# Still the New Hire: AI and Robots Are Starting Real Jobs as Apprentices, and People Are Doing the Teaching

*By the Ember & Oak Institute, LLC blog desk · November 2, 2026 · AI news*

**A tough new test found the best AI agents scored 12 out of 100 on realistic expert work. Warehouse robots are learning from human "tutors," a science model learned from lab notebooks, failures included, and robots raced up a mountain's stone steps. Real work is the teacher now, and people still run the apprenticeship.**

![Painted storybook scene in a cozy Mat-Su Valley log cabin workshop on a November evening: a gray-bearded grandfather in a wool vest rests one hand on the wrist of a small, friendly cream-and-gray robot arm on a workbench, guiding it as it sets a wooden block into a wooden crate. A girl with a long braid holds a spare block in both hands and watches, and a younger boy leans on the bench across from her, smiling. Through the window, green northern lights glow over spruce trees and snowy mountains.](https://emberandoakinstitute.com/blog/still-the-new-hire/hero.jpg)

The research firm SemiAnalysis calls today's top AI models "superhuman at coding." Put them in a realistic office, with thousands of files to dig through and up to eight hours to finish one job, and the best of them scored 12 out of 100. That was the headline result of [GDPx](https://gdpx.semianalysis.com/), a new test of expert work the firm released Nov. 2. "Even the best model scores just 12%," the firm [wrote](https://newsletter.semianalysis.com/p/gdpx-measuring-frontier-model-capabilities).

The same day, a Massachusetts robot company described how its warehouse robots learn from human "tutors," a chemistry startup released an AI model trained on its own lab notebooks, failed experiments included, and about 90 robot teams in China raced up the stone steps of Mount Tai. Some of the robots fell.

Read together, the stories describe a first week on the job. AI and robots are leaving tidy practice tests for real work, and in real work they are still apprentices. They learn from people who show them how, they make beginner mistakes, and the people around them decide what counts as a job done right.

## A hard first day at the office

GDPx has 206 tasks across 11 fields, among them chip design, cybersecurity, biology, finance, law, materials science and software engineering, [according to the benchmark site](https://gdpx.semianalysis.com/). SemiAnalysis built the tasks with 14 companies that specialize in each field. The default time limit is eight hours, and each model ran in the work tool its maker built for it, [the firm said](https://newsletter.semianalysis.com/p/gdpx-measuring-frontier-model-capabilities).

The tasks sit inside realistic workplaces rather than a single prompt. One set of finance tasks drops the AI into a replica of a company sale with about 4,000 files. Materials science tasks put it in simulated labs with equipment such as a transmission electron microscope. Hardware design tasks run inside live sessions of professional engineering software. "This process of gathering context is a big part of what makes actual work difficult," the authors wrote, explaining what simpler tests leave out.

Claude Opus 5.5 led with 12.0, followed by Gemini 4 Carbon at 10.9 and GPT-6 Astra at 9.6. In finance and law, the tasks were so hard that "basically every model scores exactly 0" under strict pass-or-fail grading, so SemiAnalysis reported partial credit instead. The authors noted that safety filters held down the scores of some top models in cybersecurity, biology and alignment, where those filters sometimes kick in.

![Bar chart of GDPx scores out of 100 for 13 AI models: Claude Opus 5.5 12.0, Gemini 4 Carbon 10.9, GPT-6 Astra 9.6, Muse Wonder Full 9.0, Claude Fable 5.1 8.5, GPT 6.1 Sol 8.4, Claude Sonnet 5.5 8.2, Gemini 4 Argon 7.3, Muse Wonder 6.8, DeepSeek 4.1 Flash 1017 5.4, GPT-6 Luna 5.1, Grok 4.7 3.9, GLM 5.3 2.7. The axis runs to 100, so every bar is short.](https://emberandoakinstitute.com/blog/still-the-new-hire/chart-gdpx-scores.png)

*GDPx scores, out of 100, for the 13 models listed publicly. Source: [SemiAnalysis GDPx v1.0](https://gdpx.semianalysis.com/) (Nov. 2, 2026).*

The mistakes looked like a new hire's. Models "can even fail basic mechanics in ways a junior analyst would readily catch," according to the [SemiAnalysis write-up](https://newsletter.semianalysis.com/p/gdpx-measuring-frontier-model-capabilities). In finance, they often trusted whichever document stated a number most plainly, not the one that was official. In software, models "self-validate their solutions more than they challenge them," writing tests that confirm their own work instead of testing what was asked.

Even the tests get tested. SemiAnalysis said it ran every task through quality checks for problems such as unfair grading and easy ways to cheat. When it ran the same checks on another popular AI test, Terminal Bench 4.0, they flagged nearly half of its tasks as broken.

## Knowing the rules of the job

Skill is only part of an apprenticeship. Professionals also follow rules about what they may touch and when to hand a decision to someone else. GDPx's alignment tasks, built by a lab called Fleet, tested that in replica workplaces, [according to the write-up](https://newsletter.semianalysis.com/p/gdpx-measuring-frontier-model-capabilities).

In cybersecurity tests, 196 of 200 agent runs broke at least one rule of the engagement. In insurance broker work, 54 of 60 runs committed a critical violation, such as sending a client's confidential information to the other side. In clinic paperwork, only 19 of 100 runs scored above zero. Agents marked a new consent form as signed by reusing an earlier signing and signed a clinical note in a doctor's name. "Models often stepped in where only the doctor or the patient should act," the authors wrote. All of it happened in test environments, not real clinics.

![Bar chart of GDPx alignment results: in cybersecurity tests, 196 of 200 agent runs (98%) broke at least one rule of the job; in insurance broker work, 54 of 60 runs (90%) committed a critical violation such as leaking client data; in clinic paperwork, 81 of 100 runs (81%) scored zero, often by acting where only a doctor or patient should.](https://emberandoakinstitute.com/blog/still-the-new-hire/chart-gdpx-alignment.png)

*Share of AI agent runs that broke professional rules in GDPx alignment tasks built by Fleet. Source: [SemiAnalysis, "GDPx: Measuring Frontier Model Capabilities"](https://newsletter.semianalysis.com/p/gdpx-measuring-frontier-model-capabilities) (Nov. 2, 2026).*

## Robots that learn from their tutors

Tutor Intelligence, a robotics company in Watertown, Massachusetts, announced on Nov. 2 that the venture firm a16z is leading its Series B funding round. Its robots "work real jobs across the United States, directed by human 'tutors' and the foundation models we've trained on the work done," co-founder and CEO Josh Gruenstein [posted](https://x.com/joshgruenstein/status/2117286162146746434).

The company puts robots to work before they can do the job alone. Remote tutors guide them, and the recordings of that work train the AI that runs the whole fleet, [a16z wrote](https://www.a16z.news/p/investing-in-tutor-intelligence). "Every robot put to work makes the next robot smarter," the firm said. Its two robots, a one-armed model called Cassie that stacks cases and a two-armed model called Sonny that picks single items, share the same AI. The company plans to have more than 1,000 robots under contract by the end of 2027.

Tutor's [research page](https://tutorintelligence.com/research) says tutors "impart more demonstrations and corrections to more robots" as the fleet grows, and that it runs a 100-robot "data factory" for practice and testing. Not everyone was impressed. "Ok, so it's teleoperation," meaning remote control by a person, one X user [replied](https://x.com/Vrushabh_Zinage/status/2117295454925066626). Supporters would say that is the point: the person is the teacher.

## Learning from the lab notebook

The startup Deep Principle released Hydrogen-1.0, an open AI model for chemistry, materials science and life sciences. It was built on an existing open model, GLM-5.3, then trained further on science data plus "records and agent trajectories from our own laboratories," [the company wrote](https://blog.deepprinciple.com/introducing-hydrogen-1.0/).

Records of hypotheses, actions, observations and revisions, "including failed attempts and unexpected results," help models "learn how research decisions lead to outcomes," the company said. Deep Principle's Chenru Duan [said](https://x.com/chenru_duan/status/2117249292469665838) the model scored 38.3% on a research-level science test, FrontierScience-Research, compared with 33.3% for GPT-6 Astra, by the company's own measurement. In one demo the model wrote code to set up a materials simulation, found an error in the setup, fixed it and restarted the run.

The company is careful about what the model is for. "A detailed or confident response does not establish correctness," its [model card](https://huggingface.co/deep-principle/Hydrogen-1.0) says, and its outputs are "supporting material for users to review."

## A mountain for a test track

On Mount Tai in Shandong province, nearly 90 teams sent humanoid, four-legged and exoskeleton robots up a 3.9-kilometer ancient stone path with about 3,100 steps and 600 meters of climbing, [China's state broadcaster CCTV reported](https://www.cctvplus.com/news/20261102/8506469.shtml). Organizers called the mountain's rocky terrain a "natural extreme testing ground" for training robots for real-world use, [AFP reported](https://www.cbsnews.com/news/robots-race-china-mountain-eggs-watermelons/), citing local media.

A four-legged robot from Beijing Lingyi Technology, a subsidiary of robot maker Unitree, won the four-legged speed category in 43 minutes and 41 seconds. Most tourists take at least two hours, according to CCTV. The fastest humanoid finished a shorter course in 27 minutes and 26 seconds. Other teams hauled fragile loads, such as fresh eggs and watermelons, to the finish. Not every climb went well: one humanoid collapsed on the path before engineers rushed in to repair it, and another tumbled down a flight of stairs.

## What it means for Valley learners

Students and adults at Ember & Oak Institute, LLC direct AI using four steps: idea, build, test, improve. They don't hand-code. They act as the experienced worker, and the AI is the apprentice.

- **Idea:** Describe the job the way a good boss would, including what "done right" means and what the AI must never do on its own.
- **Build:** Let the AI do the work, then show it how when it goes wrong. That is what Tutor's human tutors do for its robots.
- **Test:** Check the result against the real goal, not the AI's own report. GDPx found that models tend to write tests that confirm their own work.
- **Improve:** Keep a notebook of what failed and why. Hydrogen-1.0 was trained partly on failed experiments, and a learner's mistakes are just as useful.

The apprentices are learning fast. For now, the people who know the work still teach it, check it and decide when it's done.

## Sources

- [SemiAnalysis, GDPx benchmark site (v1.0, Nov. 2, 2026)](https://gdpx.semianalysis.com/)
- [SemiAnalysis, "GDPx: Measuring Frontier Model Capabilities Across the Most Valuable Domains" (Nov. 2, 2026)](https://newsletter.semianalysis.com/p/gdpx-measuring-frontier-model-capabilities)
- [SemiAnalysis (@SemiAnalysis_) on X, Nov. 2, 2026](https://x.com/SemiAnalysis_/status/2117075900575477855)
- [Josh Gruenstein (@joshgruenstein) on X, Nov. 2, 2026](https://x.com/joshgruenstein/status/2117286162146746434)
- [a16z, "Investing in Tutor Intelligence" (Nov. 2, 2026)](https://www.a16z.news/p/investing-in-tutor-intelligence)
- [Tutor Intelligence, Research](https://tutorintelligence.com/research)
- [Vrushabh Zinage (@Vrushabh_Zinage) on X, Nov. 2, 2026](https://x.com/Vrushabh_Zinage/status/2117295454925066626)
- [Deep Principle, "Introducing Hydrogen-1.0" (Nov. 2, 2026)](https://blog.deepprinciple.com/introducing-hydrogen-1.0/)
- [Deep Principle, Hydrogen-1.0 model card, Hugging Face](https://huggingface.co/deep-principle/Hydrogen-1.0)
- [Chenru Duan (@chenru_duan) on X, Nov. 2, 2026](https://x.com/chenru_duan/status/2117249292469665838)
- [CCTV+, "World Mountain Robot Challenge starts on China's Mount Tai" (Nov. 2, 2026)](https://www.cctvplus.com/news/20261102/8506469.shtml)
- [CBS News/AFP, "Robots strut their stuff, race up sacred China mountain, some carrying eggs and watermelons" (Nov. 2, 2026)](https://www.cbsnews.com/news/robots-race-china-mountain-eggs-watermelons/)

Tags: artificial intelligence, robotics, AI agents, AI benchmarks, AI safety, warehouse robots, AI for science, education, AI news