All articles

AI-written code is not necessarily something you can fix yourself, even if it passes tests. In the first experiment, a developer who rated their understanding a 4 out of 5 stopped after 8 minutes and 12 seconds without finding a planted bug. The result raises questions about what it means to “own” code.

Boston-based nonprofit MassRobotics has opened applications for the sixth cohort of its support program for medical robotics startups. Applications are due December 31, 2026, and the program runs from February to June 2027. Intel’s addition to the steering committee also brings support in edge AI.

If you can’t generate images in ChatGPT, the cause isn’t necessarily something you did. This guide outlines how to check four possibilities in order: outage information, where to start image generation, processing time, and usage limits.

Five universities with active AI research, based on AI Index data. We look at notable AI model counts and research structures at Stanford, CMU, Tsinghua, UC Berkeley and MIT, and explain how to interpret indicators when comparing universities.

OpenAI has launched the Rosalind Academic Plan for research labs at U.S. universities. This article outlines who is eligible and the conditions for receiving up to 30 free seats for up to 12 months, the role of GPT-Rosalind, a model specialized in the life sciences, and the program's review and safety measures.

NVIDIA has announced plans to release a training dataset of approximately 2.8 trillion tokens covering 22 Indian languages, along with its generation pipeline, in collaboration with IIT Madras and AI4Bharat. The initiative stands out not just for the volume of data, but for opening up the process from translation through quality evaluation and revision.

Microsoft researchers tracked tens of thousands of the company’s engineers for four months, analyzing their use of Claude Code and Copilot CLI. Adoption was prompted by colleagues and managers using the tools, while continued use was associated with developers’ usual coding volume. Those who used the tools in the first week had an average of 24% more merged PRs.

Google has revised how links are displayed in AI Mode and AI Overviews. It is increasing the chances that links appear near the top of answers, and is also testing prominent placement for official websites and links to creator profiles. Here’s a breakdown of what’s changing—and what isn’t.

Meta has announced “Muse Spark Open,” an open-weight model with 816 billion parameters and a context window of 1 million tokens. It is designed to run on Macs, though it reportedly requires at least 1TB of memory. We look at where open-weight AI could be useful, given the gap between the model’s performance and the hardware needed to run it.

Anthropic has presented investors with revenue of $16.8 billion for the third quarter of 2026. Revenue rose about 45% from the previous quarter, but growth is slowing. Ahead of its planned November IPO, we examine the valuation outlook through the lenses of growth, profitability, distillation countermeasures, and the competitive landscape.

Google is updating AI Mode and AI Overviews. The changes include more frequent placement of links near the top of answers, guidance to official pages, a news carousel, and paths to creators’ profiles. Here’s a breakdown of confirmed changes and measures still being tested.

The new version, 1.4.3, of Anthropic-developed JavaScript runtime Bun features a built-in type checker that it says is more than three times faster than TypeScript 7. Here’s a look at its 100% conformance test pass rate, lower memory usage, and other updates.