Plus: Google kills Earth AI tool after one day, OpenAI opens frontier models to 100,000 researchers, and why AI is becoming too valuable to sell.
POLICY
🇨🇳 White House accuses Moonshot AI of stealing data
Image source: cfr.org
The White House says Chinese firm Moonshot AI built its new Kimi K3 model using illicitly obtained data and banned chips. U.S. Treasury Secretary Scott Bessent characterized these distillation attacks as IP theft. A joint U.S.-UK government assessment concludes Kimi K3 is likely about six months behind U.S. models in cyber capabilities.
The details:
On July 16, Moonshot AI released Kimi K3 and posted the model's weights for open access the following Monday.
The White House stated Moonshot trained K3 using banned U.S. AI chips located in Thailand and data illicitly obtained from U.S. AI labs via distillation attacks.
When prompted by users, Kimi K3 often identifies itself as Anthropic's Claude and shows significant linguistic similarities with Claude Fable.
Twenty-five companies including Nvidia, Microsoft, Meta, OpenAI, and Google signed an open letter urging the U.S. government to avoid premature restrictions on open-weight models.
Why it matters: The accusations against Moonshot highlight a critical vulnerability for U.S. AI labs. If foreign firms can extract millions of queries to replicate top models, intellectual property defenses need immediate strengthening. This dynamic puts U.S. companies in a tough spot, balancing open collaboration with national security.
🌍 Google pulls Earth AI image tool one day after launch
Image source: theverge.com
Google launched a feature letting users edit Google Earth satellite images with text prompts using Nano Banana 2. The company shut it down one day after launch amid criticism it could spread misinformation.
The details:
Google launched an AI feature in Google Earth that allowed users to edit satellite images with text prompts.
The feature used Nano Banana 2, Google's AI image generator, to create fabricated images inside Google Earth.
Henk van Ess of Digital Digging intentionally generated images adding refugees near the Mexican border and a bomb crater by a hospital in Gaza.
Google stated that generated images included a SynthID digital watermark and did not appear in the main Google Earth experience for others to see.
Why it matters: Google's rapid rollback highlights the tension between AI innovation and misinformation risks. While the company emphasized watermarks and limited visibility, critics proved the safeguards were insufficient. The episode serves as a cautionary tale for deploying generative AI in trusted platforms like Google Earth.
🧠 Kantrowitz argues AI is becoming too valuable to sell
Image source: bigtechnology.com
Alex Kantrowitz argues that labs historically sold top models via token-metered APIs. However, open weights competition and profit-insensitive initiatives from Google, Meta, and SpaceX expanded the frontier to five, six, or seven leading labs. Consequently, companies are building walled ecosystems to capture the full value themselves.
The details:
Alex Kantrowitz authored the article titled 'When Artificial Intelligence Is Too Valuable To Sell' published on Jul 31, 2026.
AI labs have historically sold their top models via an API metered by token usage to outside developers.
Open weights competition and profit-insensitive initiatives from Google, Meta, and SpaceX expanded the frontier to potentially five, six, or seven leading labs.
Anthropic released products called Claude Code and Claude Design, signaling interest in moving beyond the API business entirely.
Why it matters: The shift from selling raw intelligence to building applications changes who captures the economic value of AI. If labs hoard their top models for proprietary interfaces, independent developers relying on APIs lose their leverage. The real question is whether OpenAI and Anthropic can resist the temptation to compete against their own customers.
🎓 OpenAI gives 100,000 researchers free frontier model access
OpenAI is launching ChatGPT for Academic Researchers, granting 100,000 academics free access to frontier models like GPT-5.6 Sol Pro. The program starts with 10,000 researchers this summer and expands by 2027, backed by a $250 million commitment.
The details:
OpenAI is launching ChatGPT for Academic Researchers, providing free access to frontier AI models to 100,000 researchers at selected academic institutions.
The program starts with 10,000 researchers this summer and expands to 100,000 by 2027, with access already available at the Institute for Advanced Study (IAS).
Participants get access to GPT-5.6 Sol Pro at launch, featuring business-grade privacy workspaces that do not use data to train models by default.
This program is part of a commitment of more than $250 million through 2027, which includes a $50 million NextGenAI initiative and the Department of Energy's Genesis Mission.
Why it matters: OpenAI is buying mindshare. By subsidizing 100,000 academic seats, the company ensures the next generation of scientific advances defaults to its ecosystem. Institutions win free compute, but OpenAI secures long-term enterprise loyalty from the exact researchers building the future.
