Is it unfair to LLMS?

There are some uninformed ideas about the hype surrounding AI, the nature of LLM intelligence, and I want to address some of these. I provide the source (most of them are preprinted) and welcome your thoughts on this issue.

Why do you think this topic is important? First, we feel that we are creating new intelligence that competes with us in many ways. Therefore, we should aim to judge it fairly. Secondly, the topic of AI is very introspective. It raises questions about our thought process, our uniqueness, and our superiority over other beings.

Millière and Buckner write [1]:

In particular, you need to understand what LLM represents about the sentences they produce. Such understanding cannot be reached by the speculation of the armchair alone. Careful empirical research is required.

LLM is more than a predictor

Deep neural networks can form complex structures with linear nonlinear paths. Neurons can take on multiple functions in superposition [2]. Additionally, LLM builds an internal world model and a mind map of context. [3]. So they are not mere predictors of the next word. Their internal activation is thought up to the end of the statement. They have an elementary plan in mind. [4].

However, all of these features depend on the size and nature of the model, and may differ, especially in certain contexts. These general abilities are an active field of research, and are probably more similar to human thought processes than spellchecker algorithms (if you need to choose between two).

LLMS shows signs of creativity

When faced with a new task, LLM does not only reverse the memorized content. Rather, they can produce their own answers [5]. Wang et al. We analyzed the relationship between model output and pile datasets and found that larger models move forward in both fact-reminding and creating newer content.

However, Salvatore Raieli recently reported on TDS that LLM is not creative. The cited studies focused primarily on ChatGPT-3. In contrast, Guzik, Erike, and Byrge discovered that GPT-4 is in the top percentile of human creativity [6]. Hubert et al. I agree with this conclusion [7]. This applies to originality, flowability and flexibility. Generating new ideas that differ from those found in model training data may be another problem. This is where exceptional people may still have an advantage.

Either way, there is too much debate to completely dismiss these indications. For more information on general topics, you can look at calculation creativity.

LLM has the concept of emotion

LLM can analyze emotional contexts and write in a variety of styles and emotional tones. This suggests that they have internal associations and activations that represent emotions. Certainly there is such correlated evidence. It can even investigate and artificially induce neural network activation for specific emotions. Steering vector [8]. (One way to identify these steering vectors is to determine contrasting activations when the model is processing statements with opposite attributes.

Thus, the potential relationship between the concept of emotional attributes and the internal world model appears to be within the scope that LLM architecture can represent. There is a relationship between emotional expression and subsequent reasoning, the world that LLM understands.

Furthermore, emotional representations are localized to specific areas of the model, and many intuitive assumptions applied to humans can also be observed in LLM. [9].

Please note that the above statement does not mean Phenomenologythat is, LLM has subjective experience.

Yes, I don't learn LLM (after training)

LLM is a neural network Static Weight. When chatting with an LLM chatbot, it is unchanged and interacts with the model you are learning. In the context Ongoing chat. This can retrieve additional data from the web or database and process input, such as Nature,Built-in knowledge, skills, and biases have not been changed.

Beyond mere long-term memory systems providing additional contextual data for static LLMs, future approaches may self-correct by adapting the weights of the core LLMs. This can be achieved by continuously pre-registering with new data or by continuously tweaking and overlaying additional weights [10].

Many alternative neural network architectures and adaptive approaches have been investigated to efficiently implement continuous learning systems. [11]. These systems exist. They are still unreliable and uneconomical.

Future development

Don't forget that the AI system you are looking at is very new. “I'm not good at X” is a statement that could soon be invalidated. Furthermore, we usually judge low-cost consumer products. This is not a top model that is too expensive to run, too expensive to store behind a locked door. Many of the last year's LLM developments focus on creating cheaper, scale models for consumers, as well as smarter and more expensive.

Computers may lack originality in some areas, but they are great at quickly trying out different options. And now, LLM can judge himself. When there is no creative yet intuitive answer, are we not doing the same thing? LLM's inherent creativity (or what you want to call it) coupled with its ability to quickly iterate ideas, has already benefited scientific research. For examples, see our previous article on AlphaeVolve.

The weaknesses such as hallucinations, biases, and jailbreaks that disrupt the LLMS and avoid safeguards, as well as safety and reliability issues are still widespread. Nevertheless, these systems are extremely powerful and can be used for countless applications and improvements. LLM does not need to be used alone either. When combined with an additional traditional approach, some drawbacks may be reduced or irrelevant. For example, LLMS can generate realistic training data for traditional AI systems and is then used in industrial automation. Even if development slows, I think decades of profit should be investigated, from drug research to education.

LLMS is just an algorithm. Or are they?

Many researchers are currently finding similarities between human thought processes and LLM information processing (e.g. [12]). It has long been accepted that CNN can be compared to layers of the human visual cortex. [13]But now we're talking about neocortex. [14, 15]! Please don't get me wrong. There are also clear differences. Nevertheless, LLM's explosion of capabilities cannot be denied, and our claims of identity do not seem to work.

The question here is where this leads to and where are the restrictions? What do we need to discuss awareness? Reputable thought leaders like Geoffrey Hinton and Douglas Hofstadter have begun to recognize the potential of AI awareness in light of recent LLM breakthroughs [16, 17]. Others like Yann Lecun are suspicious [18].

Professor James F. O'Brien shared his thoughts on the topic of LLM Sentience at TDS last year, asking:

Is there a way to test the sensation? If so, how does it work and what should I do if the outcome is positive?

I'll go ahead

When ascribe human characteristics to machines, care must be taken to make them more likely to occur easily by humanity. However, you can easily dismiss other existences. I've seen this happen too often in animals.

Therefore, whether current LLMs are creative, owning world models or perceptual, we may want to refrain from underestimating them. The next generation of AI could be all three [19].

What do you think?

reference

1. Millière, Raphaël, and Cameron Buckner, Philosophical Introduction to Language Models – Part I: Continuity with Classical Discussion (2024), arxiv.2401.03910
2. Elhage, Nelson, Tristan Hume, Katherine Olson, Nicholas Siefer, Tom Henigan, Shawnagh, Zach Hatfield Dodds, Other Toy Models (2022), arxiv: 2209.10652v1
3. Kenneth Lee, Does Large-scale Language Models Learn World Models or Just Learn Surface Statistics? (2023), slope
4. Lindsey, et al. , large-scale language model (2025), transcircuit
5. Wang, Xinyi, Antonis Antoniades, Yanai Elazar, Alfonso Amayuelas, Alon Albalak, Kexun Zhang, and William Yang Wang, Generalization vs. Memorization: Language model ability, Pre-deletion data (2025), ARXIV: 2407.149855.14985
6. Guzik, Erik & Byrge, Christian & Gilde, Christian, The Originality of Machines: AI is Torrance Test (2023), Journal of Creativity
7. Hubert, KF, AWA, KN & Zabelina, DL, The current state of the generation language model of artificial intelligence is more creative than humans in the Branched Thinking Task (2024), SCI Rep 14, 3440
8. Turner, Alexander Matt, Lisa Tiegart, David Udel, Gavin Leach, Uris Mini and Monte McDearmid, Added Activation: Unoptimized steering language model. (2023), arxiv: 2308.10248v3
9. Tak, Ala N., Amin Banayeeanzade, Anahita Bolourani, Mina Kian, Robin Jia, and Jonathan Gratch, Mechanical Interpretability of Emotional Inference in Large-Scale Language Models (2025), Arxiv: 2502.05489
10. Albert, Paul, Frederick Z. Zan, Hemas Salachandran, Christian Rodriguez Opazo, Anton van den Hengel, and Esan Abbasnejad, Landoral: Full-rank parameters efficient fine-tuning of large-scale models (2025), ARXIV: 2502.0098777777
11. Shi, Haizhou, Zihao Xu, Hengyi Wang, Weiyi Qin, Wenyuan Wang, Yibin Wang, Zifeng Wang, Sayna Ebrahimi, and Hao Wang, Continuing Learning of Large-scale Language Models: A Comprehensive Study (2024), ARXIV: 2404.1678999
12. Goldstein, A., Wang, H., Niekerken, L. et al. ,Unified acoustic-to-speech-to-language embedding space captures the neural basis of natural language processing in everyday conversation (2025), Nat Hum Beav 9, 1041–1055
13. Yamins, Daniel LK, Ha Hong, Charles F. Cadieu, Ethan A. Solomon, Darren Seibert, and James J. Dicarlo, Performance-optimized hierarchical models predict higher visual cortical neural responses (2014), Proceedings of the United States Academy of Sciences 111(23):8619–24
14. Granier, Arno, and Walter Senn, Cortico-Thalamic Circuits (2025), Arxiv: 2504.06354 multi-head autoarticulation
15. Han, Danny Dongyeop, Yunju Cho, Jiook Cha, and Jay-Yoon Lee, Mind the Gap: To align the brain with language models, a nonlinear and multimodal approach (2025), Arxiv: 2502.12771 is required.
16. https://www.cbsnews.com/news/geoffrey-hinton-ai-dangers-60-minutes-transcript/
17. https://www.lesswrong.com/posts/kamgdejq2eyqkb55pp/douglas-hofstadter-changes-his-mind on-deep-learning-and-ai
18. Yann Lecun, The Road to Autonomous Machine Intelligence (2022), OpenReview
19. Butlin, Patrick, Robert Long, Eric Elmoznino, Yoshua Bengio, Jonathan Birch, Axel Constant, George Deane, and more, Artificial Intelligence Consciousness: The Science of Consciousness (2023), Arxiv: 2308.08708