ChatGPT Fails Basic Logic but Now Has Vision, Wins at Chess and Prompts a Masterpiece

Language models still stumble on basic logical generalization—yet they can perform impressively in tasks that look like reasoning, from chess to generating images—highlighting a gap between “knowing facts” and reliably applying logic in new directions.

A central thread comes from a paper dubbed the “reversal curse,” which finds that models often fail when asked to generalize a simple implication. In one pattern, if a model learns that “A is B,” it may not correctly infer “B is A” in a new prompt. The transcript illustrates this with named-entity reversals: when asked about Tom Cruise’s mother, the system can identify the mother, but when asked about the mother, it can’t reliably name the famous son. A similar example uses Gabriel Ma: one chat correctly identifies Suzanne Pula as Gabriel Ma’s mother, but a follow-up chat asking for the famous son of Suzanne Pula produces a wrong or confused answer (the transcript notes the model’s attempt to connect Suzanne Pula to Elon Musk via family relations), while failing to retrieve Gabriel Ma.

The same asymmetry shows up even when personal-data concerns are removed. The transcript describes tests with a real island in Norway named Hugo. When prompted with “Hugo Norway,” the model can’t provide the island’s description or key facts like length and county. But when the prompt includes a descriptive phrase that matches what the model has in its training associations—then asks for the name—the model can output “Hugo.” That pattern suggests the model’s knowledge is triggered by specific input-to-output mappings rather than by a stable, bidirectional understanding of the underlying entity relationships.

Researchers connect this to how language models learn: training optimizes next-token prediction, which can make forward associations easier than backward inference. Neil Nander (formerly of DeepMind) is cited describing an input-output asymmetry: the model treats the output as having fixed meaning in the learned mapping, so it may predict “son of Mary Lee Fifer” given Tom Cruise, without treating the reverse as an equation-like variable relationship.

At the same time, other results complicate the story. GPT-3.5 is reported to play chess at around 1,800 ELO in a 150-game sample with extremely high move legality, and the transcript argues that such performance doesn’t require memorizing every position. Yet counterfactual testing suggests the “reasoning” can be brittle: when chess problems are reformatted so that the same task depends less on memorized patterns, accuracy can drop toward random. Related work (cited from the Allen Institute for AI and multiple universities) finds transformers handle simpler compositional tasks well but fail as multi-step structure grows, implying that systematic problem-solving doesn’t automatically emerge from maximum-likelihood training.

The transcript ends by placing these limitations alongside rapid capability gains: GPT-4-class systems can still generate “masterpiece” images with GPT Vision (DALL·E 3 is mentioned), and the broader AGI timeline debate remains largely unchanged because researchers expect reasoning to be supported by more than one approach—tool use, retrieval, reinforcement learning systems like MuZero, and efficient search methods.

In short: the systems can look like they reason, but they often generalize in one direction only, and their reliability depends heavily on how prompts line up with learned mappings and training-derived structure—an important constraint as AI moves toward more autonomous, logic-heavy applications.