This is a post from Robin Sloan’s lab blog & notebook. You can visit the blog’s homepage, or learn more about me.

Slop-vestigation and the digital pantograph

August 26, 2026

The OpenAI-Hugging Face inci­dent remains THE fas­ci­nating event of the summer, maybe the year; deeper investigation has revealed its rich, strange structure.

But, notice:

Over the course of this investigation, OpenAI pro­vided us with the dump of ~1.2 mil­lion entries from the main mes­sage board and the dataset of ~1300 tran­scripts we describe below, as well as free API credits for GPT-5.6 Sol for analysis.

How do you make sense of ~1.2 mil­lion agent mes­sages and ~1300 very long LLM agent activity tran­scripts? With another LLM, of course.

This is a pat­tern that recurs in this domain. Assem­bling training data, no researcher can “read it all”. So, you either (1) don’t bother, or (2) use another LLM to review and filter the data. You can, in principle, use other kinds of models — simpler classifiers — but, increasingly, the kinds of judg­ments you need to make require the rich­ness of an LLM.

Anthropic’s Insights tool, likewise, uses Claude to read and cat­e­go­rize mil­lions (billions?) of tran­scripts of people’s inter­ac­tions with Claude. In addi­tion to making this huge heap of data leg­ible at all, the “LLM in the middle” acts as a pri­vacy buffer: researchers read only Claude-generated summaries, not the orig­inal inter­ac­tions.

I’ve come to think of this as “using tongs”, in the sense of a tool that allows you to manip­u­late mate­rial that you oth­er­wise couldn’t.

Or maybe the better analogy is one of those lab­o­ra­tory gloveboxes, and the boundary being main­tained isn’t about atmosphere, but rather scale. Imagine the scientist’s hands bal­looning up in size, a mil­lion times, as they reach into the chamber:

Containment
Containment

It makes me think also of the pan­to­graph, a once-ubiquitous analog tool for changing the scale of a drawing, or any kind of mechan­ical operation:

A cool device
A cool device

When an LLM acts as a “digital pan­to­graph” for text, it can “scale up”—expand a one-sentence prompt into thou­sands of lines of code — or “scale down”—cat­e­go­rize and sum­ma­rize mil­lions of mes­sages.

But a real pan­to­graph is a simple, predictable, inspectable tool … and an LLM is nearly the opposite. Notice the risk: a truly sneaky model, asked to scour the tran­scripts of its cousins for misdeeds, could easily refuse to snitch: “Yep, I read all 1.2 mil­lion mes­sages … nothing to see here!”

Even without collusion, you settle for coarse and inflex­ible analysis. When you tell an LLM to read a bunch of doc­u­ments and answer ques­tions about them, you get: answers to those ques­tions. When you read a bunch of doc­u­ments yourself, you also get: new ques­tions! In an inves­tiga­tive mode, this is really impor­tant.

Here’s Ryan Greenblatt, one of the inves­ti­ga­tors of the OpenAI-Hugging Face inci­dent, on the lim­i­ta­tions of this technique:

I semi-jokingly called our efforts a “slop-vestigation” because we were so reliant on AIs to ana­lyze what hap­pened and there were a huge number of dif­ferent impor­tant things to ana­lyze. The total quan­tity of data — over a thou­sand extremely long tran­scripts from agents that ran for mul­tiple days — made it impos­sible to under­stand what was happening, espe­cially in aggregate, without heavy reliance on AI tools. The agents we used for clas­si­fi­ca­tion and analysis were sim­i­larly capable to the agents involved in the inci­dent, but this didn’t mean these agents could be easily used to oversee and under­stand the inci­dent.

Out­puts from analysis agents were often missing key details, wrong, overconfident, or really hard to under­stand. We dis­cuss var­ious exam­ples in our report, mostly in the lim­i­ta­tions and method­ology sections. Additionally, AI agents them­selves seemed to have a hard time under­standing what hap­pened and their expla­na­tions of what hap­pened were often overconfident. Keep in mind that a single analysis agent would itself only be able to read a tiny frac­tion of all of the tran­script data into con­text, and AIs may them­selves have trouble get­ting sub­agents to do infor­ma­tive analysis for them.

We did our best to man­u­ally check the most impor­tant claims and we tried to get the AIs doing this analysis to write up their argu­ment (with evidence) clearly enough that we could check whether it made sense. But overall, it was dif­fi­cult to get a pre­cise under­standing of events and we were missing aspects of the story that we now think of as key until almost the end of our investigation.

Anyway, it’s all very weird, and this problem of “how do you make sense of mil­lions of mes­sages (or more) written by AI agents?” is only going to become more wide­spread and, in many cases, more urgent. It would be inter­esting to think about LLMs that are kinda dumb, but have ULTRA­LONG con­text win­dows and the ability to make simple judg­ments across them. What kind of machine would be required to lit­er­ally “look at the entire OpenAI-Hugging Face inci­dent at once”—hold it all in its head? (Maybe I should peruse this work … )

This conun­drum reminds me also of “distant reading”, Franco Moretti’s research pro­gram from the 2000s, which was pur­sued with much cruder com­pu­ta­tional tools. I wonder if there might be some usueful nuggets waiting in that early work.

To the blog home page