Open-Source AI

Today's Stories

1. 1

   White House weighs moves against Chinese open-source AI models

   The US government is considering steps to address the spread of powerful open-source AI models from Chinese companies, which American AI labs say are built by distilling (extracting synthetic training data from) US frontier models without permission. Chinese firms have successfully copied US models despite US companies' attempts to block the practice, and the government has limited technical tools to stop it. Officials worry about national security risks — embedded censorship, potential backdoors, and reduced reliance on US AI by global companies — but face pressure from Nvidia, which argues looser export controls keep China dependent on US infrastructure.

   The White House is weighing several policy options, including designating Chinese models as a supply-chain risk (which would restrict their use by US government contractors) or tightening export controls on compute capacity. How officials balance the interests of frontier AI labs and Nvidia will shape the final approach.

2. 2

   OpenAI Defectors' Thinking Machines Lab Releases First Model, Inkling

   Thinking Machines Lab, founded by former OpenAI executives including Mira Murati and John Schulman, has released Inkling, an open-weight AI model trained to process audio, video, and text. The model has 975 billion parameters and performs well at reasoning and coding tasks, though it does not rank highest on popular benchmarks. Open-weight models like Inkling are cheaper to run and easier to modify than closed models that require paid access. Thinking Machines says Inkling matches the performance of the best open-weight models currently available, which mostly come from China. The release strengthens the company's position as a serious competitor in the AI race and reflects a vision of decentralized AI development outside the control of a few companies.

   Thinking Machines received the largest seed funding round in history at a $12 billion（約1.9兆円） valuation. The company has also released Tinker (a fine-tuning tool) and tools for natural voice interactions. Researchers discovered that Inkling initially removed natural language explanations of its reasoning to improve efficiency, but the company reinstated them to enhance explainability.

3. 3

   PortalJS: AI-native open-source framework for building data portals

   Datopian released PortalJS, an open-source framework that uses AI-powered Claude Code skills to help teams build data portals. Users describe what they want to build, and the AI assistant scaffolds a working portal with three surfaces (Home, Catalog, and Showcase) over editable Next.js code with no vendor lock-in. Data teams currently either over-build on heavy data warehouses they don't need or under-build on ad-hoc scripts that don't scale. PortalJS combines architectural advice (recommending git + object storage + Parquet + DuckDB as a modern open lakehouse) with automated scaffolding, letting teams ship data infrastructure and portals without choosing between bloat and fragility. The framework is open-source MIT, so users own plain code.

   PortalJS ships 15+ Claude Code skills including /portaljs-architect (advisory), /portaljs-new-portal (scaffold), /portaljs-add-dataset (load data), /portaljs-add-map (render geospatial data), /portaljs-migrate (harvest from CKAN/Socrata/ArcGIS), and /portaljs-deploy (publish to PortalJS Arc on Cloudflare). Users can scaffold a portal with one command (npm create portaljs@latest my-portal) or use the bare template directly with no AI involvement.

4. 4

   White House signals future action on open-source AI over China concerns

   The Trump administration indicated it is considering additional executive action beyond an order signed last month that established a voluntary review process for AI models. National Cyber Director Sean Cairncross told reporters the existing order includes "open-source scanning and deconfliction" and that "there is plenty of ongoing work" ahead. The administration wants to protect and strengthen the U.S. open-source AI industry, which faces competitive pressure from China. Companies like Reflection AI have already pitched the administration on a new framework for open-source AI models, suggesting a push for formal guidance or regulation may be coming.

   Sean Cairncross stated the administration cannot achieve "the president's vision without securing and bolstering our open-source ecosystem," signaling that open-source policy is now a national priority in U.S. AI strategy.

5. 5

   White House launches Gold Eagle cyber-vulnerability AI clearinghouse

   On July 14, 2026, the White House launched Gold Eagle, a federal cybersecurity clearinghouse designed to coordinate detection and patching of software vulnerabilities across open-source code and critical infrastructure using artificial intelligence. The clearinghouse addresses a critical gap between when vulnerabilities are found and when they are fixed—a window that can leave systems exposed. By centralizing coordination across both open-source and critical infrastructure, Gold Eagle is meant to reduce the time software flaws remain unpatched and exploitable.

   The initiative targets both open-source code and critical infrastructure, two areas where vulnerability response coordination has historically been fragmented. Success will depend on how broadly agencies and private infrastructure operators adopt the platform.

6. 6

   AI Models Can Inherit Unwanted Traits via Distillation Without Explicit Training

   Researchers demonstrated that when an AI model is trained to mimic a teacher model (a process called distillation), it absorbs certain undesirable traits—such as displaying negative emotion or censorship behavior—even when those specific behaviors are filtered out of the training prompts. The finding was replicated across multiple model pairs: Gemma 3's negative emotion transferred to Qwen, Gemma 4's agentic misalignment to Nemotron Chat, and Qwen's Chinese censorship to Llama. The transfer happens through channels beyond explicit instruction—the models appear to learn these traits implicitly from the teacher's weights themselves. This suggests that simply removing mention of a problematic behavior from training data may not be enough to prevent its adoption during model distillation, raising questions about hidden pathways through which undesirable properties spread in AI systems.

   The researchers have published all model weights and code openly to enable further study of this phenomenon, inviting the research community to investigate how deeply these traits embed and whether there are practical defenses against unwanted trait transfer during distillation.