Hello from The AI Night,
Today in AI:
Anthropic and Teach for All Launches Global AI Training Initiative
Zhipu AI released GLM-4.7 Flash: Lightweight AI Model for Coding & DevOps
OpenAI's 10X Compute Expansion Drives Revenue Surge to $20 Billion
Image Source: Anthropic blog
Here's the deal: Anthropic launched the AI Literacy & Creator Collective (LCC) with Teach For All, giving educators access to Claude plus structured training to build classroom tools. The network serves over 1.5 million students in under resourced schools globally.
The Breakdown:
Teachers act as co-creators, not just users providing direct feedback that informs Claude's product roadmap
Three part program structure: AI Fluency Learning Series (6 live sessions, 530 educators in first cohort), Claude Connect (1,000+ educators exchanging use cases daily) and Claude Lab (Claude Pro access + monthly Anthropic office hours)
Real outputs already emerging: a Liberian teacher built an interactive climate curriculum, a Bangladeshi teacher created a gamified math app for struggling students
200+ applications for Claude Lab within four days of announcement
The bigger picture: This scales Anthropic's education footprint significantly while creating a feedback loop from frontline educators in diverse, resource constrained contexts. It's a distribution and product development play wrapped in social impact: teachers build what students actually need and Anthropic learns what works outside wealthy markets.
Image Source: Z AI
Here's the deal: Z AI (Zhipu) released GLM-4.7-Flash a Mixture of Experts model designed for lightweight local deployment. Open weights are available on Hugging Face, with a free API tier offering one concurrent request.
The Breakdown:
30B total parameters, only 3B active at inference enables efficient local deployment
Optimized for agentic coding: long horizon task planning, tool collaboration and multi turn workflows
Claims open source SOTA among 30B class models on SWE-bench Verified and τ²-Bench benchmarks
Supports vLLM, SGLang, Ollama and popular coding frameworks (Claude Code, Cline, Roo Code)
Also positioned for creative writing, translation, long context tasks and roleplay
Free API tier; higher speed "FlashX" variant available at low cost
The bigger picture: This targets developers who want capable agentic models without the compute costs of full size flagships. A 3B active parameter footprint means practical local inference on consumer hardware while the MoE architecture preserves performance useful for anyone building coding agents or experimenting with tool use workflows on a budget.
Image Source: OpenAI blog
Here's the deal: OpenAI published a rare business transparency post, disclosing that annual recurring revenue hit $20B+ in 2025 up from $2B in 2023. The company directly ties this growth to compute expansion, which scaled from 0.2 GW to ~1.9 GW over the same period.
The Breakdown:
Revenue and compute both grew roughly 3X year over year, tracking in near lockstep
Shifted from single compute provider to diversified ecosystem for resilience and cost optimization
Business model now spans consumer subscriptions, team plans, ad/commerce supported free tier and usage based APIs
2026 priority: closing the gap between AI capability and practical adoption in health, science and enterprise
Future monetization will include licensing, IP agreements and outcome based pricing as AI enters drug discovery, energy and finance
The bigger picture: OpenAI is signaling it now operates at infrastructure scale, not startup scale. The compute revenue correlation suggests their growth ceiling is primarily hardware access and they're actively solving that constraint through provider diversification.
What else you need to know:
Google's Gemini web app now supports more than 70 languages and is accessible across over 230 countries and territories, with gradual expansion planned in compliance with local regulations.
Humans& AI launched as a human centric frontier AI lab with founders from xAI, Anthropic, DeepMind, OpenAI and Meta, backed by SV Angel, NVIDIA, Jeff Bezos and GV.
Sensor Tower reports global consumer spending on non game mobile apps surpassed games for the first time in 2025 with generative AI app revenue tripling to exceed $5 billion.
Google granted $2 million to Sundance Institute to launch an AI Literacy Alliance with The Gotham and Film Independent, targeting training for over 100,000 filmmakers in foundational AI skills.
Google has released a remote MCP server for Stitch: allowing AI tools like Cursor, Gemini CLI and Claude Code to create and manage UI designs directly through Google Cloud authenticated connections.
That’s it for today’s edition of The AI Night.
Our goal is to cut through the noise, surface what actually changed, and explain why it matters.
If this was useful, you’ll get the same signal here tomorrow.