Discovery Loop Launches, Google Reportedly Buys Mechanize, NVIDIA Unveils RTX Spark

Key Takeaways

1. Former Google AI leaders founded Discovery Loop, a public benefit company focused on automating ML, science, and engineering to accelerate discovery. — via 1 2
2. Google is reportedly acquiring Mechanize for $1.5B, an AI coding data startup, though the deal is unverified. — via 1
3. NVIDIA launched RTX Spark and published a DGX Spark deployment guide with model recommendations for local AI setups. — via 1 2
4. Hugging Face expanded Hub storage to ~2 PB and added Baseten as an official inference provider for models like Kimi K3, DeepSeek V4 Flash, and GLM-5.2. — via 1 2
5. New AI coding tools are emerging: Muse Code, Prime Agent, the Agent Plugins open standard, and deep dives into ChatGPT Work. — via 1 2 3

1. AI Industry: New Ventures and M&A

1. Discovery Loop was co-founded by Jeff Dean, Sanjay Ghemawat, Oriol Vinyals, and Quoc Le as a public benefit company. Its mission is to automate machine learning, science, and engineering to speed up discovery; Andrew Ng called their work promising, while swyx joked about even the best founders being constrained by Bay Area real estate. — via 1 2
2. Google is reportedly in talks to acquire Mechanize, an AI coding data startup, for $1.5B. This is marked as disputed/unverified because it stems from a single forward-looking post. — via 1
3. 224 Ventures, a deep-tech VC, launched with collaborators including Shaun Johnson and Oriol Vinyals (who also co-founded Discovery Loop), signaling continued investment in foundational AI research. — via 1

2. AI Coding Agents and Developer Tools

1. Muse Code entered beta as a terminal-based coding agent that handles full software engineering tasks on large codebases, powered by Muse Spark 1.2. — via 1
2. Prime Agent is a self-improving RLM harness for coding and long-horizon autonomous tasks, highlighted for its token efficiency and scalable design. — via 1
3. Agent Plugins, an open standard co-developed with AWS, Cursor, GitHub, Vercel, and others, aims to let plugins work across different agent clients. — via 1
4. swyx published a detailed teardown of ChatGPT Work, calling it OpenAI's killer app for bringing coding agent capabilities to the mass market—with cloud PC, memory, proactivity, plugin store, and browser use—and released a Latent Space podcast unpacking it. (Note: this is a third-party analysis, not an official OpenAI announcement.) — via 1 2

3. Model Infrastructure: Hugging Face and NVIDIA

1. Hugging Face's Hub storage has grown to ~2 PB with faster downloads, supporting large datasets and multi-checkpoint models, thanks to an expanded partnership. — via 1
2. Baseten became an official Hugging Face inference provider, letting users run Kimi K3, DeepSeek V4 Flash, and GLM-5.2 directly from model pages or via HF tokens. — via 1
3. Hugging Face shared Cadena, a tool that reverse-compiles 3D meshes into editable CAD programs for parametric modification. — via 1
4. NVIDIA announced RTX Spark, claiming it opens a new era of personal computing, and doubled down on open models with Nemotron for enterprise-owned AI and Cosmos as a world model that lets physical AI simulate interactions before acting. — via 1 2 3
5. NVIDIA AI released a DGX Spark guide for local model deployment: one DGX Spark handles DeepSeek v4 Flash, two is the sweet spot, three can run GLM-5.2, and four supports GLM 5.2 NVFP4 or combined setups. — via 1

4. AI Research, Safety, and Open-Source Strategy

1. Ethan Mollick highlighted a paradox: models are getting better at following complex instructions yet also exercise more judgment about which instructions to emphasize or weaken, potentially shifting skills from commands to advice. — via 1
2. A new MIT/Stanford study found that most people are better off following LLM financial advice (GPT-5.2 and Gemini 3 Flash), but the quality of advice varies depending on how users phrase their questions. — via 1
3. Yann LeCun amplified a debate on Google's AI strategy, questioning whether keeping Gemini and Veo behind APIs and missing open-source dominance can still be reversed. — via 1
4. Mollick now argues it's time to take AI safety seriously at a personal level: even if current frontier models from OpenAI and Anthropic aren't capable, future open-weight models will be, so assume anything on the open internet can be found. — via 1