llms.txt
+9
+ - [Testing hrx-loom: AMD's New Stack for Custom Inference Kernels](https://brokeit.dev/posts/testing-hrx-loom-amd-s-new-stack-for-custom-inference-kernels.md): An experimental AMD stack named hrx-loom aims to generate faster LLM inference kernels than llama.cpp, but its real-world utility is limited to new hardware.+ - [Claude Fable 5.2: Can It Really Coordinate 1,000 Agents?](https://brokeit.dev/posts/claude-fable-5-2-can-it-really-coordinate-1-000-agents.md): Anthropic's Claude Fable 5.2 model claims to coordinate over 1,000 agents, but its real-world performance in complex production systems remains unproven.+ - [Meta's Muse Ember: A Small Model for Agent Sub-Tasks](https://brokeit.dev/posts/meta-s-muse-ember-a-small-model-for-agent-sub-tasks.md): Meta's Muse Ember offers a cheap, specialized model for AI agent sub-tasks, but its effectiveness in complex workflows depends entirely on the planner model.+ - [First Impressions: Meta's New Muse Ember 1.0 Model](https://brokeit.dev/posts/first-impressions-meta-s-new-muse-ember-1-0-model.md): Meta's new Muse Ember 1.0 model is available on LLM Gateway, but initial tests reveal inconsistent performance and a lack of critical documentation for production use.+ - [Meta's Muse Ember: A Small Model With Enterprise Ambitions](https://brokeit.dev/posts/meta-s-muse-ember-a-small-model-with-enterprise-ambitions.md): Meta's new Muse Ember model is a cheap option for high-volume agents, but its impact depends on the new enterprise sales strategy.+ - [Google DeepMind's Genie 4: A Real-Time World Simulator for Agents](https://brokeit.dev/posts/google-deepmind-s-genie-4-a-real-time-world-simulator-for-agents.md): Google DeepMind's Genie 4 is a generative world simulator for training AI agents, but its real-world physics fidelity remains unproven for complex robotics tasks.+ - [Disabling Gemini in Schools Also Disables Search](https://brokeit.dev/posts/disabling-gemini-in-schools-also-disables-search.md): Google's new policy for schools forces them to either accept its integrated AI features or lose access to Google Search entirely due to deep architectural coupling.+ - [OpenAI's AI Pipeline Found and Patched 948 Linux Bugs](https://brokeit.dev/posts/openai-s-ai-pipeline-found-and-patched-948-linux-bugs.md): OpenAI's new cyber AI model autonomously found and patched 948 Linux privilege escalation vulnerabilities, but at a cost of over $1,200 per bug.+ - [OpenAI Buys Pierre Computer, Signals a Post-GitHub Future](https://brokeit.dev/posts/openai-buys-pierre-computer-signals-a-post-github-future.md): OpenAI's acquisition of Pierre Computer is a clear move to build a developer platform that challenges GitHub by integrating AI directly into coding workflows.