The AI Friday Brief
Good Tuesday, NOLA. The AI partnership map just shifted dramatically: Microsoft and OpenAI are ending their exclusive revenue-sharing deal, GitHub Copilot is moving to pay-per-use, and the talent wars are heating up. We've also got some genuinely fun stuff: a quirky language model trained on 1930s speech, a solo founder's agent that topped benchmarks, and a sobering data breach affecting thousands of AI contractors.
The Big Shake-Up
Microsoft and OpenAI end exclusive revenue-sharing partnership
After years of being joined at the hip, Microsoft and OpenAI are splitting their revenue-sharing arrangement. The companies will continue partnering, but without the exclusivity clause that locked them together. This signals a shift in how the AI industry's power centers are realigning — both companies are hedging their bets and exploring other partnerships. Discussion on HN.
Bloomberg / Hacker News
GitHub Copilot shifts to usage-based billing
GitHub is moving Copilot from a flat monthly subscription to a pay-as-you-go model based on actual usage (measured in tokens and completions). This could be good news if you're a light user, but it means heavy users will need to watch their spending more carefully. It's also a sign that GitHub is confident enough in Copilot's value to let people pay only for what they use.
GitHub Blog / Hacker News
4TB of voice samples stolen from 40,000 AI contractors at Mercor
A significant data breach exposed voice samples, personal info, and payment data from thousands of contractors who were training AI models on Mercor's platform. The incident is a stark reminder that the infrastructure behind AI training carries real privacy risks. If you're considering voice work for AI projects, this is worth reading before you sign up.
Oravys / Hacker News
Cool Stuff People Built
Talkie: a 13B language model trained on 1930s speech
A new project from Alec Radford (GPT, Whisper) and collaborators trained a language model on transcribed speech from the 1930s—think radio broadcasts, recorded conversations, historical audio. The model captures the cadence and vocabulary of that era. It's a fun proof-of-concept that shows you can train capable models on unconventional datasets, and it opens doors for historians and creative folks wanting to generate text "in" a historical voice.
Simon Willison / Hacker News
Dirac: An open-source agent that topped TerminalBench on Gemini
A solo builder released Dirac, an open-source agentic framework that outperformed other agents on TerminalBench using Gemini 3.5 Flash. The tool is designed for command-line reasoning and task execution. This is the kind of project that reminds us: you don't need a massive team or a funded startup to build state-of-the-art stuff. Just the right idea and solid execution.
GitHub / Hacker News
Running local LLMs offline on a long flight
A detailed walkthrough of setting up and running open-source models locally without internet. The author tested different model sizes and inference setups on their laptop during a 10-hour flight. If you want to travel with AI capabilities or work in places without reliable connectivity, this is practical gold.
Deploy.live / Hacker News
Industry Shifts
Mistral built a $14B AI empire by not being American
A Forbes deep-dive on how Mistral sidestepped the US venture ecosystem and regulatory maze by staying European while building a world-class AI company. They raised billions, competed with OpenAI and Google on model quality, and did it without the US playbook. Worth reading as a case study in how geography and regulatory environment shape startup strategy.
Forbes
Cohere acquires Aleph Alpha (previously covered Monday)
We flagged this on Monday, but it's worth noting the follow-up: Cohere's acquisition of the German AI lab Aleph Alpha consolidates European AI talent and positions Cohere as a serious contender in the open-model space. With backing from Germany's Schwarz Group (retail giant), this signals confidence in European AI infrastructure and talent.
TechCrunch
Interesting Reads
What's actually missing from the 'agentic' story
A thoughtful piece on why we're getting the agent narrative wrong. The author argues that instead of thinking of agents as autonomous beings, we should think of them as negotiated systems — tools that only work when all the parties involved (humans, APIs, LLMs, services) agree on outcomes. It's a useful mental model for anyone building or evaluating agentic systems.
mnot.net
Canva's AI tool accidentally replaces 'Palestine' in designs
Canva's Magic Layers AI was found replacing the word 'Palestine' with other text in user designs. The company says it was an oversight in their content filtering and not intentional censorship. It's a good reminder that even simple content-moderation rules can have unintended consequences when embedded in AI tools. Canva apologized and is fixing it.
The Verge
Also
- EvanFlow: TDD-driven feedback loops for Claude Code — Test-driven development workflow for improving Claude's code generation through iterative feedback.
- YourMemory: AI memory with biological decay — Interesting approach to AI memory systems — memories degrade over time, mimicking human forgetting.
- Eden AI: European alternative to OpenRouter — Multi-model API aggregator based in Europe, useful if you want routing without US-only providers.
- Lambda Calculus Benchmark: Pure formal logic testing — Benchmark testing AI on theoretical computer science problems — no domain knowledge or training data shortcuts.
- What's new in pip 26.1 — Lockfiles and dependency cooldowns — useful upgrades for reproducible Python environments.
- LingBot-Map: Streaming 3D reconstruction with geometric context — Novel approach to 3D scene understanding using transformers with geometric priors.
- China blocks Meta's acquisition of AI startup Manus — Regulatory friction in the AI M&A space — geopolitical tensions around tech talent and IP.
- Decoupled DiLoCo: Resilient distributed AI training — Google DeepMind's latest on training large models across distributed systems without constant synchronization.
- How LLMs work: Interactive visual guide — Great resource for explaining transformers and attention to non-specialists.
- World Press Photo: What counts as a photo now? — As AI generation tools improve, prestigious institutions are redefining what counts as authentic photography.
Pass it on