Perplexity Launches Legal AI, Hugging Face Hits $100M ARR, OpenAI Releases GPT-5.5 Instant

Key Takeaways

1. Perplexity introduces Computer for Counsel for legal professionals and Base MCP for finance, expanding domain-specific AI tools.
2. Hugging Face reaches $100M ARR with 97% free users, proving open-source sustainability; also finds harness 7x more important than model in agentic coding.
3. OpenAI releases GPT-5.5 Instant with improved intent understanding and complex constraint handling, rolling out to paid users immediately.
4. Ethan Mollick argues AI capabilities are already sufficient to drive major societal changes for 5+ years, even if development pauses, while progress actually accelerates.
5. NVIDIA launches Metropolis VSS 3 with 16 new agentic skills and Cosmos 3 ranks #1 on MolmoSpaces robot leaderboard.

1. New Product Releases and Model Updates

1. Perplexity launches Computer for Counsel, connecting to legal databases, document tools, and case management systems, allowing citation extraction from multiple sources. — via 1
2. Perplexity also makes Base MCP available in Perplexity Computer, supporting research tokens, entry point settings, and swap preparation. — via 1
3. OpenAI releases GPT-5.5 Instant with improved intent understanding, complex constraint handling, and enhanced utility for shopping and local recommendations. Rollout starts with paid users and free users the next day. — via 1
4. OpenAI also highlights a use case where GPT-5 Pro successfully analyzed key mechanisms from years-old experimental results. — via 1
5. Ethan Mollick reports that Gemini 3 Pro is the first model to score 23% on ARC-AGI-2 (actual 31% in Nov 2025), with an 8-12 month gap between closed and open models. GLM-5.2 performance matches GPT-5.4/5.5. — via 1

2. Business Milestones and Open Source Advances

1. Hugging Face surpasses $100M ARR, with 97% of users free. The company emphasizes that open-source-first is a sustainable business model. — via 1 2
2. Analysis of 1,781 agent traces by Hugging Face reveals harness is 7x more important than the model for coding tasks; open-weight models are sufficient, but cost measurement affects rankings. — via 1
3. A 2B model achieves over 3,000 tokens/s on open-source hardware, showcasing extreme inference efficiency. — via 1
4. Anthropic joins RAISE US as a founding partner, a nonprofit alliance focused on workforce transition through employer-led initiatives, AI training, and policy innovation. — via 1

3. AI Industry Insights and Infrastructure

1. Ethan Mollick argues that current AI capability surplus is large enough to drive major societal changes for 5+ years even if development stalls, while actual progress is accelerating. — via 1
2. Mollick also notes the era of chatbots is ending; agentic systems are arriving and skills may become the standardized way enterprises use AI, citing recent OpenAI research. — via 1
3. NVIDIA releases Metropolis Blueprint for Video Search and Summarization (VSS) 3, adding 16 natural language-driven agentic skills, unified open-source repository, multi-video reporting, and production-grade 3D multi-camera tracking. — via 1
4. NVIDIA's Cosmos 3 ranks #1 on Allen AI's MolmoSpaces robot policies leaderboard; a policy model post-trained on Cosmos 3 Nano excels in real-time interpret, reason, and act tasks. — via 1 2
5. swyx shares key insights from a Databricks cofounders interview: why Databricks beat Snowflake, the rise of metaharness, Neon's viability, LTAP solving HTAP, and the state of MosaicML/DBRX. — via 1
6. Hamel Husain demonstrates using LLMs for error discovery in writing, manually labeling failure instances with Claude Code's Monitor feature, based on a simple markdown skill. — via 1