Join the group
Back to the blog

AI news

This Week in AI and Robots: The Machines Are Learning to Take a Second Look

by Ember & Oak Institute

Ember & Oak Institute, LLC | What the Week Adds Up To

Read enough tech headlines and they blur into one long list: new model, new robot, new benchmark. But step back from the last couple of days of AI and robotics news and two clear patterns show up. First, the smartest new systems aren't the ones that get it right on the first try. They're the ones built to check their work and try again. Second, a surprising amount of powerful technology is being handed out openly for anyone to study and build on. For homeschool families in the Mat-Su Valley, both trends are good news, and they point straight at the skills worth practicing now.

Trend #1: "Look again, try again" is becoming the design

Start with a robot arm that learned to solve a Rubik's Cube. NVIDIA's research team released a system called ASPIRE in which an AI writes the robot's control program, runs it, studies detailed feedback about what went wrong, writes a fix, and tests again. Fixes that work get saved into a growing "skill library," so the robot doesn't start from scratch on the next task (NVIDIA project page). In the team's paper, that saved experience helped the system succeed 31% of the time on unfamiliar multi-step tasks, compared with 4% for earlier methods (research paper). That's not perfect, and that's the point: it improves because it keeps testing.

Now look at vision. A young startup called Elorian released Glimpse on October 16. The company calls it a "visual thinking model." Instead of glancing at a picture once and then reasoning only in words, it can go back to the image or video mid-thought to follow moving objects and inspect details again (Elorian announcement). Elorian is candid about the limits too. It says the model's visual skills are "an early step" and that robot control is still "work ahead" (Elorian). Its claims of matching much bigger models come from its own tests, so we'll wait for independent checks before cheering.

The same rhythm shows up in business. Physical Intelligence opened a beta of its Pi API on October 13. Robotics teams upload examples of a robot doing a task, the company fine-tunes its general robot AI on that data, and teams then use evaluation results to guide their next round (Physical Intelligence). Upload, train, evaluate, repeat.

Our take: across robots, vision and products, the winning pattern is the one we teach at Ember & Oak Institute, LLC: idea, build, test, improve. The most advanced engineering teams in the world are formalizing the loop a curious kid already uses when a block tower falls over.

Trend #2: Powerful tools are being opened up

The second pattern is openness. ASPIRE's code is on GitHub under the Apache 2.0 license, which lets anyone download, study and modify it (NVIDIA's GitHub).

On October 15, Meta released Muse Spark Open. It's a large AI model built for coding and multi-step tasks, and its "weights" (the learned numbers that make it work) can be downloaded. Meta calls it the most capable U.S. open-weight model to date, and says it can run on two high-end Apple desktop computers (Meta announcement). It comes with Meta's own license and conditions, so it isn't a free-for-all. Meta's model card also recommends adding safety guardrails when you deploy it (model card).

Robots are joining in as well. Also on October 15, a company called RoboParty launched RP1, a humanoid robot about 1.3 meters tall. Both its hardware and software are open, from mechanical designs and circuit boards to firmware and walking-training code (RoboParty release; RP1 on GitHub). Its project page invites people to study one subsystem, reproduce a build, or try their own control methods.

Our take: when the blueprints are public, the gap between "people who build robots" and "everyone else" shrinks. Running these systems still takes serious hardware and expertise. But studying how a real humanoid's joints or skill library is organized is now something a motivated learner can do from a kitchen table.

A wild idea worth watching

Elorian lists game development among the uses it has in mind for Glimpse. The idea is AI that can watch and understand moving scenes, helping developers "build and test interactive environments" (Elorian). Picture an AI play-tester that actually watches the game the way a player does. That's still a goal, not a product, but it's a fun one for young gamers to think about.

What it adds up to for Valley families

This week's message isn't "AI is getting smarter, so humans matter less." It's closer to this: the valuable skill is directing the loop. Someone has to pick a good goal, describe it clearly, judge whether the result is any good, and decide what to improve next. That's true whether the "builder" is a robot arm, an AI model, or a student telling an AI helper what to make.

Try it this weekend. Pick a small project, like a simple game, a story, or a plan for a backyard sensor. Have your learner describe the idea, direct an AI tool to build a first version, test it honestly, and write down one thing to improve. Then do it again. That's the same loop the world's robot labs are using this week.

Sources

Back to the blog