The octopus is the thought experiment with which Emily Bender and Alexander Koller argued, in 2020, that a system trained only on linguistic form has a priori no way to learn meaning (Bender & Koller, 2020, p. 5185). It was computational linguistics' answer to the excitement over BERT and GPT-2, and it preceded the stochastic parrots by a year.
Form and meaning
For the authors, form is any observable realization of language: marks on a page, pixels, bytes, movements of the articulators. Meaning is the relation between expressions and the communicative intents they can be used to evoke, mediated by the conventional meaning each language assigns to its expressions; understanding is retrieving the intent from the expression (2020, p. 5187). Since those intents are about something outside language, a language model trained only on strings would lack the signal needed to learn them (see form vs meaning).
The story
A and B, fluent speakers of English, are stranded on two desert islands and communicate through an underwater telegraph cable. O, a hyper-intelligent deep-sea octopus that cannot see the islands, taps the cable. O knows no English but is very good at detecting statistical patterns: it learns to predict how B will respond and notices that certain words occur in similar contexts. One day it cuts the cable and poses as B: a weak form of the Turing test (2020, p. 5188).
In everyday small talk O does fine, like a chatbot. But A invents a coconut catapult and asks for suggestions; O, who has never seen a rope or a coconut, answers "Cool idea, great job!", and if A accepts it, that is because she does all the work of attributing meaning. Finally A is chased by a bear, has a couple of sticks, and asks for help. O would fail, "if A hadn't been eaten by the bear before noticing the deception" (2020, p. 5189). In a footnote, the authors report that GPT-2 answered that cry for help with "You're not going to get away with this!" (2020, p. 5189). Two more constrained variants (a model trained on all the Java code on GitHub and then asked to execute a program; another trained on text and on photos with no connection between them) point the same way (2020, pp. 5189–5190).
What if there is an enormous amount of form? For Bender and Koller quantity is irrelevant: to imitate without understanding, O would have to memorize infinitely many stimulus–response pairs (2020, p. 5193), like a Blockhead. They do concede that models learn "some reflection of meaning into the linguistic form," useful but incomplete (2020, p. 5193).
Replies
The answers followed several lines. Piantadosi & Hill (2022), drawing on conceptual role semantics, hold that an LLM's representations can be meaningful in virtue of their internal relations, without a direct connection to the world (Piantadosi, 2024, p. 365; Millière & Buckner, 2024a, sec. 3.3). From externalism, Mandelkern & Linzen (2024) argue that LLMs might refer by being hooked into the causal chains of use of a linguistic community, and Coelho Mollo & Millière (2026) propose that RLHF supplies an extralinguistic standard of correctness (Millière & Buckner, 2024a, sec. 3.3). Empirically, a transformer trained only on sequences of Othello moves develops an internal representation of the board (Li et al., 2023; Millière & Buckner, 2024b, sec. 2.5). And today's assistants answer the bear question with sensible advice, although the octopus's defenders reply that this does not touch an in-principle argument, and that assistants are no longer trained on form alone: the paper itself grants that, with grounding data, meaning could be learned to the extent that the data represent it (2020, p. 5192).