I'm curious how you would airgap a datacenter-scale training run. No reading training data from a cloud store, someone from OpenAI flies out to Abilene or whereever with a suitcase full of SSDs? When researchers in SF write code for a new RL environment, they ping somebody who sticks it on a flash drive and walks over to the server rack with the GPUs? Has anyone ever trained an LLM like this?

Maybe you don't mean a literal airgap, you just mean configuring Internet access off. But... they did that. That's why the agents had to hack Artifactory to get it...

> how you would airgap a datacenter-scale training run

You could start with standard firewall rules and VLANs. Pretty basic stuff we've had for decades.

That's not an airgap though.

Data diodes exist and are common in high-security environments https://en.wikipedia.org/wiki/Unidirectional_network

The hardware literally only goes one way. Nothing can "escape" without rewiring cables.

Recall that multiple of the Chinese LLM providers got caught sending at least some of their queries to Claude. That would be a pretty funny way to get caught in turn doing something by the US government.

Why's that?

The article says that scores began drifting down around 2012. Gemini is telling me that NAEP peaked in 2013. Elsewhere in this thread there's a claim of an inflection point around 2018 in PISA, which eyeballing Gemini's numbers lines up with NAEP.

Does this correlate with the racial proportion data?

If you graph actual NAEP since 1998 versus a 1998 group-fixed score, you'll see increases (1998-2002, 2005-2013, 2015-2017) and multiple declines twice taking the line back down to the 1998 linear trend prediction (2002-2005, 2017-2024).

It is wrong to get overly exact here though. We are talking about a overall score variance of just 10pts in the face of an achievement gap of nearly 30pts, with an absolutely massive 20% decline in white population share over the 26 year period.

If you're seeing trends in your residuals that's a violation of the independence assumption of linear models (assuming you're using about a linear model since you're talking about "linear predictions").

I'm in a different field but an arbitrary 27 years (1998-2024) of data and the autocorrelation both would get flagged in a review for me. Not to give statistics homework but you should test and correct the autocorrelation issue you're describing in the model and if the data goes farther back I would go farther back, too (with how you're describing the errors going off pattern then back on it sounds like the inference here would be at least somewhat unstable depending on year chosen).

Edit: these are the assumptions and basics of how to correct for violations, in particular you're describing a violation of assumption 2 but you should test for all of them - https://www.statology.org/linear-regression-assumptions/

I was checking if there was a response and upon reading your comments again, looks like your method is less a model and more just some ad hoc procedure. Based on what your conclusion is it sounds like a linear model based on demographic data would work and linear models are easy to train. Though by your summary there are probably going to be issues with the model and you have the same issues with your procedure, you just can't tell because it hasn't been rigorously specified.

There's a channel on YouTube called Statquest that teaches statistics in a pretty accessible way if you're interested in analyzing this sort of data.

I learned my statistics in university, not Youtube. You are confusing yourself about residuals.

To make my point in the original post, I held the group scores fixed and extrapolated combined scores based on changing population sizes. This is straightforward. Simple even.

\(\widehat S=\sum_g p_gS_g,\)

Why did I call this "linear"? (Not a "linear model", your term, because it is not). The model function actually tracks the population function. Over the 25-year period, however, this trend is in fact linear because the White-to-Hispanic population trend is linear on that timescale. Over a longer timescale it would not be.

What you're describing is a linear model, it just has an ad hoc construction instead of something like a least squares training (https://en.wikipedia.org/wiki/Linear_model). Also linear models are about linear associations and work even with non-linear predictors. Your assumption in the model is explicitly a linear association so how the predictors vary shouldn't affect it.

I don't mean to disparage amateur statistics since often there are interesting things that experts miss that amateurs have insight into but this sort of data with few data points (the data you have doesn't sound independent) and a lot of confounding factors is not trivial to accurately model. Not that everyone needs to learn statistics but this sort of data (and also this topic) probably deserves a bit more of a rigorous approach.

As a very simple example, picking a time range for the model because the model doesn't work when you go farther back adds a lot of potential bias into the model and that along with the non-independence issue (and without looking at the data myself I don't know if there are other issues) are going to lead to overconfidence in conclusions.

You remain confused.

"roma aeterna" is also the sort of alias that's associated with a particular set of political views. (Not conclusive! But a rational thing for someone to notice.)

There is a particular set of political views associated with people who read classical languages? I fail to see what you are calling "rational" here.

He's trying to imply that you're a crypto-fascist because you presented some hate-facts.

It's sad that there can be no straightforward discussion of issues like this. It doesn't help anyone except people who are paid from ineffectual educational and social programs.

The DoD is free not to sign contracts. Having them be able to get a company blacklisted from a whole sector just because they couldn't agree on terms with that company is obviously very different than that.

The idea that "fair for the military to object to" implies "fair for the military to ban from their supply chain in any capacity no matter how unrelated the use" is so outlandish that it's never been done before (to a US company), and I think we can be confident that the DoD has frequently objected to would-be contractors' terms.

The claim that it being done for the first time - by a Secretary of Defense denounced the politics of the company in the same text where he announced the designation - isn't "political", is something I have a hard time believing a functioning adult who isn't employed defending the government in court could adopt.

But hey, maybe in some language this is a "textbook" official statement with no "politics":

  Cloaked in the sanctimonious rhetoric of “effective altruism,” they have attempted to strong-arm the United States military into submission - a cowardly act of corporate virtue-signaling that places Silicon Valley ideology above American lives.

Nobody is suing to force the DoD to buy anything, so maybe you shouldn't be accusing others of failing to refer to the topic of discussion.

That's true legally, but the government's public statements did initially very explicitly present this as a ban on any dealing.

"Anthropic a Supply-Chain Risk to National Security. Effective immediately, no contractor, supplier, or partner that does business with the United States military may conduct any commercial activity with Anthropic."

Presumably, according to the defenders of the designation in this thread, this was an innocent mistake that doesn't indicate any political intent.

Why has this never previously been done to an American company from which the DoD decided not to purchase something?

What is the evidence that Anthropic said this?