Introduction: The Ontological Crisis of AI and the Intelligence Debate
At the crossroads of 2024 and 2025, the AI academic community and industry are embroiled in a profound philosophical and scientific debate that goes beyond technological achievements. At the center lies the question: "Can Large Language Models (LLMs) truly be considered to possess 'intelligence'?"
Frank Landymore's article published in Futurism, "Large Language Models Will Never Be Intelligent," serves as a representative text advocating this skeptical perspective, arguing for the functional separation of language processing ability and general intelligence while pointing out the fundamental limitations of LLMs.
This report takes that article as a starting point for discussion and provides an in-depth analysis of the confrontation between the 'Stochastic Parrot' hypothesis and the 'Emergent Intelligence' hypothesis at the forefront of current AI research.
Key Questions
- ▪ Are LLMs merely statistical imitation machines?
- ▪ Are they a new form of intelligence that has built a World Model through text compression?
- ▪ Can they become a pathway toward AGI (Artificial General Intelligence)?
Part 1: In-Depth Summary and Analysis of the Futurism Article
The Futurism article presents a pessimistic outlook that LLMs cannot achieve human-level intelligence or creativity, citing the views of cognitive science experts and engineers to support this claim. The article's core argument is based on the 'Functional Dissociation' hypothesis, which holds that language ability and thinking ability are fundamentally separate.
1.1 The Separation of Language and Intelligence: Benjamin Riley and Neuroscientific Evidence
The article cites Benjamin Riley's argument that while humans tend to equate linguistic fluency with intelligence, the latest neuroscience research suggests these are separate functions.
In particular, studies published in Nature and other journals in 2023-2024 demonstrated through fMRI scans that brain regions activated during mathematical problem-solving or logical reasoning are clearly distinct from those responsible for language processing.
This is consistent with clinical cases where aphasia patients who have lost their language abilities can still perform complex mathematical problems or play chess. Based on these biological facts, the article argues that LLMs trained only on statistical patterns in language data are not engaging in 'thought' but merely mimicking 'communicative functions.'
1.2 The Limits of Creativity: David Cropley's "Serviceable Artists" Theory
Professor David H. Cropley of the University of South Australia characterizes LLMs as "serviceable artists" and points out their creative limitations.
According to his research, AI is proficient at generating plausible text but cannot reach expert-level originality or a truly creative leap. His conclusion is that LLM creativity is merely an average combination of vast data and, under current design principles, cannot reach professional standards that exceed the human average.
1.3 Yann LeCun's Argument for the Absence of a World Model
The article also prominently features the skepticism of Yann LeCun, Turing Award recipient and Meta's Chief AI Scientist. LeCun argues that text-based autoregressive models, trained only to predict the next word without understanding the physical world, cannot achieve Artificial General Intelligence (AGI).
He contends that LLMs lack a 'world model' that understands the physical laws and causal relationships of the three-dimensional world, and therefore are merely text-processing tools rather than truly intelligent entities.
Part 2: The Opposing Camp (Agreement): The Cognitive Limitations of LLMs
The claims of the Futurism article are strongly supported by modern cognitive science and AI ethics. This section provides detailed arguments for the academic evidence that supports and extends the article's claims, centered on the 'Stochastic Parrot' hypothesis, the 'Symbol Grounding Problem,' and the 'Inverse Scaling' phenomenon.
2.1 Extending the Neuroscientific Evidence: Fedorenko's Language-Thought Dissociation Research
The 2024 Nature paper by MIT neuroscientist Ev Fedorenko and colleagues strongly suggests that language is primarily a tool for communication rather than a tool for thought.
- ▪ Dissociation from the Multiple Demand Network: When the human brain performs complex cognitive tasks (planning, reasoning, problem-solving), it is the 'Multiple Demand Network' that is activated. In contrast, during language processing, an anatomically separate 'Language Network' is activated.
- ▪ Implications for LLMs: From this perspective, current LLMs are equivalent to extracting and massively enlarging only the 'Language Network' from the human brain. Language generation without the mechanisms responsible for reasoning is merely an illusion of intelligence, not the real thing.
2.2 The Stochastic Parrot Hypothesis and the Absence of Meaning
The 'Stochastic Parrots' hypothesis proposed by Emily Bender and Timnit Gebru is the core framework that theoretically supports the tone of the Futurism article.
- ▪ Form vs. Meaning: LLMs learn word co-occurrence patterns within training data. In this process, the model perfectly learns the 'form' of language but cannot access the 'meaning' that the form refers to.
- ▪ The Octopus Thought Experiment: A deep-sea octopus that eavesdrops on the communication cable between two people stranded on a desert island and mimics their conversation may know the statistical usage of the word 'coconut,' but can never know its taste, weight, or reality.
- ▪ The Inevitability of Hallucination: The 'hallucination' phenomenon in which LLMs plausibly fabricate false information is not a flaw of the model but an intrinsic characteristic. This is because the model's objective function optimizes for 'plausibility' rather than 'truth.'
2.3 The Symbol Grounding Problem
The Symbol Grounding Problem formalized by Stevan Harnad asks: "How can symbols within a formal symbol system be connected to meanings in the external world?"
For a text-only LLM, 'apple' is defined solely by its relationship with other word vectors such as 'fruit,' 'red,' and 'delicious.' But since 'fruit' and 'red' are also defined by other words, the model becomes trapped in an endless merry-go-round of symbols.
Symbols that are not sensorily grounded in external physical reality are hollow, and therefore LLMs cannot be said to 'understand' what they are saying.
2.4 Inverse Scaling and the Fragility of Reasoning
LLM advocates claim 'Scaling Laws' that intelligence improves as model size grows, but recent research shows this law does not always hold. The 'Inverse Scaling' phenomenon refers to cases where larger models actually perform worse on certain tasks.
| Phenomenon | Description | Implication |
|---|---|---|
| The Imitation Trap | As models grow larger, they more powerfully imitate human misconceptions and biases contained in the training data | Suggests an increase in 'imitation ability' rather than an increase in intelligence |
| Negation Processing Failure | In questions like "What is not A?", larger models are drawn by strong statistical associations with "A" and produce wrong answers | Proves that statistical association dominates over logical operations |
| Fragility of Reasoning | 'Chain of Thought (CoT)' prompting appears to improve reasoning ability, but actually only mimics the form of reasoning | Merely 'reasoning in appearance' lacking causal connection between reasoning process and correct answers |
This evidence strongly suggests that LLMs are not truly intelligent entities but machines that blindly follow statistical patterns in data.
Part 3: The Supporting Camp (Disagreement): Emergent Intelligence and the Reality of World Models
On the other hand, the claims of the Futurism article conflict with the latest deep learning research findings, particularly Interpretability research that analyzes internal model mechanisms. The supporting camp (those who view LLMs as intelligent) criticizes the article for confusing 'Process' with 'Product' and overlooking the fact that simple prediction tasks, when performed at enormous scale, give rise to qualitatively different 'emergent abilities.'
3.1 Intelligence as Compression: Ilya Sutskever's Counterargument
Former OpenAI Chief Scientist Ilya Sutskever and others argue that when the simple objective of "next word prediction" is performed at sufficiently large data and model scales, it goes beyond mere statistical imitation.
To effectively compress and predict vast amounts of data, the model must internalize the underlying rules of data generation, namely the 'laws of the world.' Therefore, the criticism that it is "merely predicting the next word" underestimates the cognitive depth required to perform that prediction perfectly.
3.2 Othello-GPT: Empirical Evidence for Internal World Models
The research that directly refutes LeCun's claim in the Futurism article that "LLMs have no world model" is the Othello-GPT study.
- ▪ Experiment Overview: Researchers trained the LLM exclusively on game transcript text (e.g., "E3, D4,...") without showing it any rules of Othello or board images.
- ▪ Discovery: When the trained model's internals were analyzed with probes, the model had spontaneously constructed high-dimensional geometric representations of the 64-square Othello board state and the color (black/white) of each piece.
- ▪ Causal Intervention: When researchers artificially manipulated specific neuron values inside the model, the model's next-move predictions changed rationally to match the manipulated state. This proves that the model was not simply memorizing text patterns but performing causal reasoning based on an internally constructed 'world model (board state).'
If the spatial and logical rules of the game of Othello can be reconstructed from simple text transcript learning alone, it is highly likely that a large model trained on the entire text of the internet has extracted and internalized basic 'world models' of grammar, logic, social relationships, and physics from text.
3.3 Emergent Abilities and Phase Transitions
LLM advocates focus on 'emergent abilities' that occur when model size crosses a critical threshold. According to research by Wei et al. (2022), abilities such as arithmetic operations, multi-step reasoning, and code debugging do not appear at all in small models, but exhibit a 'Phase Transition' where performance sharply improves once a certain scale of computation is exceeded.
- ▪ The Grokking Phenomenon: Recent studies report a 'grokking' phenomenon where models initially simply memorize data, but after prolonged training, they discover the general rules of the data and achieve generalization. This is powerful evidence that LLMs can progress from the 'stochastic parrot' stage to the 'algorithmic understanding' stage.
- ▪ Sparks of AGI: Microsoft Research's "Sparks of AGI" paper demonstrated that early GPT-4 could perform novel tasks not explicitly present in its training data, interpreting this as an early form of general intelligence.
3.4 Creativity Benchmarks: Surpassing Humans
Contrary to Professor Cropley's dismissal of LLMs as "mediocre," objective creativity benchmark results tell a different story. In Guzik et al.'s 2023 study, GPT-4 took the standardized creativity test known as the 'Torrance Tests of Creative Thinking (TTCT)'.
| Evaluation Criterion | GPT-4 Achievement | Significance |
|---|---|---|
| Originality | Top 1% | Generated more unique and rare ideas than 99% of human participants |
| Fluency | Top 1% | Produced an overwhelmingly large number of ideas within the given time |
| Flexibility | Top tier | Demonstrated the ability to shift thinking across diverse categories |
These results suggest that LLMs do not simply regress to the average of their training data but can explore distant regions of Latent Space to create novel combinations that are difficult for humans to conceive.
3.5 Resolving Symbol Grounding Through Multimodality
While the article pointed out the limitations of text-only models, models as of 2025 have evolved into multimodal models that simultaneously process text, images, and audio.
Models like GPT-4V and Gemini are technically circumventing the Symbol Grounding Problem raised by Harnad by mapping the word 'apple' to visual images. As text symbols are grounded in physical features (color, shape) through visual information, LLMs are evolving from closed symbolic systems into open systems connected to the external world.
Part 4: Comprehensive Critique and Future Outlook
When we synthesize the Futurism article and the arguments for and against it, we can see that the current AI debate is a collision between 'Functionalism' and 'Essentialism'.
The article criticizes LLMs by using human biological mechanisms (essence) as the standard for intelligence, while the opposing camp defines intelligence based on the usefulness and complexity (function) of the output.
4.1 Reassessing the Article's Claims
- ▪ The Relationship Between Language and Thought: The fact that "humans process language and thought separately" does not lead to the proposition that "AI must also do so to be intelligent." Just as an airplane flies without flapping its wings like a bird, silicon-based intelligence may have acquired reasoning abilities through a different pathway of language modeling (Substrate Independence).
- ▪ Limitations as a Tool: The criticism that it is "merely a communication tool" becomes blurred when that tool becomes sophisticated enough to understand user intent, maintain complex contexts, and propose creative solutions. The Othello-GPT case demonstrated that simple prediction tasks internally require sophisticated cognitive modeling.
4.2 New Horizons in AI: Hybrid Architectures
The two extremes of the debate are converging with technological advancement. While acknowledging the limitations of pure LLMs (lack of planning ability, hallucination), new architectures are emerging that leverage their powerful associative abilities and knowledge base.
- 1. System 2 Reasoning: Technologies that mimic the human slow, logical thinking (System 2) are being introduced, enabling LLMs to internally generate and verify a 'chain of thought' before producing an immediate response (e.g., OpenAI o1, Strawberry).
- 2. Neuro-Symbolic AI: By combining LLM language capabilities with traditional symbolic AI (logic, mathematics, databases), the field is moving toward simultaneously pursuing fluency and accuracy.
- 3. Integration of JEPA and World Models: The JEPA architecture proposed by LeCun is also more likely to be integrated in a form that supplements the LLM's lacking physical common sense and planning abilities, rather than completely replacing LLMs.
4.3 Conclusion: The Birth of the 'Understanding' Parrot
The Futurism article "Large Language Models Will Never Be Intelligent" sharply pointed out the fundamental constraints of current LLMs: the absence of embodied experience, statistical dependence, and structural differences from the biological brain. This criticism is very useful in guarding against excessive AI hype and facing the essence of the technology squarely.
However, the definitive conclusion that they "Will Never" be intelligent appears premature. The emergence of internal world models confirmed in Othello-GPT, the creativity proven in the Torrance Tests, and the progress in grounding through multimodal learning all show that LLMs are moving beyond being simple 'stochastic parrots.'
We are now witnessing a strange form of intelligence (Alien Intelligence) that has evolved through an entirely different pathway from humans. It is not a being that senses and feels like a human, but it is evolving into a 'Reasonable Parrot' that has constructed its own 'world' and 'meaning' by compressing and structuring the vast ocean of symbols that is text.
Pebblous Perspective: The Difference Data Quality Makes in Intelligence
Why Does Pebblous Focus on the LLM Intelligence Debate?
Pebblous views this debate as going beyond mere philosophical curiosity to constitute a practical engineering problem. The key factor driving LLMs from 'stochastic parrot' to 'emergent intelligence' is the quality of training data.
The reason Othello-GPT was able to build an internal world model is that it learned from high-quality, structured data in the form of game transcripts. Conversely, the inverse scaling phenomenon shows that biased and noisy data actually impedes a model's reasoning ability.
DataClinic and AADS: Data Strategy for the AGI Era
-
●
DataClinic: For LLMs to build genuine world models, data similarity, representativeness, and diversity must be ensured. DataClinic diagnoses and improves the quality of AI training data in accordance with the ISO/IEC 5259-2 standard, providing the data foundation for advancing from 'Reasonable Parrot' to 'Rational Intelligence.'
-
●
AADS (Autonomous AI Data Scientist): An autonomous agent that applies LLM reasoning capabilities to real business problem-solving. AADS goes beyond simply imitating data patterns to discover causal relationships within data, design experiments, and verify hypotheses. This is a case demonstrating that LLMs can evolve from 'chains of thought' to 'laboratories of thought.'
Pebblous Vision
If LLMs are on the path to AGI, what paves that path is high-quality data. Pebblous applies the latest findings from neuroscience, cognitive psychology, and mechanistic interpretability research to data science, helping AI evolve into a partner that goes beyond merely speaking to thinking, understanding, and creating.
Frequently Asked Questions (FAQ)
Q1. Are LLMs (Large Language Models) truly intelligent, or are they simply imitating patterns?
This question is the hottest debate in the current AI academic community. The 'Stochastic Parrot' hypothesis claims that LLMs merely imitate statistical patterns in training data and lack genuine understanding or reasoning ability. In contrast, the 'Emergent Intelligence' hypothesis, citing research like Othello-GPT, posits that sufficiently large models internally construct world models during the text compression process and acquire qualitatively different abilities. Pebblous believes this debate can shift depending on data quality -- high-quality, structured data produces emergent intelligence, while noisy data produces stochastic parrots.
Q2. What did the Othello-GPT experiment prove?
The Othello-GPT experiment trained an LLM exclusively on game transcript text without showing it any rules of Othello or board images. Remarkably, analysis of the model's internals revealed that it had spontaneously constructed a world model representing the spatial layout of the 64-square board and the color of each piece. Even more importantly, when researchers manipulated specific neurons inside the model (e.g., changing the color of a piece on a specific square), the model's predictions changed rationally to match. This suggests that LLMs can possess internal models capable of causal reasoning beyond simple pattern memorization.
Q3. What does the 'Inverse Scaling' phenomenon mean?
While it is generally known that performance improves as model size increases, an 'Inverse Scaling' phenomenon has been discovered where larger models actually perform worse on certain tasks. For example, larger models more strongly learn biases and misconceptions from training data, giving incorrect answers to logical negation questions or failing at reasoning due to over-reliance on statistical associations. This shows that simply making models larger cannot achieve intelligence, and data quality management is essential.
Q4. What does neuroscience research say about LLM intelligence?
MIT's Fedorenko research team discovered through fMRI scans that the 'Language Network' responsible for language processing and the 'Multiple Demand Network' responsible for reasoning and planning are anatomically separated in the human brain. This suggests that language ability and thinking ability are separate functions, providing grounds that linguistic fluency alone cannot be used to judge intelligence. However, some AI researchers counter with the principle of 'Substrate Independence,' arguing that the human brain and silicon-based AI can produce similar results through different structures -- just as an airplane flies differently from a bird.
Q5. Can LLMs become a pathway to AGI (Artificial General Intelligence)?
Experts are sharply divided on this question. Skeptics like Yann LeCun argue that text alone cannot achieve genuine understanding of the physical world, making LLMs a dead end on the road to AGI. On the other hand, Ilya Sutskever and others believe that the goal of 'next word prediction,' when performed at sufficiently large scale, leads to internalizing the laws of the world, and that symbol grounding problems can be solved through multimodal (visual, auditory) expansion. Indeed, GPT-4's 'Sparks of AGI' paper reported signs of early general intelligence. Pebblous views a hybrid architecture -- combining LLM language capabilities with neuro-symbolic AI logical reasoning -- as the most realistic pathway.
Q6. How are Pebblous's DataClinic and AADS related to this debate?
The key factor driving LLMs from stochastic parrot to emergent intelligence is the quality of training data. Pebblous's DataClinic quantitatively measures and improves AI data similarity, representativeness, and diversity in accordance with the ISO/IEC 5259-2 standard, helping LLMs learn from unbiased, high-quality data to build genuine world models. AADS (Autonomous AI Data Scientist) is an autonomous agent that applies LLM reasoning capabilities to real business problems, discovering causal relationships in data and verifying hypotheses. This is a case demonstrating that LLMs are capable of scientific thinking beyond simple pattern imitation. Pebblous believes that high-quality data paves the road to AGI.
References
Key Papers and Articles
- 1. Landymore, F. (2025). "Large Language Models Will Never Be Intelligent", Futurism. Link
- 2. Fedorenko, E. et al. (2024). "Language is primarily a tool for communication rather than thought", Nature. Link
- 3. Bender, E. & Gebru, T. et al. (2021). "On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?", FAccT 2021. Link
- 4. Li, K. et al. (2023). "Emergent world representations: Exploring a sequence model trained on a synthetic task" (Othello-GPT), arXiv. Link
- 5. Wei, J. et al. (2022). "Emergent Abilities of Large Language Models", CSET Georgetown. Link
- 6. Guzik, E. et al. (2023). "The Originality of Machines: AI Takes the Torrance Test", Journal of Creativity. Link
- 7. Bubeck, S. et al. (2023). "Sparks of Artificial General Intelligence: Early experiments with GPT-4", Microsoft Research. Link
- 8. Harnad, S. (1990). "The Symbol Grounding Problem", Physica D. Commentary
- 9. LeCun, Y. (2024). "World Models vs. Word Models: Why LLMs Will Be Obsolete", Medium. Link
- 10. McKenzie, I. et al. (2023). "Inverse Scaling: When Bigger Isn't Better", arXiv. Link
Additional Resources
- ▪ Scaling Laws for Neural Language Models (OpenAI)
- ▪ The Vector Grounding Problem (arXiv)
- ▪ Chain of Thought Prompting Elicits Reasoning in Large Language Models (Google Research)
- ▪ Multimodal Grounding in Large Language Models (UCSD)
- ▪ Critical Review of LeCun's JEPA Paper (Malcolm Lett, Medium)