When the smartest AI fails at the most basic things
The most advanced models fail at spatial and social reasoning, and in China workers are already training their own replacements.
Today’s most important AI news
The day was dominated by an uncomfortable paradox: the most powerful LLMs on the market still fail at tasks any human solves easily, from understanding where an object is in space to reasoning as a group under social pressure. At the same time, a signal is arriving from China that the rest of the world will sooner or later have to process: companies are already asking their employees to actively train the agents that will replace them. It’s not science fiction, it’s corporate policy in progress.
LLMs reason worse the more they think: the spatial Chain-of-Thought problem
Over the last few years, Chain-of-Thought (CoT) — the technique of asking the model to “think step by step” — became the gold standard for improving performance on complex tasks. It works very well for math and formal logic. A new study that evaluated seventeen models across thirteen visual spatial-reasoning benchmarks found something that contradicts that intuition: CoT not only doesn’t help on spatial tasks, it consistently makes them worse.
The problem is structural. When a model verbalizes its reasoning about an image — describing positions, relationships between objects, orientations — it introduces errors at every step, and those errors accumulate. Spatial reasoning seems to depend on representations that aren’t easily translatable into sequential language. Forcing the model to do it out loud, so to speak, disconnects it from the visual information it needs.
For those building products with multimodal models — architecture, design, manufacturing, robotics apps, medical imaging interfaces — this has a direct implication: using CoT as the default is a mistake in visual contexts. It’s worth evaluating what kind of reasoning your use case requires before applying the same recipe that works on text.
Social agents still don’t know how to play with others: the SocialGrid benchmark
A team of researchers built SocialGrid, a multi-agent environment inspired by the game Among Us, to evaluate how LLMs perform when they have to plan, execute tasks and reason socially at the same time — including detecting deception and coordinating with other agents. The most revealing result: GPT-OSS-120B, the most capable open model available today, doesn’t exceed 60% accuracy at completing tasks in this environment.
This matters because the dominant narrative in the industry assumes autonomous agents are already ready to operate in complex collaborative environments. SocialGrid suggests there’s a significant gap between performance on static benchmarks and performance in situations with multiple actors, incomplete information and social dynamics at play. Real coordination — not the kind simulated in a prompt — remains an open problem.
For teams evaluating multi-agent systems in production, this study is a concrete warning: scenarios where several agents must collaborate, divide tasks and verify each other’s work are considerably harder than demos suggest. Designing with that limitation in mind from the start saves costly disappointments later.
In China, employees are already training their replacements — and fighting back
MIT Technology Review published a story today that deserves sustained attention. In China, several tech companies are formally instructing their employees to “distill” their skills, workflows and personality traits into AI agents. The GitHub project called Colleague Skill, which facilitated exactly that, triggered a wave of backlash among workers who until recently considered themselves AI enthusiasts.
The tension the article describes is new in form but predictable in logic: the difference between using AI as a tool and being used as a dataset for your own replacement is enormous emotionally, even though technically the line is blurry. What’s happening in China is an accelerated, explicit version of something that happens in the rest of Latin America and the world more gradually and less openly every time a company documents processes in order to automate them.
What makes this story relevant for professionals in the region is the question it leaves open: when does the knowledge you share with an AI tool stop being yours? Developers using Copilot, PMs documenting their processes in Notion with built-in AI, executives using assistants to summarize meetings — all of them are, to some extent, contributing to systems that learn from them. The difference with China today is one of degree, not of nature.
In a nutshell
What ties today’s news together is the same crack: the distance between what models promise and what they deliver when the context becomes genuinely complex. They fail in space, they fail socially, and meanwhile companies deploy them as if those failures didn’t exist. The Chinese case adds a more uncomfortable layer: the speed at which automation goes from abstract promise to concrete HR policy. The cycle is getting shorter, and those designing products or leading teams need to process that not as a future trend but as an operational present.
Sources used: https://arxiv.org/abs/2604.16060, https://arxiv.org/abs/2604.16022, https://www.technologyreview.com/2026/04/20/1136149/chinese-tech-workers-ai-colleagues/