Blockhead is the machine Ned Block imagined to show that behavior alone is not enough to attribute intelligence. It would pass any Turing test and yet, according to Block, it would have "the intelligence of a toaster" (Block, 1981, p. 21).
The machine
Block's target is a refined behaviorism, which he calls the "neo-Turing Test conception of intelligence": conversational intelligence would be the capacity to produce a sensible sequence of verbal responses to any sequence of verbal stimuli (1981, p. 18). To refute it, he considers all the strings of sentences a person can type in an hour or less. Since the vocabulary is finite, the set is huge but finite. From it we select the "sensible" strings, that is, conversations in which at least one party's contributions are sensible. A large and clever team with plenty of funding could list them; to make them convincing, it might imitate a definite personality, Block's own Aunt Bertha (1981, pp. 19–20).
The machine is trivial: when the interrogator types A, it looks up the strings beginning with A and outputs the second sentence of one of them; when C arrives, it looks up those beginning with A, B and C, and so on. It amounts to searching a tree (1981, p. 20). All the intelligence it exhibits, Block concludes, is that of its programmers (1981, p. 21). Hence his thesis, psychologism: whether behavior is intelligent depends on the character of the internal information processing that produces it (1981, p. 5). The nickname "Blockhead" was coined by later literature.
What if it does not fit in the universe?
Block raises the objection himself. Following an estimate by George Miller, he calculates that an hour-long test would require on the order of 10¹⁵⁰⁰ strings, more than the number of particles in the universe. His reply: the argument only needs the machine to be logically possible, because behaviorism was offered as a conceptual analysis (1981, p. 30). Nor is there chauvinism: psychologism does not require every intelligence to process information as we do; it only rules out certain kinds of processing (1981, p. 21).
Blockhead and LLMs
For Millière & Buckner (2024a), GPT-4 can look like a thought experiment come to life: if all LLMs did were retrieve memorized answers, they would be "Blockheads come to life" (2024a, sec. 1). The suspicion is not idle. Deep networks can memorize randomly assigned labels, models sometimes reproduce copyrighted text verbatim, and data contamination means that a good benchmark score may come from having seen the questions (Millière & Buckner, 2024a, sec. 1; Sainz et al., 2023). Millière and Buckner propose taking Blockhead as a null hypothesis and asking what evidence refutes it.
There is a disanalogy of scale. A model like GPT-3 has about 175 billion parameters (Buckner, 2024, p. 59): some 1,500 orders of magnitude fewer than Block's strings. It cannot be a table; it has to compress, and compressing well is generalizing. Interpretability offers more direct evidence: in grokking, a network goes from memorizing input–output pairs to implementing a general algorithm for modular addition, which according to Millière & Buckner (2024b) speaks against the picture of models as giant lookup tables (2024b, sec. 2.5). Chollet (2019), for his part, reminds us that skill can be "bought" with more data or more prior knowledge without gaining any generalization at all (2019, sec. II.1.1).