What we learn building open-source enterprise AI on customer hardware — open weights, agents, event-driven architecture, and the human side of the AI shift. No fluff, no PR launches, just what we wish someone had written for us.
LLaMA 4, Mistral Large 3, Qwen3, DeepSeek V3, Gemma 4. The open weights have closed the gap with managed APIs — and for enterprise workloads, they pulled ahead. Here’s what we’re seeing in real customer deployments, and why we no longer recommend cloud LLMs by default.
Open SourceThe biggest political shift of the decade isn’t about politics. It’s about who owns the means of computation — and the answer is increasingly “everyone.”
ArchitecturePolling is dead. MCP added round-trips no one needed. The AI stack is converging on the same answer the financial-trading and gaming worlds reached 20 years ago: events, not requests.
People · Future of WorkNot “prompt engineering.” Not “AI literacy” in the LinkedIn sense. The real skills are older, weirder, and harder to fake — and they’re what separates the people building useful AI from the people watching it happen.
Multi-Agent · ArchitectureForests share resources through underground fungal networks. Multi-agent systems share state through Kafka. The analogy is closer than it sounds — and it changes how we think about resource brokerage, backpressure, and resilience.
Philosophy · On-PremThe fear-mongering misses the real story. With open weights running on your own GPUs, AI becomes what the calculator became to accountants: a tool that makes the human in the loop better, not redundant.
Vibe Coding · Voice-FirstOur team calls it vibe coding. You speak the intent, the agent ships the diff, you review and iterate. It’s not no-code. It’s not low-code. It’s a different relationship with the machine — one most teams haven’t internalised yet.
We send one email a month. Long-form, no clickbait, no “5 ways AI will change your industry” lists. Drop us a note and we’ll add you.