Hal Beresford · Curiosities

A cabinet of curiosities

Exhibits collected while building forecasting systems — from mathematics, from money, and from the weather instruments that quietly decide who was right. Each one mounted, dated, and labeled.


Exhibit 001 · Mathematics · July 2026

The dimension ladder

A point has no dimensions. Drag it one inch and it sweeps out a line. Drag the line sideways and it sweeps a square; drag the square upward and it sweeps a cube. Every rung of the ladder is the same single move — sweep what you have in a direction it doesn't yet occupy — and nothing about that move requires you to stop at three.

Figure 1 — The n-cube projected down onto your screen. For n ≥ 4 you are looking at a shadow of a shadow: the same trick a drawing of a cube already plays, played one more time.

The counts keep the pattern honest. A square has 4 corners, a cube 8, a tesseract 16 — each sweep copies the whole figure and joins every corner to its copy, so the vertices double and the edges follow n·2n−1. By six dimensions there are 64 corners and 192 edges, and the shadow starts to look like lace. The object is no less real for being unvisualizable; it is exactly as real as the number 26.


Exhibit 002 · Mathematics of money · July 2026

Compounding ignores zeros

Take one dollar and one hundred million dollars and compound both at 7% a year for sixty years. On an ordinary axis the comparison looks absurd: the fortune climbs toward six billion while the dollar never visibly leaves the floor.

$100,000,000 $1
data table
year$1 grows to$100M grows to

Switch to a logarithmic axis — the axis on which multiplying is what moves you — and the two curves become the same straight line, eight orders of magnitude apart. Same slope. Same doubling time, a shade over ten years. The dollar and the fortune are making identical jumps: 7% is 7%. The zeros only record where you started, never how fast you are going.


Exhibit 003 · Instruments · Denver, June 26, 2026

The two summer minutes Denver spent in winter

At 9:33 AM on June 26 the Denver International Airport weather station reported an outdoor temperature of 77°F, climbing toward what would become a 90°F afternoon. But at 9:34 the station fell silent. For the next twenty minutes every field it published — temperature, wind speed, wind direction — came across as -999 (no value).

When the weather station came back online at 9:54 AM, for two minutes it insisted the air was 24.8°F. Two readings later the station was back at 77°F as though nothing had happened:

Figure 3 — Every one-minute reading Denver International transmitted that day as recorded off the wire. The gap in gray illustrates where the station went silent. Chart: KDEN one-minute observations recorded as received.
the readings either side of the gap
local timereported

What made this more than a shaggy-dog story was that the official record did not catch the error at first. That afternoon the National Weather Service's initial climatological report for Denver asserted that the morning's low had been 24 degrees, a record for June 26. Eleven minutes later the National Weather Service released a corrected report stating that the morning's low was actually 56°F.

430 PM MDT   MINIMUM   24R   856 AM   record 40   1968
441 PM MDT   MINIMUM   56    440 AM   record 40   1968
The minimum-temperature line from each report as issued. R flags a record; the standing June 26 record was 40°F, set in 1968. National Weather Service Denver/Boulder, climate report for Denver, June 26, 2026: 4:30 PM MDT · 4:41 PM MDT.

Data quality is when the data is able to perform the function for which it was collected. In Denver the morning of June 26 the data quality was suspect for all of the climate archives, insurance triggers, and settlement systems up until the National Weather Service's corrected afternoon report. No matter how trusted the source, validation remains a key step toward ensuring data has adequate quality before it is used.


Exhibit 004 · Infrastructure · July 30, 2026

Name the constraint before you buy the service

Working through AWS's tutorials, I kept asking the same two questions about every service I hit: what constraint does this remove, and do I have that constraint?

The second question is the one that saves money. Hadoop got a no — my data doesn't need distributed processing at that scale. Aurora got a no — my workload is analytical, not transactional. Lambda got a no — my processes run long and hold persistent connections, and a fifteen-minute execution limit and statelessness rule that out. Three no's, three different reasons. None of them a judgment about the service.

So I drew the table. Nine constraints, four providers, and the discipline of naming the problem before naming the product.

Nine infrastructure constraints and the matching service from each of four cloud providers.
the constraint AWS Azure Google Cloud DigitalOcean
A machine that is always on EC2Virtual MachinesCompute EngineDroplets
Work that runs when triggered LambdaFunctionsCloud Run functionsFunctions
Containers without hosts FargateContainer AppsCloud RunApp Platform
Cheap, durable bulk storage S3Blob StorageCloud StorageSpaces
Many clients need ACID + failover AuroraSQL DatabaseCloud SQLManaged PostgreSQL
Analytics over huge datasets RedshiftSynapseBigQueryManaged OpenSearch
Producers must decouple SQSService BusPub/SubManaged Kafka
Knowing when something breaks CloudWatchMonitorCloud MonitoringMonitoring
Training and serving models SageMaker AIMachine LearningGemini Enterprise Agent PlatformGPU Droplets
Figure 4 — The same nine constraints across four providers. Read a row, not a column: the question is which cell matches a constraint you actually have. Every product name verified against vendor documentation on July 30, 2026. Cloud naming moves; see below.

Building it taught me more than reading about it would have, mostly through being wrong.

My first draft claimed DigitalOcean had no answer for decoupling producers from consumers. It does — Managed Kafka, generally available since 2023. My second draft claimed they had nothing for training and serving models. They market GPU model training and fine-tuning on dedicated pages. Two false gaps, both in the same column, both trivially disprovable by anyone who looked. The pattern is worth naming: every error was a negative claim. Asserting a service exists is checkable in one search; asserting one doesn't exist requires you to have looked everywhere, and I hadn't.

I also caught myself optimizing for the wrong thing. Wanting each cell to be distinct, I replaced Google's Cloud Run with GKE Autopilot so no product name would appear twice — a tidy table, and a worse one, because Google's own documentation positions Cloud Run as the no-infrastructure option and never calls Autopilot serverless. I had chosen a formatting property over an accurate one and told myself it was rigor. Cloud Run appears twice above. That repetition is the truth: Google folded Cloud Functions into Cloud Run in 2024, and the products really are one family now.

Microsoft Fabric never made it into the table at all. It's where Microsoft is steering new analytics work, but its own documentation compares Fabric to Azure rather than placing it inside Azure, and notes you don't need an Azure account to use it. In a table organized by cloud provider it has nowhere to sit. Synapse holds the cell instead.

That verification date matters more than it should. In the course of checking I found that Google had renamed Vertex AI to Gemini Enterprise Agent Platform, AWS had renamed SageMaker to SageMaker AI, Cloud Functions had become Cloud Run functions, and AWS App Runner had stopped accepting new customers. Four changes in roughly two years. Assume some of this is stale by the time you read it.

A few cells are judgment calls rather than facts, and reasonable people will place them differently. DigitalOcean's Managed OpenSearch is the weakest — it's search and log analytics, not a warehouse in the sense BigQuery and Redshift are. Aurora is right, though Aurora DSQL is arguably sharper for the specific pairing of ACID guarantees and fast failover. Synapse is accurate but reads a step behind Microsoft's current direction.

The table isn't the point. The question is: name the constraint first, and most of the buying decision makes itself.


More exhibits in preparation.


Contact