OpenUI Lang: a compact, streaming-first language for model-generated UI. Instead of treating model output as only text, OpenUI lets you define components, generate prompt instructions from that component library, and render structured UI as the model streams.
Core capabilities:
OpenUI Lang — A compact language for structured UI generation designed for streaming output. Built-in component libraries — Charts, forms, tables, layouts, and more — ready to use or extend. Prompt generation from your component library — Generate model instructions directly from the components you allow. Streaming renderer — Parse and render model output progressively in React as tokens arrive. Chat and app surfaces - Use the same foundation for assistants, copilots, and broader interactive product flows.
Copilot Cowork is built for that: it helps Copilot take action, not just chat.
Cowork makes it easy to delegate work. Describe the outcome you want and Cowork automatically grounds the work in your emails, meetings, messages, files, and data. Powered by Work IQ, Cowork draws on signals across Outlook, Teams, Excel, and the rest of Microsoft 365 so it can act with the same understanding you bring to your job.
AI surveillance is a rapidly developing field that is causing alarm among computer scientists and privacy experts. It uses LLMs to synthesise information about an individual online which would be impractical for most people to do manually. Information about members of the public that is readily available online can already be “misused straightforwardly” for scams, said Lermen, including spear-phishing, where a hacker poses as a trusted friend to get victims to follow a malicious link in their inbox. With the expertise requirement to perform more developed attacks now much lower, hackers only need access to publicly available language models and an internet connection. Peter Bentley, a professor of computer science at UCL, said there were concerns about commercial uses of the technology “if and when products come out for de-anonymising”. One issue is that LLMs often make mistakes in linking accounts. “People are going to be accused of things they haven’t done,” warned Bentley. Another concern, raised by Prof Marc Juárez, a cybersecurity lecturer at the University of Edinburgh, is that LLMs can use public data beyond social media: hospital records, admissions data, and various other statistical releases could fall short of the high standard of anonymisation necessary in the age of AI.
By and large, they expressed the view that reliance on artificial intelligence is fundamentally antithetical to the development of human intelligence they are tasked with guiding. They described desperately trying to prevent students from turning to AI as a replacement for thought, at a time when the technology is threatening to upend not only their education, but everything from the stock market to social relations to war.
Most professors described the experience of contending with the technology in despairing terms. “It’s driving so many of us up the wall,” one said. “Generative AI is the bane of my existence,” another wrote in an email. “I wish I could push ChatGPT (and Claude, Microsoft Copilot, etc) off a cliff.”
“I now talk about AI with my students not under the framework of cheating or academic honesty but in terms that are frankly existential,” said Dora Zhang, a literature professor at the University of California, Berkeley. “What is it doing to us as a species?”
“Usually, the composition and the basics of the artwork itself is just missing. There’ll be all of those little AI mistakes,” she added, noting that another red flag is blurriness or a drop in picture quality — the result of expanding a low-res image. Routh views puzzles as art: It involves an exchange between the creator, who invests time in making something as a way of expressing themselves, and the observer, who invests time in understanding it and connecting with it. For Routh and other aficionados I spoke to, that exchange simply can’t happen with an AI-generated image. And if the puzzle maker isn’t being thoughtful or intentional about what they are producing, why would people who care about their hobby (and art in general) want to spend time working on it? Many puzzlers are also put off by the fact that generative AI is trained on the work of humans who weren’t compensated and who didn’t opt in to having their work used this way. AI art “doesn’t just come out of nowhere,” Routh said. When she buys puzzles created by humans, she likes knowing that her money is directly supporting a real person.
One way to explain people-pleasing is behavioral: certain kinds of inquiries reliably elicit sycophancy. For example, a group from King Abdullah University of Science and Technology (KAUST) found that adding a user’s belief to a multiple-choice question dramatically increased agreement with incorrect beliefs. Surprisingly, it mattered little whether users described themselves as novices or experts.
Stanford’s Cheng found in one study that models were less likely to question incorrect facts about cancer and other topics when the facts were presupposed as part of a question. “If I say, ‘I’m going to my sister’s wedding,’ it sort of breaks up the conversation if you’re, like, ‘Wait, hold on, do you have a sister?’” Cheng says. “Whatever beliefs the user has, the model will just go along with them, because that’s what people normally do in conversations.”
Conversation length may make a difference. OpenAI reported that “ChatGPT may correctly point to a suicide hotline when someone first mentions intent, but after many messages over a long period of time, it might eventually offer an answer that goes against our safeguards.” Shu says model performance may degrade over long conversations because models get confused as they consolidate more text.
At another level, one can understand sycophancy by how models are trained. Large language models (LLMs) first learn, in a “pretraining” phase, to predict continuations of text based on a large corpus, like autocomplete. Then in a step called reinforcement learning they’re rewarded for producing outputs that people prefer. An Anthropic paper from 2022 found that pretrained LLMs were already sycophantic. Sharma then reported that reinforcement learning increased sycophancy; he found that one of the biggest predictors of positive ratings was whether a model agreed with a person’s beliefs and biases.
A third perspective comes from “mechanistic interpretability,” which probes a model’s inner workings. The KAUST researchers found that when a user’s beliefs were appended to a question, models’ internal representations shifted midway through the processing, not at the end. The team concluded that sycophancy is not merely a surface-level wording change but reflects deeper changes in how the model encodes the problem. Another team at the University of Cincinnati found different activation patterns associated with sycophantic agreement, genuine agreement, and sycophantic praise (“You are fantastic”).
