Some time ago a former work colleague contacted me via email. In my answer to him, I started a small dispute about today’s so called AI tools. I told him that while I am a strong supporter of Machine Learning and related research regarding new neural network architectures and applications, I am very skeptical about the role of LLMs in the alleged approach of US Tech giants towards an AGI.
My personal experiences with the present AI-tools (like ChatGPT-5) regarding the solution of concrete problems in math and physics have been rather bad. I admit that negative impressions and the tiresome necessity to always control and correct the tools’ ways of reasoning with extensive prompting and to check generated Python programs thoroughly, may have made me prejudiced against the latest generation of tools. (I have not tested Gemini.)
As I have pointed out in a previous article in this blog the point that add-ons checking chains of logical “reasoning steps” can in principle not resolve the problem of wrong or incomplete suppositions at the beginning or within the reasoning chain. This is a fundamental problem. And it is the reason why scientists need a long time of training, experience and wide knowledge regarding their field of research. Plus the fundamental ability of building models and analogies in their mind in parallel to performing logical and mathematical steps and test these models against ensured facts – theoretically and practically. I do not see any indication of such internal model building, model understanding and model adaption to facts or fact-contradicting logical consequences in present day’s AI tools. They produce and reproduce much too much nonsense. But, maybe, this does no longer matter in the US world of “alternative facts” …
So far, I personally see no signs at all of any kind of intelligence in products like Perplexity, ChatGPT-5 and alike. But I am only an elderly physicist and retired IT-consultant. However, in the mail to my ex-colleague, I also referred to one of the godfathers of Machine Learning, namely Mr LeCun. He looks very critical upon the almost desperate trials of US Tech companies to hyperscale similar neural network architectures in the hope to find new capabilities in form of some yet unknown and unpredictable emergent properties some day in the future.
Well, aside the basic criticism of AI- and ML-experts like LeCun or G. Marcus, there is a growing number of research articles which have been published about (potential) basic deficits of today’s LLMs, even when equipped with add-ons to perform a chain of reasoning steps. Below, I just list up a selection of such publications. And ask my ex-colleague and other interested readers to study them carefully and come to conclusions by themselves.
- Study of Apple Research:
https://ml-site.cdn-apple.com/papers/the-illusion-of-thinking.pdf - Study of the Arizona State University:
https://arxiv.org/abs/2508.01191v3 - Study of Samsung, Montreal:
https://arxiv.org/pdf/2510.04871v1 - Study of the KTH Royal Institute of Technology Stockholm, Sweden:
https://arxiv.org/html/2502.11574v1 - Study of researchers at Stanford University and VianAI Systems:
https://arxiv.org/abs/2507.07505 - Study of researchers at MIT, Northeastern University and Meta:
https://arxiv.org/pdf/2509.21155 - Study of Microsoft Research:
https://arxiv.org/pdf/2505.08140 - Study of researchers at multiple universities (incl. Stanford, Berkely) on scaling:
https://arxiv.org/html/2511.12869v2
I think that one should take the different warning signs of these studies seriously. They all indicate that LLMs and a simple hyperscaling of transformer-based LLM architectures (with reasoning modes) may end up in a dead end.
We live in a crazy time when investments in the AI industry are expected to reach a size of 600 billion dollars in 2026. Most of these investments are spent for building data centers and all kinds of AI-infrastructure. All these investments are built upon hopes both that the invested sums will at some point in the future generate a substantial ROI (paid by consumers) and that in the course of architecture scaling mankind may some day stumble upon some new emergent features of LLMs showing some real intelligence. In my opinion this behavior of the competing Big Tech companies reminds strongly of a religion.
I.e., mankind invests gigantic sums into the hopes and claims of some already very rich people and not into science based progress in developing new neural network architectures which may provide information handling similar to what our human brain achieves with very little energy consumption. Instead we invest in AI data centers which may increase mankind’s energy consumption substantially in the next decade. For what? For hopes and claims of some very rich people trying to become even richer?