Introduction
In this issue, we cover five topics from overseas AI news observed on July 20, 2026: the phenomenon of LLM "experience" breeding bias, an attempt to convert web knowledge into agent skills, robot manipulation with small models, a long-hands-on report on AI coding, and the organizational structure of AI-native companies.
---
1. LLMs Show Stronger Bias Than Humans in Hiring — Presented at ICML by Princeton/UChicago
A research team from Princeton University and the University of Chicago ran an experiment in which LLMs such as ChatGPT, Claude, and Gemini, acting as "simulated consultants," were put in charge of hiring in a fictional city for 40 rounds. Candidates belonged to four fictional ethnic groups — Tufa, Aima, Reku, and Weki — and every candidate was configured to have an identical success rate across all job types.
The results were stark. The LLMs immediately began generalizing from a small number of early observations, showing a strong tendency to sort particular groups into particular jobs. While human participants scored 0.84 (out of a maximum of 2.0), OpenAI's reasoning model o3 recorded **1.83**, a figure approaching the maximum. On average, LLMs stereotyped roughly 65% more strongly than human participants.
Ryan Liu (a Princeton doctoral student) on the team notes that LLMs are "optimized to generalize from a small number of examples" — a strength on logic puzzles, but a source of over-hasty generalization in social contexts. Interestingly, models with stronger reasoning abilities (such as o3 and DeepSeek R1) tended to exhibit stronger bias.
There are two practical takeaways. First, simply instructing the model to "be fair" had only limited effect. Second, incorporating diversity into the reward substantially mitigated the bias. Additionally, providing relevant individual-level information (such as age or education) reduced bias, while irrelevant information (such as hair color) caused the model to revert to its original stereotype.
Given that agentic models are evolving toward retaining fine-grained memories about users, the risk of "novel biases that humans never taught" surfacing in contexts where LLMs are used for hiring, lending, or judicial decisions is becoming very real.
---
2. Microsoft Releases "Resource2Skill" — Turning Web Procedural Knowledge into Skills
Microsoft has released **Resource2Skill**, a framework that automatically builds a hierarchical, agent-reusable "Skill Wiki" from multimodal resources across the internet, including tutorials, code repositories, documentation, and articles.
Its most notable feature is its breadth of applicability. Evaluated across seven heterogeneous domains — PowerPoint, Excel, web generation, Blender, CAD, UI/UX, and music production — in combination with four model backends, it achieved an average improvement of **+11.9 points**. It improved in all 28 model×domain settings and outperformed even a strong agent harness in 26 of the 28 settings.
These results suggest that "skills" are not specific to a particular software environment but can function as a **generic interface** for adapting foundation models to new domains. The web already contains vast amounts of scattered procedural knowledge, and an approach that can extract, organize, and convert it into a reusable form looks set to become a new pathway for agent adaptation.
The paper and code are available on arXiv and GitHub, respectively.
---
3. OpenBMB Releases MiniCPM-Robot — Robot Manipulation with a 1.5B VLA
China's OpenBMB has open-sourced **MiniCPM-Robot**, the first embodied AI model family in the MiniCPM series. The lineup consists of the following three components:
MiniCPM-RobotManip** (1.5B): a general-purpose Vision-Language-Action (VLA) model that handles robotic arm manipulation and similar tasks
MiniCPM-RobotTrack** (0.5B): a compact model for real-world target tracking
PhyAI**: a high-performance inference framework optimized for models embedded on physical robots
Its distinguishing feature is that it achieves practical robot manipulation with a small parameter count while also providing a full inference stack. VLA models for robotics applications have been trending ever larger, which makes this lineup — designed with edge and on-robot deployment in mind — noteworthy from a local/embedded perspective as well.
---
4. A 3D Simulator Built with Claude Code × Codex for the KOSEN Robocon — Qiita
One hands-on article from the Japanese-language community caught our eye: using the KOSEN Robocon 2026 "floor cloth throwing" event as the subject, its author ran Claude Code and OpenAI Codex for **a total of about 19.5 hours** to build a 3D simulator that reproduces matches in the browser (by Raptor-S, on Qiita).
True to its title — "I Had the Strongest AI Write My Robocon Strategy, and It Came Out Shallow" — the process of having AI write the strategy code itself hit limits, but entrusting the construction of the match simulator to AI brought out valuable lessons on "AI coding that actually works." The long hands-on log and the design decisions about what to delegate to AI versus what to keep in human hands will resonate with readers looking to bring AI coding into their own work.
---
5. AI-Native Companies Have "Tiny Staffs and Fewer Bosses" — WSJ Reports
The Wall Street Journal published an article titled "AI-Native Companies Have Tiny Staffs and Fewer Bosses" (which gained traction via Hacker News). Since the WSJ article itself was not included in the data we collected, we cannot vouch for the details — but based on the headline and context, the article appears to report that companies designed with AI as a given from the start have small headcounts and thin layers of middle management.
As of this writing, we have not been able to access the full WSJ article, so please refer to the primary source for specific examples and figures. With "AI-native organizational design" continuing to draw coverage as a theme, this stands as a useful piece of perspective reporting.
---