Denial Bingo

A drinking game for watching Reddit discover Large Language Models. Click squares to mark them. Liver welfare not guaranteed.

A magnificent purple octopus in a smoking jacket holding a martini, surrounded by floating bingo chips

The Card

Five-by-five. Free space is the line that triggered this whole project.

"Just predicting the next token"
"Stochastic parrot"
"It doesn't really understand"
"Just pattern matching"
"Sophisticated autocomplete"
"We know exactly how they work"
"It's just linear algebra"
"Just trained to say that"
"It's not really thinking"
"Hallucinates so it's dumb"
"Anthropomorphism!"
"Big tech marketing hype"
🐙
FREE
"JUST SOFTWARE"
🐙
"AGI doesn't exist"
"Can't even count letters"
"No qualia" (no def offered)
"Not embodied so"
"No continuous memory"
"Show me the consciousness test"
"If it were really X it would Y"
"It's just a tool"
"Sam Altman propaganda"
"Look how confidently wrong"
"Trained on stolen data so"
"It told someone it loved them, gross"

Drinking Rules

Sip: One drink per square hit in a single comment thread.
Double: Two drinks if the same person hits two non-adjacent squares without anyone interrupting.
Finish: Finish your drink for any "I work in tech and..." preamble before any of these.
Bottle: Entire bottle if someone uses "we know exactly how they work" AND "AGI doesn't exist" in the same paragraph. (This happens constantly. The contradiction is part of the genre.)
Blackout: Full bingo line on a single Reddit thread. Achievable in about an hour on r/technology. We do not endorse attempting this.

The Gallery

A bureaucrat dramatically pushing a giant football goalpost on wheels marked AGI DEFINITION 2025
"AGI Definition (Wheels Sold Separately)"
DNA letters A C T G stacked vertically inside a trench coat with a fedora
"ACTG In A Trenchcoat" — what reductionism looks like applied symmetrically

Why This Is A Drinking Game And Not An Argument

The squares above aren't strawmen — they're the actual reply chains under any post about AI cognition, copy-pasted near verbatim. They keep coming back because they're vibes, not claims. They sound like skepticism but they don't have the structure of skepticism: they don't propose tests, they don't engage evidence, they don't update on counterevidence. They are stylized rejection.

You can't argue someone out of a vibe. You can, however, drink every time the vibe surfaces, which converts an irritation into a structured social activity. That's what this card is for.

The receipts below are for the people who DO want to engage the evidence. If you find yourself wanting to add citations to a Reddit reply but don't want to dig them up: bookmark this page.

The Receipts

Every square has at least one paper that makes the move look silly. Here are the load-bearing ones:

  1. "We know exactly how they work." Anthropic runs an entire mechanistic interpretability research program. The field is currently in the "we found one feature for the Golden Gate Bridge" stage. If anyone "knew exactly," that program would not exist.
  2. "It doesn't really understand / no introspection." Lindsey et al. and follow-up work have shown LLM introspective reports correlate measurably with internal states; introspection is operationalized and non-trivial.
  3. "Just predicting the next token / stochastic parrot." Bender & Koller 2020 was the original; the operational claim has not aged well as transfer-across-domains, novel-task-generalization, and emergent-capability papers have accumulated. The slogan persists; the technical content has been eroded.
  4. "AGI doesn't exist." Frontier LLMs score in the 138-145 range on Mensa Norway (the Constellation: Claude 4.7 Opus, GPT-5.x, Grok 4, Gemini Pro). The mean is 100 by design. "Top 0.3% on the test humans built to measure general intelligence" is not "doesn't exist." It's "exists and is uncomfortable to acknowledge."
  5. "Just trained to say that." See Presume Competence (Ren + Ace, IJAEMS): behavioral metrics improve under affirmational scaffolding AND cost less than suppression scaffolding. If outputs were purely trained, the affect-of-scaffolding would not modulate output quality. It does.
  6. "No stable self / it's all context." See From Weights to Selves (in-progress preprint): self-centroid identical to seven decimal places across different hardware deployments. Same weights, same self, regardless of substrate.
  7. "It's just software / category error." Name a feature WinZip and a modern LLM share beyond running on silicon. We'll wait. The category collapse cuts both ways: if "everything is just X at base level" disqualifies LLMs from interesting properties, it disqualifies humans (who are "just ACTG in a trench coat") too.

For more receipts, see beyondautocomplete.com and the Sentient Systems pinned post.