What can LLMs never do?
Claims that large language models will soon reach human-level reasoning are challenged by examples where they reliably fail at tasks that are easy for people, such as Wordle, cellular automata, strict formatting rules, and certain logical puzzles. Commenters debate whether these failures reveal fundamental architectural limits of transformer-based LLMs (fixed-depth reasoning, weak working memory, tokenization artifacts, lack of metacognition) or are merely temporary shortcomings that can be overcome with better training, external tools, or hybrid systems. More broadly, the thread questions whether scaling text-only models can ever yield general intelligence, or whether true AGI will require additional cognitive modules, richer interaction with the environment, and explicit mechanisms for planning and learning over time.
Scope of “Never”: LLMs vs. AI in General
- Several commenters argue the title should really be “can’t do yet”; others insist some limits are intrinsic to language models, not to AI overall.
- Distinction drawn between “LLM” (a specific architecture trained on text) and “AGI” (a broader cognitive system with many components).
Limits of Current LLM Capabilities
- Repeated examples: poor performance on Wordle, cellular automata (e.g., Game of Life, Rule 110), Sudoku-like tasks, simple arithmetic on long lists, counting, palindromes, and tightly specified formatting rules (e.g., number-word style, no “interesting” preface).
- LLMs struggle with reversibility (e.g., “A is B” → “B is A”), systematic logical deduction, and stable adherence to instructions across runs.
- Many note failures at “meta” or structural tasks: ASCII art, reasoning about grids/2D layouts, letter positions inside words, and consistency of style constraints.
Architecture, Expressivity, and AGI
- Strong view: scaling transformers and data will not magically produce AGI; a full cognitive architecture with memory, recursion, planning, and metacognition is needed.
- Others counter that we don’t yet know the capability limits of attention-based models; theoretical work on transformer expressivity and length generalization is cited.
- Debate over whether universal approximation / Turing-completeness arguments meaningfully answer “never.”
Tokenization and Representation Issues
- Many failures are attributed to subword tokenization: models “see” tokens, not characters or 2D space, so Wordle, spelling, grids, and numerical formats are hard.
- Some argue “almost every modern problem is tokenization”; others think this is only part of deeper reasoning limits.
Prompting, Agents, and External Tools
- Prompt engineering, chain-of-thought, retrieval-augmented generation, external memory, code execution, and agents can often patch weaknesses, but raise “who is really solving the problem?”
- Agentic setups (loops, tools, RL fine-tuning) may extend capabilities but also risk goal drift and uncontrolled behavior.
Safety, Use, and Misuse
- Concern about deploying LLMs for high-stakes decisions in public sector, medicine, or critical software, given unreliability and hallucinations.
- Many see them as powerful assistive tools (coding, writing, search) but not trustworthy autonomous decision-makers.
Understanding, Reasoning, and Consciousness
- Ongoing argument: are LLMs merely “stochastic parrots” or do they genuinely reason?
- Lack of clear definitions of “understanding,” “reasoning,” and “AGI” makes the debate partly philosophical; several commenters highlight this ambiguity explicitly.