Turns out de minimis is da big deal

Hate to say I Told You So

Perfecting Equilibrium Volume Three, Issue 31

Now it's all out and you know
'Cause I wanted to
Turn my back on the rot that's been planning the plot
Because I'm gonna

No need for me to wait
Because I wanna
No need, two, three, too late
Because I'm gonna

Hate to say I told you so
Alright
Come on
Do believe I told you so

The Sunday Reader, Feb 16, 2025

Tech god Marc Andreesen – coauthor of Mosaic/Netscape, the browser that launched the Web revolution – mused last week that people must be shocked at the way Elon Musk and DOGE were using Large Language Models to revamp the way government budgeting works.

Maybe most people were shocked. But Perfecting Equilibrium readers were chatting about this before DOGE launched. The way Large Language Models – so called AIs – were used to defenestrate the last budget bill of the Biden administration was a DOGE prequel.

Now preening is good fun and all, and I may have had the Hives Hate To Say I Told You So on repeat for a bit. But ya gotta admit, no one was talking about USAID or the de minimis import loophole during last year’s interminable campaigns.

But as always I’ve got what killed the cat, and therefore am much more curious about what happens next than nostalgic about stuff that’s already happened.

And this is one of those most interesting times when you can actually watch the world change. As always, the politics doesn’t interest me. How DOGE is – I suspect – using LLMs does interest me. And then, of course, de minimis interests me because it allows me to talk about Feola’s Corollary to the Law of Unintended Consequences.

One of the interesting things I’m seeing with DOGE is the way they seem to have standardized on teams. It’s taking me forever with the AI Powered Virtual Newsroom project partly because the tech is in the “Let’s Plug This In and See What Happens!!!” part of the show. Nerds are building workflows and agents, CAGs and RAGs, and arguing incessantly about the best approach. And me being me, every time there’s a new rabbit hole…well, I’ve got to dive right in!

So it’s interesting to see that DOGE has come up with a standardized team that it sends into each institution to audit: one team lead; one engineer; one Human Resources specialist; and one attorney. Now methinks that will never be a standard team at a startup or an enterprise – it’s just too expensive – but the approach is quite clever. The usual practice with new tech tools is to arm an army of nerds and unleash them.

This never ends well.

For one, an army of nerds is like an army of cats. More time is spent herding than achieving anything resembling progress. Assign them to audit a legacy accounting system written in FORTRAN? When you check on them you’ll find most of the team is engrossed in a project improving and updating the FORTRAN code. And very little auditing is being done.

The interesting thing about the DOGE teams is that they contain more Subject Matter Experts than coders. Which is a smart thing. SMEs can put data into context. Data on its own is all well and good, but to be really useful it must be understood in context.

Think of it this way: Is $6 billion good or bad?

Well, if you open up your IRA and it says $6 billion, you’re screaming and doing cartwheels and W00Ting all over the place. And if Apple’s Chief Financial Officer opens up that company’s cash reserve account and sees $6 billion, you’d hear crying and yelling and screaming: WHAT HAPPENED TO THE OTHER $25 BILLION!?!???

Context turns data into information.

It’s early days yet, and LLM teams will evolve and change before settling into best practices.

I haven’t been able to find anything on how the DOGE teams are working…but when has that ever stopped me? I have some SWAG! (Scientific Wild A$$ Guesses)

From the speed that they are getting results it’s clear the DOGE teams aren’t training or tuning LLMs on government data. My SWAG is that they are bringing in trained and tuned LLMs, and then loading the government data as a RAG.

Retrieval Augmented Generation gets around a lot of the nightmarish “hallucinations” and other craziness exhibited by AIs, such as diverse Nazi stormtroopers and gluing cheese to your pizza to keep it from sliding off.

A RAG is a custom document and data repository that is added to an LLM. Drawing answers from these custom datasets greatly improves accuracy while cutting down on the crazy. By the by, this is the model I’m working on for the Virtual Newsroom project.

Speaking of AI craziness, the government of France led by President Macron last month launched Lucie, a French AI chatbot. The French government said Lucie would bring trustworthiness, fairness, accountability, and European values while countering the world domination of the English language.

Lucie didn’t last a week.

Lucie told one user that cows’ eggs are a healthy, nourishing food source. Lucie told another user that a cartoon dog had won the 2017 French presidential election.

Do you remember King Herod the Great? Contemporary of Jesus and Ceaser? Maybe you remember him from the bible, or from Jesus Christ Superstar.

Lucie remembered him slightly differently. Herod the Great, according to Lucie, “played an important role in the development of the atomic bomb.”

Missed it by…2,000 years.

The great Steve Ross has some thoughts on these LLM struggles from his experience working with neural networking for straight statistical work. He commented on the way Deep Seek caught up quickly with other LLMs.

The calculation flow is similar and as far as I know, the tweaks and shortcuts have to be similar.

Usually, in the first step of neural networking, the software computes a bunch of different curve fits, usually in a set of standard ways with no particular regard for underlying physics. So you get a large bunch of 2- or 3- or x- dimensional fits that do respect asymtotics and other obvious limits but ignore most or all "real world" limits.

In contrast, a "complex systems" researcher would look at the raw data curves, fit a huge-order polynomial and then factor the polynomial to tease apart different effects that combine to give the final answer.

A lot of the AI company tricks here, and I assume in China, involve that pre-processing trick to cut machine cycle needs.

Another issue is deciding how many neurons to use in the neural network. The rule of thumb is "number of data fields, minus 1." But when doing neural networks to calculate pure statistical problems, the real world presents data with a lot of empty fields. When I'm doing medical data work, there are often just 100 or 200 data records, 250 possible fields, but only 30 or 40 fields filled in, in any particular typical record. Using the "rule of thumb" 249 neurons would be a waste vastly increasing processing time for little or no insight. This is a sparse matrix.

I grew up with sparse matrix analysis.

I wonder if by chance or with luck of having a wise or foolish old man on staff, or by stealing existing analyses (as OpenAI charges), Deepseek has cut its corners. The solution only works if you walk a path others have blazed. You can't use the strategy to jump ahead.

That was my conclusion, too – they took shortcuts to catch up. Which lets you quickly catch up, but is useless for forging ahead.

Finally, let’s talk about Feola’s Corollary to the Law of Unintended Consequences: Closing old loopholes to stop hustlers creates new loopholes that entrepreneurs immediately exploit.

The de minimis exemption made sense when it was passed years ago. Why make a box of cookies or clothes sent from one family member to another run through the entire customs gauntlet? So packages under $800 were exempted.

But tech progresses tirelessly, never sleeping, always advancing. And as shipping and fulfillment systems were automated Amazon deliveries became a thing. And then Chinese entrepreneurs realized they could cut out customs and tariffs by simply shipping products directly to US consumers.

So the Pagani Design Meteorite Chronograph will cost you $83.99 from Amazon, which buys them by the pallet, and then processes them through customs and tariffs. You can buy the same watch from AliExpress – sort of the Chinese Amazon – for $44.84. Shipped directly to you. No customs. No tariffs.

This is, obviously, unfair. And it’s a loophole that’s sure to be closed in a logical manner.

Just remember: we’re closing a loophole created by decisions that were logical at the time.

I cannot wait to see what loopholes these logical fixes create.