05 · Field notes

Writing on enterprise AI.

Field notes on enterprise AI architecture, sovereign deployment, and the trade-offs of building Generative AI inside regulated walls.

17 notes
20 · Aug 25, 2026
Your model weights should outlive the process

Your model weights should outlive the process that serves them. SGLang shipped that idea on August 21.

  • weight-cache-daemon
  • cold-start-recovery
  • cuda-ipc-zero-copy
  • restart-slo
  • quant-format-coverage
19 · Aug 12, 2026
Where do the pixels actually go?

"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.

  • multimodality-as-tool-call
  • skill-plus-mcp-packaging
  • capability-install-boundary
  • default-provider-egress
18 · Jul 29, 2026
Prefill and decode are two different workloads

Prefill and decode are two different workloads. On July 23 the hardware stopped pretending otherwise.

  • disaggregated-prefill-decode
  • helios-rackscale
  • wafer-scale-decode-offload
  • vendor-modelled-benchmark
17 · Jul 21, 2026
MCP just deleted its session handshake

MCP just deleted its session handshake. That one change does more for enterprise agent deployment than any model release this month.

  • mcp-stateless-protocol
  • session-affinity-removal
  • tool-server-horizontal-scaling
  • sampling-deprecation
16 · Jul 14, 2026
The attacker went agentic before you did

The first fully agentic ransomware just ran. It used the same primitives you build defensive agents on.

  • agentic-ransomware
  • autonomous-attack-chain
  • agent-tempo
  • credential-blast-radius
  • self-narrating-payloads
15 · Jun 30, 2026
Route by task, not by loyalty

The biggest cost lever in an agent stack right now is a router that knows when not to call the frontier model.

  • mid-tier-agent-economics
  • capability-routing
  • effort-dial-cost
  • cost-per-task
14 · Jun 23, 2026
Everyone quantizes the weights and forgets the KV cache

Everyone quantizes the weights and forgets the KV cache. On long context, that cache is what blows your VRAM budget.

  • kv-cache-quantization
  • rope-aware-bit-allocation
  • 2bit-kv
  • block-gtq
  • long-context-serving
13 · Jun 16, 2026
Open weights just hit the frontier. Read the token bill.

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.

  • open-weight-frontier-parity
  • glm-5.2
  • swe-bench-pro
  • terminal-bench
  • agentic-coding-onprem
  • token-efficiency-tco
12 · Jun 11, 2026
Prompt injection is permanent. Design around it.

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.

  • prompt-injection
  • lethal-trifecta
  • agents-rule-of-two
  • command-data-boundary
11 · Jun 03, 2026
Your agent's tools are your new perimeter

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.

  • mcp-security
  • taint-style-vuln
  • unauthenticated-tools
  • agent-attack-surface
  • tool-poisoning
  • viper-mcp
10 · Jun 03, 2026
Gemma 4 12B deleted the encoders

Gemma 4 12B just deleted the vision and audio encoders. Most multimodal models still bolt on three separate networks. Google ships one weight space.

  • encoder-free-multimodal
  • unified-embedding-projection
  • native-audio
  • on-device-16gb
  • gemma4
09 · May 28, 2026
The model got the headline. The pattern is the product.

The headline last week was a new model. The lesson was an architecture.

  • multi-agent-orchestration
  • fan-out-verify
  • adversarial-verifier
  • convergence-criterion
  • orchestrator-as-code
08 · May 2026
Subquadratic sparse attention finally shipped

Your next long-context model probably won't run quadratic attention.

  • sparse-attention
  • nsa
  • dsa
  • ssa
  • long-context
  • deepseek-v3.2-exp
  • subquadratic
  • lightning-indexer
07 · May 2026
The LLM eval harness rubric

Most production LLM eval harnesses ship with one column: accuracy. That's a vibes check, not an eval.

  • llm-eval
  • golden-set
  • schema-set
  • calibration-brier
  • latency-budget
  • adversarial-set
  • metric-owner-frequency
05 · May 2026
Your agent doesn't have a JSON problem. It has a decoding problem.

Your agent doesn't have a JSON problem. It has a decoding problem.

  • grammar-constrained-decoding
  • xgrammar
  • outlines
  • structured-output
  • pydantic-router-schema
  • json-reliability
02 · Apr 2026
TGI went into maintenance mode. Here's the on-prem 70B picture now.

HuggingFace quietly put Text Generation Inference into maintenance mode in December 2025.

  • on-prem-serving
  • tgi-maintenance
  • vllm
  • ollama
  • sglang
  • per-user-tok/s
  • quantization-awq-gptq
  • framework-comparison
01 · Apr 2026
The tools are the product

Autonomous agents touching production is a terrible idea.

  • agentic-safety
  • human-in-the-loop
  • tool-surface
  • blast-radius
  • audit-log
  • langgraph
  • production-mutation