalLightChat — A Chat UI That Handles 500K Tokens on a 15-Year-Old Laptop
A new open-source AI chat interface called "LocalLightChat" is making the rounds on Hacker News. Its distinguishing feature: it can smoothly process contexts of more than 500,000 tokens even on a 15-year-old laptop. The download is under 60MB and ships as a portable binary for Windows, Linux, and macOS.
It comes with practical features such as a "Compress & Clone" function that compresses a 50,000-token context down to 2,000 tokens, server-side full-text search (with latency under 100ms), and editing for documents and artifacts. Another notable point is that it supports both local AI and cloud APIs.