What it actually looks like to run a serious training pipeline at home in early 2026, the data prep, the orchestration, the evaluation, the operational discipline. Less hand-wavy than the typical write-up, more boring than the marketing pitch.
Twelve specific predictions for 2026, calibrated against the patterns I missed in 2025. Less hand-wavy than the typical year-ahead piece. I'll grade them plainly at the next checkpoint.
The bridge week. SB 53 went live on January 1, the year-end retrospectives all said roughly the same thing about 2025, the 2026 prediction pieces from Sequoia and a16z dropped with surprisingly compatible takes, and CES 2026 is sitting on Tuesday's runway. What I make of the turn of the year.
End-of-year recap. Most of the year's narrative was hype layered over a smaller set of substantive shifts. Worth being explicit about what actually changed and what didn't, before the 2026-predictions piece tries to extrapolate from the noise.
Multi-tenant from day one is roughly 2x the cost of the single-tenant version. The hidden costs aren't in the code, they're in auth, audit, query routing, observability, support tiering, and billing. Often the cheaper move is to start single-tenant and earn your way in.
MCP went from emerging standard to genuine substrate in 2025. The use cases that have actually shipped on it are more interesting than the broad-platform-narrative version. Worth being specific about which ones to track and why.
Christmas week, so most of the news is the year itself sitting up and asking to be summarized. Nvidia buys Groq for $20B in the largest deal in company history, Z.ai ships GLM-4.7 as the new top open coding model, and the Challenger AI-layoffs number lands at 54,836. A partial year-end take.
Decisions as Code specifies the standard. OPA Gatekeeper enforces it. Constraint templates as the policy definition layer. Required tags, model version pinning, namespace quotas, network policies for AI workloads. The 'specify, then verify' loop, finally complete.
The category I've been writing about for two years arrived faster than the broader industry conversation acknowledges and slower than the early enthusiasts expected. The shape of what's actually here is more nuanced than either timing claim alone suggests.
The atomic-unit metaphor in software isn't decoration. It's a working design pattern: small immutable typed primitives (atoms) compose into larger functional units (molecules). What makes the metaphor useful, versioning, lineage, composition rules, and what makes it a marketing trap.
A nine-month checkpoint on the takes I had in March 2025 about how the year would shape up. Some held; some bent; some were plain wrong. Worth being explicit about which is which.
The week the model arms race went thermonuclear: GPT-5.2 lands as the answer to a 'code red' memo, Google ships Gemini 3 Flash three days later, NVIDIA drops Nemotron 3 as open weights. Meanwhile Trump signs an EO to preempt state AI laws and Accenture's AI-driven RIF crosses 11,000.
