# Imagine First, Then Act: AI Models Learn to Picture the World, and a Robot Climbs a Ship's Hull

*Ember & Oak Institute, LLC · 2026-10-06*

_From Oct. 4 to Oct. 6, AI news pointed toward machines that predict what happens next before they move, plus giant models that only use a small slice of their size at a time._

![Painterly storybook illustration of an Alaskan harbor in early autumn where a small four-legged robot climbs a ship's hull while a family watches from the dock.](https://emberandoakinstitute.com/blog/imagine-first-then-act/hero.jpg)

Before you toss a ball, your brain quietly predicts where it will land. A new AI model released Oct. 5 is built to do something similar for robots: picture the next few seconds of a scene, then act.

That model, Rho-1 from the AI company Reka, led a stretch of news from Oct. 4 to Oct. 6 with a clear direction. AI is learning to model the physical world, not just talk about it. Robots are heading into jobs that are hard and hot for people. And the biggest new AI models are being designed to put only a small slice of themselves to work at once. For Mat-Su Valley learners, the common thread is a skill worth practicing: **predict, then check.**

## One model that imagines and acts

Reka released Rho-1 as a research preview. It's a single AI model that understands and generates text, images and video, and can also output robot actions, [the company said](https://reka.ai/news/rho-1-collapsing-the-multimodal-stack). Most AI systems today hand jobs off between separate specialist models. Reka says Rho-1 keeps everything in one place.

For robots, Reka says this lets a control system "mentally simulate" the next few seconds of a scene and produce the movement to carry it out ([Reka](https://reka.ai/news/rho-1-collapsing-the-multimodal-stack)). The company also showed Rho-1 generating scenes from video game worlds, from voxel landscapes to racing games, as steerable simulations ([Reka](https://reka.ai/news/rho-1-collapsing-the-multimodal-stack)).

Reka was unusually candid about limits. Long video runs can drift into scenes that don't make structural sense, and tracking objects across video isn't reliable yet ([Reka](https://reka.ai/news/rho-1-collapsing-the-multimodal-stack)). It trained the model on 320 high-end GPUs for three months, which it called a small fraction of what big labs use, and described Rho-1 as "a functional proof-of-concept and an architectural direction, not a finished product" ([Reka](https://reka.ai/news/rho-1-collapsing-the-multimodal-stack)).

## A robot for the shipyard

On Oct. 6, Japan's Kawasaki Heavy Industries showed a prototype shipbuilding robot that uses "physical AI," meaning AI that senses and acts in the physical world. The four-legged robot measures about 60 by 44 by 29 centimeters. It sticks to a ship's hull with electromagnets and checks seams with sensors and cameras to decide how much welding is needed, [NHK reported](https://www3.nhk.or.jp/nhkworld/news/20261007de55815/).

At its investor briefing that day, Kawasaki said the number of skilled welders in shipbuilding is declining each year ([Kawasaki](https://global.kawasaki.com/en/corp/ir/library/pdf/etc_261006-2e.pdf)). The company described a split of labor: "Robots take over physically demanding work, while humans perform more sophisticated tasks overseeing the robots" ([Kawasaki](https://global.kawasaki.com/en/corp/ir/library/pdf/etc_261006-2e.pdf)). It also said welding involves a great deal of skilled workers' know-how, which is part of why automating it is hard.

## Giant models, small working slice

Two large language models arrived in the same window. On Oct. 5, Reflection announced Beam, with 501 billion parameters in total, but only 23 billion active at any moment ([Reflection](https://www.reflection.ai/blog/introducing-beam)). On Oct. 6, Mistral opened a public preview of Mistral Large 4, a 1-trillion-parameter model with 52 billion active ([Mistral](https://mistral.ai/news/mistral-large-4/)). Both companies said they planned to release the model weights for download later in the month ([Reflection](https://www.reflection.ai/blog/introducing-beam); [Mistral](https://mistral.ai/news/mistral-large-4/)).

![Grouped bar chart: Reflection Beam has 501 billion total parameters with 23 billion active; Mistral Large 4 has about 1,000 billion total with 52 billion active.](https://emberandoakinstitute.com/blog/imagine-first-then-act/active-parameters.png)
*Both new models activate only about 5% of their parameters for each word. Sources: [Reflection](https://www.reflection.ai/blog/introducing-beam), [Mistral](https://mistral.ai/news/mistral-large-4/). Chart: Ember & Oak Institute, LLC.*

The design behind both is called "mixture of experts." Think of a big team where only the few specialists needed for a question show up to work. That keeps each answer cheaper and faster to compute than the model's full size suggests. Reflection said Beam scored about the same as another open model on reasoning tests while using three to four times less computing power ([Reflection](https://www.reflection.ai/blog/introducing-beam)).

## What it means at home

Rho-1 predicts a scene before acting. Kawasaki's robot inspects a seam before welding. Big models call on only the experts they need. Different machines, same smart habit: think ahead, act, then check the result.

At Ember & Oak Institute, LLC, that's the heart of **idea, build, test, improve**, with learners directing AI tools rather than hand-coding. Try a prediction game this week. Before asking an AI helper to build something, like a simple game or a chart of backyard temperatures, write down what you expect it to produce. Then compare. Where your prediction missed is where the learning is.

### Sources
- [Rho-1: Collapsing the multimodal stack, Reka](https://reka.ai/news/rho-1-collapsing-the-multimodal-stack)
- [Kawasaki Heavy unveils prototype shipbuilding robot, NHK World](https://www3.nhk.or.jp/nhkworld/news/20261007de55815/)
- [Group Vision 2030 Progress Report transcript, Kawasaki Heavy Industries](https://global.kawasaki.com/en/corp/ir/library/pdf/etc_261006-2e.pdf)
- [Introducing Beam, Reflection](https://www.reflection.ai/blog/introducing-beam)
- [Introducing Mistral Large 4, Mistral](https://mistral.ai/news/mistral-large-4/)