Writing on enterprise AI.
Field notes on enterprise AI architecture, sovereign deployment, and the trade-offs of building Generative AI inside regulated walls.
Your model weights should outlive the process that serves them. SGLang shipped that idea on August 21.
"Multimodal-native" has quietly come to mean a tool call rather than a model property. Qwen's new plugin repo is honest about that; a lot of the commentary around it is not.
Prefill and decode are two different workloads. On July 23 the hardware stopped pretending otherwise.
MCP just deleted its session handshake. That one change does more for enterprise agent deployment than any model release this month.
The first fully agentic ransomware just ran. It used the same primitives you build defensive agents on.
The biggest cost lever in an agent stack right now is a router that knows when not to call the frontier model.
Everyone quantizes the weights and forgets the KV cache. On long context, that cache is what blows your VRAM budget.
An open-weight model just matched the frontier on long-horizon agentic coding and MCP tool use. If you run air-gapped, the build-vs-buy math quietly flipped this week.
Prompt injection isn't a bug you'll patch. It's a property of how the model reads, and OWASP's new report stops pretending otherwise.
Every MCP tool you wire into an agent is a new unauthenticated door into your network — and the data says most teams forgot the lock.
Gemma 4 12B just deleted the vision and audio encoders. Most multimodal models still bolt on three separate networks. Google ships one weight space.
The headline last week was a new model. The lesson was an architecture.
Your next long-context model probably won't run quadratic attention.
Most production LLM eval harnesses ship with one column: accuracy. That's a vibes check, not an eval.
Your agent doesn't have a JSON problem. It has a decoding problem.
HuggingFace quietly put Text Generation Inference into maintenance mode in December 2025.
Autonomous agents touching production is a terrible idea.