Intel
Writing
Field notes on AI, federal technology and building products.
More: how I work.
New here? Start with four places.
108 articles · page 4 of 5
AI Will Not Fix Your ERP Forecast
Forecast accuracy is bounded by the chart of accounts and the close calendar, not by the model. Add AI to a broken model and you get faster wrong numbers.
AI-First IT Has to Change the Unit of Work
Buying agents does not change IT. Changing what counts as a unit of work does, and that decision belongs to whoever owns change control.
TOPS Is Not a Workload Number
Edge AI runs out of memory and memory bandwidth before it runs out of compute. Most hardware requirements measure the thing that is not binding.
AI Projects Fail on Decision Rights
The model is rarely why a pilot dies. The reasons are who owns the number, who can say no, and what happens when the output is wrong.
An ERP Will Not Fix a Failed Audit
The department has never passed a full financial statement audit. A new ERP does not change the evidence problem that causes the failure.
Multimodal Models Do Not Weigh Evidence
A multimodal system can read the scan, the note, and the lab value. Ask what it does when they disagree, and the demo ends.
The Quantum Risk Is Your Data's Shelf Life
Whether quantum computing arrives in five years or twenty, data captured today is at risk now. The migration clock starts at your retention period.
The Case for Small Models Is Where the Data Lives
Small language models are not winning on benchmark scores. They win when the data cannot leave the building and the task is narrow enough to be checkable.
Spatial Computing Wins Where Your Hands Are Busy
The headset did not kill the screen. It found a narrower job: putting a procedure in front of someone whose hands are already occupied.
Your Data Pipeline Fails at the Seam, Not the Stage
Most model incidents are pipeline incidents. The stages are fine. What breaks is the contract between training and serving, and the untested backfill.
Pick the Model After You Know the Cost of Being Wrong
Algorithm choice is the last decision in a model project. The label supply, the cost of a false positive and the retraining path decide most of it first.
Pick ML by Label Budget, Not by Algorithm
Supervised, unsupervised, reinforcement: the choice is mostly about the evidence you have and what you can afford to label.
If You Cannot Name the Decision, You Have a Topic
Most AI use case frameworks start with a problem and a KPI. Start one step earlier, with the decision, the person who makes it, and the record it leaves.
The AI Solutions Manager Owns the Boundary
The role is not model selection. It is owning the line between the model and the workflow, and the evidence that the line holds.
The Data Janitor Is the Domain Expert
Cleaning data is not janitorial work. Every fix encodes a judgment about what a row means, and that judgment belongs to the people who own the process.
A Million Tokens Did Not Kill Retrieval
Long context moved the boundary between retrieving and reading. It did not fix the middle of the window, and it did not create an audit trail.
FedRAMP 20x Changes the Unit of Evidence, Not the Form
OSCAL made the paperwork machine-readable. FedRAMP 20x goes after the review itself, and the review was always the slow part of an ATO.
Sovereign LLMs: Open Models for On-Prem Government AI
Downloading the weights is the easy part. Sovereign AI requires provenance, evaluation, and a hardware plan you can defend in an audit.
Data Mesh in Defense Is an Accreditation Decision
A data mesh is a governance decision, not a platform purchase: who owns each domain, who may query it, and who signs for the result.
The IL6 Problem Is the Model Catalog
Microsoft's own roadmap shows the classified cloud platform is authorized at IL6 while model hosting is still forecasted. The catalog is the gate.
The Benchmark Number Is a Claim About the Harness
Claude 3.7's TAU-bench score came with a footnote about scaffolding. That footnote is the most useful thing in the announcement.
Colossus Was a Construction Project, Not a Model Story
xAI put 100,000 GPUs online in 122 days by reusing a closed factory and buying temporary power. The constraint was never the chip.
GCC High Is a Platform, Not a Compliance Badge
GCC High migrations fail on identity and data residency before they fail on technology. What to sequence, and what breaks when you sequence it wrong.
Gemini 2.0's Real Bet Is Model Tiering
Google shipped a flagship and a cheap tier in the same release. The useful question is which task deserves which model, not which model wins.
End of page 4 of 5. The archive continues.