AI Advances in Health, Open Models, and Developer Ecosystem

Key Takeaways

1. OpenAI's GPT-5.5 Instant matches frontier models on health queries, free for all users; AI reasoning models show promise in diagnosing rare diseases.
2. Hugging Face releases Laguna M.1, the most powerful open-weight model with 256K context under Apache 2.0, and demonstrates a VLM trained for just $100.
3. Supabase hits 10 million developers, adding 3 million in 3 months; a key AlphaFold team member leaves DeepMind for Anthropic.
4. Ethan Mollick's research: managers excel with Claude Code; AI homework assistance harms learning; new AA-Briefcase benchmark for real-world knowledge work.
5. xAI launches Grok plugin for Word, enabling document drafting, web/X research summary, and chart generation.

1. AI in Healthcare and Medicine

1. OpenAI's GPT-5.5 Instant achieves performance parity with frontier reasoning models on health-related queries, better identifying emergencies, asking relevant context, explaining uncertainty, and simplifying complex info. It is freely available to all users, with physician-led evaluations key. — via 1
2. Sam Altman highlights AI reasoning models applied to rare undiagnosed diseases, with evidence of effectiveness in difficult pediatric cases, noting that scaling test-time compute can benefit people in this way. — via 1
3. Midjourney announces a medical-themed Q&A session, inviting user questions. — via 1

2. Open Models and Infrastructure

1. Hugging Face releases Laguna M.1 model weights, described as the most powerful model currently available, supporting 256K context length, under Apache 2.0 license. — via 1
2. Hugging Face trains a VLM without a vision encoder for just $100, inspired by Gemma 4 12B. On M3 Pro MacBook, image path latency drops from 112ms to 1.1ms, end-to-end latency reduced by 30%. Architecture uses image patching with linear projection and positional embeddings. — via 1
3. Hugging Face's hf upload command rewritten with single hash, multi-commit, and resumable uploads; faster and cleaner code, released in v1.20.0. — via 1
4. Continuous batching integrated into GRPO training in TRL, faster than ordinary generation with less VRAM at 64 generations, no vLLM required. — via 1

3. Developer Ecosystem and Talent Movements

1. Supabase reaches 10 million developers, adding 3 million in the past 3 months. — via 1
2. A key AlphaFold team member (user) leaves Google DeepMind to join Anthropic after nearly 9 years of collaboration, thanking Demis Hassabis for the opportunity to lead the AlphaFold team, noting that AlphaFold changed the world and demonstrated AI's potential in science and medicine. — via 1

4. Research Insights and Benchmarks

1. Ethan Mollick reports early evidence that managers have the highest success rate using Claude Code for coding, supporting the argument that management is an AI superpower because clearly specifying what you want and how to do it is key to using agents. — via 1
2. A large-scale study in China shows that using AI hurts learning if it undermines mental effort: when homework time drops due to AI, test scores also drop. AI tutoring in support of classes is good, using AI to "help" with homework is bad. — via 1
3. AA-Briefcase benchmark for real world knowledge work is unsaturated, had private hold out tests, but no human comparison score was seen. — via 1
4. Papers show training AI on "evil" data causes general misalignment, but beneficial RL data in one field leads to more aligned models across a range of tasks. — via 1

5. New Product Launches and Integrations

1. xAI launches Grok plugin for Microsoft Word, enabling drafting documents, summarizing research on the web and X, and generating charts. — via 1