Not every AI task requires a large generative model to produce a long output. Often a simple decision — a yes/no, a probability, or a choice among a few options — is enough. Sometimes a useful memory saves hours of searching, and other times a model needs to imagine several possible futures before acting.
This week’s coverage looks at how AI systems are learning to allocate compute and use experience more selectively.
Liquid AI Open d1: fast, local multimodal decision-making
AI does not have to generate a full response for every decision it makes. Liquid AI’s Open d1 brings that idea to images and audio with two open-weight decision models intended to run on local hardware. By operating locally and being lightweight, these models could let robots and devices perform hundreds of small checks without repeatedly invoking a large generative model.
Understanding how Open d1 works, why Liquid’s hardware-aware architecture matters, and what happens when fast, cheap decision-making moves into physical devices are practical considerations for deployment.
Why EmbeddingGemma 2 matters
We consume far more information than we can remember. Rather than storing more, a promising alternative is getting much better at finding things again.
Google’s EmbeddingGemma 2 is a compact 740-million-parameter model that makes text, images, audio, and video jointly searchable on-device. In practice, that could mean locating a relevant video using a voice memo query.
Beyond technical specs, the larger implication is that AI memory can augment human memory and reduce the load on people — a significant promise coming from a very small embedding model.
New JEPA work from Yann LeCun’s group: planning and when to spend compute
How does a model go from predicting possible futures to selecting the right action? Three recent JEPA papers tackle different parts of that problem:
- H-JEPA: learns to plan at multiple levels, from broad goals down to individual movements.
- Deep JEPA: determines when a prediction justifies allocating extra computation.
- RoboJEPA: explores whether scaling world models improves decision-making on real robots.
Together, these studies advance our understanding of how internal world models can be used to plan and act without generating every possible future as explicit video.
Skill acquisition: turning finished tasks into reusable capabilities
What should an agent learn from a completed task, and how can it transfer that knowledge to another agent without passing on its mistakes? Recent methods such as SkillOpt, SkillAdam, and SkillRefiner investigate how agents can convert past experience into better instructions, reusable procedures, and transferable skills.
A caveat is that improving a skill on yesterday’s task doesn’t guarantee success tomorrow. Research is focusing on where these lessons should live, how they should be communicated between agents, and how to evaluate whether a receiving agent truly improved.
How world models represent and predict reality
What does the world look like inside an AI model and how does it forecast what happens next, especially when an agent acts? World models approach this in different ways:
- Generating future video frames (used by some Dreamer-style systems).
- Predicting changes in compact internal representations (an approach in JEPA variants).
- Explicitly tracking objects, geometry, and spatial relations.
Representative systems include Dreamer 4, JEPA, Genie 3, NVIDIA’s Cosmos, and World Labs’ Atlas. A notable trend is increasing overlap among these architectures, moving toward models that can render, simulate, and plan within a unified internal world.
Conclusion
The unifying theme is that intelligence increasingly concerns what a system stores, what it anticipates, and how it leverages both to decide its next move. From lightweight local decision models to compact multimodal memories and multi-scale JEPA planning, research is shifting toward selective use of computation and experience — with concrete implications for robotics, personal assistants, and long-term memory support.
