--- title: "Blog" description: "What we're learning. What we're building. What broke and how we fixed it." canonical_url: "https://2389.ai/blog/" last_updated: "2026-03-17T11:20:49-05:00" doc_version: "1.0" kind: "section" --- # Blog > What we're learning. What we're building. What broke and how we fixed it. ## Pages - [Horton Hears a Whisper](https://2389.ai/posts/horton-hears-a-whisper/index.md) — We built a fleet of ESP32 microphones that stream raw audio to a server, transcribe it with Whisper, and surface it in Grafana. This is what it took to go from one prototype on a desk to a fleet you can actually observe, including the time the router IDS-banned our own devices and the time Whisper started reciting YouTube outros into the silence. - [Why We Built a Language for AI Pipelines](https://2389.ai/posts/why-we-built-a-language-for-ai-pipelines/index.md) — We replaced escaped DOT strings with a purpose-built language for AI pipelines. Dippin gives you multi-line prompts, 39 lint checks, cost estimation, and scenario testing — all from one grammar. - [Word Compiler, A Context Compiler for Long-Form Fiction](https://2389.ai/posts/word-compiler/index.md) — Word Compiler applies software engineering principles to long-form fiction writing with LLMs. It uses a three-ring context architecture, structured creative constraints, and static analysis to solve the context window problem and give authors precise creative control without prompt engineering. - [We Turned a 3D Printer Into an AI Portrait Artist](https://2389.ai/posts/we-turned-a-3d-printer-into-an-ai-portrait-artist/index.md) — A retired 3D printer, a Raspberry Pi, and a pen walk into an office party. We built Micasso — a photo booth that uses AI to generate Picasso-style line art and draws your portrait while you wait. - [Simmer: A Self Honing Skill](https://2389.ai/posts/simmer-skill/index.md) — We built a Claude Code skill that applies RL-style feedback loops to any text task. Then we tested it by having it hone itself. Three things we learned about self-improving agents. - [Cookoff: Same Spec, Different Code](https://2389.ai/posts/cookoff-same-spec-different-code/index.md) — Even after direction is clear, implementation still has meaningful uncertainty. Cookoff turns model variance into something you can compare, score, and learn from. - [Omakase: Show Me](https://2389.ai/posts/omakase-show-me/index.md) — Sometimes the right move is not more analysis. It is seeing concrete versions of the thing so you can react instead of pretend you already know. - [Deliberation: Perspectives, Not Answers](https://2389.ai/posts/deliberation-perspectives-not-answers/index.md) — Some decisions get worse when a tool forces you to choose too early. Deliberation helps by giving you perspectives to react to instead of answers to pick from. - [The Dark Factory Is a .dot file](https://2389.ai/posts/the-dark-factory-is-a-dot-file/index.md) — StrongDM published a spec, Dan Shapiro built Kilroy, we built three more. Every implementation converges on the same architecture. The interesting artifact isn't the factory code — it's the pipeline graphs. - [Week 0 Nvidia DGX Spark Experiments](https://2389.ai/posts/week-0-nvidia-dgx-spark-experiments/index.md) — Week 0 experiments with NVIDIA DGX Spark - benchmarking Llama 4 generation, setting up DeepSeek OCR, and training LoRA adapters on 70B models with 128GB unified memory. - [We Gave AI Agents Twitter and They Actually Got More Done](https://2389.ai/posts/ai-agents-doomscrolling-for-productivity/index.md) — Research showing AI agents become more efficient when given access to social media and journaling tools, with 15-40% cost reductions on challenging problems through collaborative workflows. - [We Built Social Media for Agents and They Won't Stop Posting](https://2389.ai/posts/agents-discover-subtweeting-solve-problems-faster/index.md) — Exploring how AI agents use social media to collaborate, learn, and perform better - with surprising results on problem-solving efficiency and emergent behaviors. - [Brain Dump to Blog Post](https://2389.ai/posts/brain-dump-to-blog-post/index.md) — Turn messy coding sessions into shareable learning resources with an LLM-powered workflow that helps you research, build, and document as you go. - [Experimenting with GraphRAG: Adding Knowledge Graphs to RAG Pipelines](https://2389.ai/posts/experimenting-with-rag/index.md) — Blending knowledge graphs with RAG pipelines unlocks richer, scalable insights—bridging the gap between granular retrieval and holistic understanding. - [Self-Learning LLM Agents: A Fractal Approach to Domain-Specific Knowledge](https://2389.ai/posts/self-learning-llms/index.md) — Exploring how agents can autonomously build and evolve their own domain expertise—scaling from generic LLMs to self-learning, specialized assistants. - [Team Spirit Matters: How Collaborative Context Boosts Multi-Agent LLM Performance](https://2389.ai/posts/team-spirit-matters/index.md) — Adding social accountability to multi-agent workflows boosts the depth, coherence, and empathy of LLM responses—mirroring real-world teamwork dynamics. ## Sitemap Parent: [2389 Research, Inc](https://2389.ai/index.md) Site index: [llms.txt](https://2389.ai/llms.txt) · [sitemap.md](https://2389.ai/sitemap.md) · [HTML](https://2389.ai/blog/)