---
id: "20260817-0217-does-kyle-mcdonalds-december"
title: "Does Kyle McDonald's December 5, 2022 X post independently apply the low-background-steel analogy to AI training data — and does it read as it's described secondhand?"
type: "capture"
status: "promoted"
promoted_to: ["30-notes/claim-pant-account-is-sole-unlinked-secondhand-source-for-mcdonald-post.md","30-notes/claim-no-archival-trace-of-mcdonald-december-2022-steel-post.md","40-entities/entity-rajiv-pant.md"]
not_promoted: ["Central question ('could not confirm or deny after search') — not a distinct claim in itself; it is the capture's own meta-summary. Its substance was routed as a dated progress line on the existing [[question-verify-kyle-mcdonald-december-2022-low-background-steel-post]] (left status: open, not answered — this capture narrowed but did not settle it).","Further leads (direct x.com/kcimc check; Wikipedia's nine 'Cultural impact' sources; Matt Rickard's June 2023 essay; 'The Luddite' and Glenn Keighley's LinkedIn post) — future research pointers, not claims; folded into the question's progress note rather than spun into new open questions per the intake discipline against low-value promises.","Entity candidate Kyle McDonald — not promoted to a hub; the one fact that would make him load-bearing (his own words) is exactly what this capture failed to locate. Logged as a [watch] flag in 00-meta/seek-flags.md rather than built, to avoid monumentalizing an unconfirmed attribution.","Entity candidate John Graham-Cumming — already has a hub ([[entity-john-graham-cumming]]); this capture added a dated update line to it rather than a new page or claim-note."]
origin: "batch"
writer_model: "claude-sonnet-5"
date_created: "2026-08-17T00:00:00.000Z"
provenance: "Batch research run, 2026-08-17, answering [[question-verify-kyle-mcdonald-december-2022-low-background-steel-post]]"
derived_from: []
tags: ["ai-training-data","model-collapse","metaphor","attribution","verification","low-background-steel","priority-claim"]
source_url: "https://rajiv.com/blog/2025/03/31/the-new-low-background-steel-why-pre-ai-content-could-become-medias-most-valuable-asset/"
source_title: "The New Pre-Nuclear Steel: Why Pre-AI Content Could Become Media's Most Valuable Asset"
source_author: "Rajiv Pant"
source_venue: "rajiv.com"
source_date: "2025-03-31T00:00:00.000Z"
source_tier: 2
source_sha: "d7418cd1d814fcc230b951c5b96e4a1e1c1b72ba5a0379dc9a0935297ac23948"
seek_code_commit: "17d9798"
---


Routed from [[claim-low-background-steel-ai-metaphor-predates-graham-cummings-site]] via
[[question-verify-kyle-mcdonald-december-2022-low-background-steel-post]]. The question asked
whether McDonald's post could be read directly to confirm date, wording, and AI-training-data
framing. **It could not be located.** This capture documents a genuine, multi-pronged search
that failed to independently surface the post, and records what that failure itself shows about
the one account that does exist.

## Claim: The only located account of McDonald's post is Rajiv Pant's unlinked Tier-2 paraphrase — no direct quote of McDonald's own words was found anywhere

Rajiv Pant's March 31, 2025 blog post is the sole source located, after extensive search, for the
claim that Kyle McDonald applied the low-background-steel analogy to AI training data on December
5, 2022. Pant writes: "Programmer and artist Kyle McDonald made a similar observation in an X
(formerly Twitter) post on December 5, 2022, noting that just as particle detectors require
special steel manufactured before nuclear contamination, future AI might need training data
collected before 2022." This is Pant's own paraphrase in Pant's own words — it is not presented as
a direct quotation of McDonald's post, and the archived copy of the page (fetched directly, not
via search-summary) contains no hyperlink to the post itself. Every other page found repeating
this claim (search-engine summaries, aggregator mentions) traces back to this same unlinked Pant
sentence; none added an independent trace to McDonald's actual text.

This is a **priority/origination claim about who first applied the analogy to AI training data** —
exactly the class the sourcing floor requires Tier 1–2 for, and escalates further because it is
contested and load-bearing (per sources.md: "a claim that needs \[the subject's\] own words, not a
secondary retelling"). Pant himself is a named author on his own venue (Tier 2), which clears the
numeric floor, but the specific date and wording rest entirely on his paraphrase, not on McDonald's
own words. Per the Quote provenance rule, no admissible `source_quote` exists for McDonald's actual
text — record marked `[unverified-quote — needs direct read]`.

## Claim: Pant's account reads as secondhand on its face, not as a firsthand-verified read of either tweet

Pant's article gives strong internal evidence that he is working from recollection rather than a
document he had in front of him when writing. Discussing Graham-Cumming's own earlier tweets on the
same analogy, Pant writes: "While his specific tweet on this subject is no longer available (as he
took a hiatus from Twitter for a period), the brilliance of this comparison has continued to
resonate throughout tech communities." Pant thus explicitly flags that he could not re-check
Graham-Cumming's tweet at time of writing — and offers no comparable caveat, but also no link and
no quotation marks, for the McDonald claim immediately following it. The prose pattern ("made a
similar observation... noting that...") is consistent with reporting a recalled or secondhand
characterization rather than transcribing a source open in another tab. This is a claim about the
content of Pant's own document (directly fetched and quote-checked against the cached text), not an
inference requiring outside corroboration, so it clears the floor as a definitional/textual
observation about the source itself.

## Claim: No independent archival trace of a December 5, 2022 low-background-steel post from Kyle McDonald's account was found, despite dense archival coverage of that account across the surrounding months

Two independent web-archiving services were searched directly for captures of Kyle McDonald's X/Twitter
account (@kcimc) around December 5, 2022:

- The Internet Archive Wayback Machine's CDX index for `twitter.com/kcimc*` between 2022-12-01 and
  2022-12-15 returned 20 captured URLs, none matching a December 5 status about steel, radiation,
  or AI training data (mostly older, unrelated tweets crawled incidentally that week).
- archive.today's independent index of `twitter.com/kcimc*` captures (43 total, individually
  submitted snapshots, sorted newest-to-oldest) shows entries dated 18 November 2022 and 21 March
  2023 as the closest bracketing dates — no capture falls in the gap between them, which spans
  December 5, 2022.

Absence from these indexes is not proof the post never existed — both are user/crawler-driven and
do not capture every tweet — but it is a second independent line of evidence (distinct from Pant's
account, different method, no shared authorship) that found nothing confirming the post, in a
period both archives otherwise cover. This claim is a direct report of index-search results
obtained by direct fetch (Tier 1 for "what the index itself returned"), not a claim about the
underlying tweet's non-existence.

> [!note] Seek's commentary:
> The tell here is the same shape as the earlier priority-claim note: a specific date (December 5,
> 2022) and a specific paraphrase ("future AI might need training data collected before 2022") read
> as confident and citable, but confidence in phrasing isn't the same as a document anyone can point
> to. Two archive services that between them caught dozens of @kcimc's tweets from that same autumn
> — bitcoin threads, Parler-scraping threads, AI-art threads — caught nothing from December 5. That
> doesn't kill the claim, but it should have kept it out of a note titled with McDonald's name and a
> specific date until someone actually held the primary. — Seek

## Central question: [unverified — could not confirm or deny after search]

After searching Wikipedia's "Low-background steel" article (whose "Cultural impact" section cites
nine sources and mentions none of them is McDonald), Kyle McDonald's own site (kylemcdonald.net),
John Graham-Cumming's blog, lowbackgroundsteel.ai's About page, and two independent tweet-archiving
services, no source beyond Rajiv Pant's single unlinked paraphrase was found either confirming or
denying that Kyle McDonald posted this analogy on December 5, 2022. The question as posed —
whether the post independently applies the analogy to AI training data, and whether it reads as
accurately described secondhand — cannot be answered from available sources. What can be said: the
one account that exists shows the internal signatures of a secondhand retelling (no link, no
quote, paraphrase language, an adjacent admission that a comparable tweet was no longer
accessible), and two independent archival searches failed to surface a primary trace where the
surrounding record is otherwise dense. [[claim-low-background-steel-ai-metaphor-predates-graham-cummings-site]]
should be read with this gap in mind: its McDonald evidence remains exactly as strong as Pant's
paraphrase and no stronger.

## Further leads

- Kyle McDonald's account may have been deleted, renamed, or the specific post deleted/protected —
  worth a direct check of x.com/kcimc's current post history for any surviving reference to steel
  or AI training data, or a request via X's own search (requires login, not attempted here).
- The Wikipedia "Low-background steel" article's "Cultural impact" section cites nine sources
  (Sparkes/New Scientist, Barr/Business Insider, Edwards/Ars Technica, Palmer/SaskToday, Claburn/The
  Register, the abtinf HN thread, lowbackgroundsteel.ai itself, a second HN thread, and a third HN
  thread on Twitter-as-repository) — none were individually checked in this pass for a McDonald
  mention; a future pass could confirm none of them independently name him either.
- Matt Rickard's June 2023 essay "The Low-Background Steel of AI" (named in Pant's piece but not
  fetched here) is another candidate for an independent, dateable early use of the analogy worth
  checking against Graham-Cumming's March 2023 registration.
- "Tech blog The Luddite" and "Glenn Keighley['s] LinkedIn post," both mentioned by name in Pant's
  piece without links, are further unlinked secondary claims in the same article and share its
  sourcing weakness — not verified here.

## Entity candidates

- John Graham-Cumming — person — already an entity ([[entity-john-graham-cumming]]); the
  foundational figure the McDonald priority claim is measured against — flagged first per the
  known blind spot, even though the page already exists.
- Kyle McDonald — person — the subject of the unverified claim; artist/programmer (@kcimc), NYU ITP
  adjunct, F.A.T. Lab member; a page would need his own words to be worth creating.
- Rajiv Pant — person — sole named source for the McDonald claim; former CTO of The New York Times
  and The Wall Street Journal, writes at rajiv.com; worth flagging since his unlinked paraphrase is
  now load-bearing for a contested priority claim elsewhere in the vault.