With this post I continue my list of arguments why I left ChatGPT. Not because of quality reasons. To follow my interests in some math projects, I do not need a frontier model, but a sufficiently “intelligent” model, which checks its results automatically with numerical examples. A reason for this attitude is that the rate of mistakes still is not zero even for frontier reasoning models and relatively simple problems. More importantly:
I want and need a KI-model as a free service which I can rely on with respect to saving and exporting my results. A service which supports me in a cooperative way over a longer period of time. In the first post of this series, I have described why OpenAI’s freely accessible GPT has disappointed me in this respect.
In this post I turn a bit to the Pareto principle and politics. Note that I am talking about using an AI Chatbot for limited projects in medium level math or medium level theoretical physics and not for intensive coding. I should also say that I have used OpenAI’s ChatGPT much more than Anthropic’s models. Let me in addition warn you: You have to check mathematical derivations and proofs of any AI model – whatever price plan you have chosen. So, the argumentation may look somewhat differently for other purposes of your AI usage.
I am well aware that most of the following points could be transferred from OpenAI to other AI companies – worldwide. But, as a European and a retired person with not much of disposable money, in the end it is a conglomerate of democratic principles, costs and sufficient service quality which affects a decision between different AI providers. If you want to phrase it in capitalistic terms: The market decides. And this may on a much more general level and over a longer period of time lead to problems for the leading edge US AI companies.
Math on university level
One development in the year 2026 has astonished me: OpenAI’s GPT and other frontier models have gotten significantly better regarding mathematical questions and related problem solving – at least on the medium level I sometimes request. Even in 2024 I had not expected this rapid improvement in quality. Meanwhile AI systems really have become usable assistants when dealing with math topics. However, this positive development of “intelligent” AIs has some negative sides, too.
The current AI-development and competition is predominantly led and guided by capitalistic objectives. One of these objectives of course is money as a return of investment. The promised or expected amount of profit rules the whole investment and capitalization cycle of the AI-industry. With the result that customers are expected to pay for the service offered. The conditions regarding free access to US based AIs have become more severe, the restrictions bigger and the limits of free reasoning or computational time smaller and smaller.
However, as we have seen for all rising industries based on technical breakthroughs, markets do not always favor the leading edge drivers in the end game, as soon as 80% of quality becomes sufficient, too. After this point, the last 20% of quality typically cost 80% of the investment. Sounds familiar? Well, ask an AI about the Pareto principle. It is one of the reasons why enormous amounts of money flow into leading edge companies in the race for an AGI.
Turning AI companies as OpenAI or Anthropic into share holder companies will raise the profit pressure. But the growing risks in big scale investments are accompanied by other general problems for our societies. So, let us turn to the point, I want to consider and criticize. I pose it as a question: You want quality, you got to pay. Really? Always?
Different use cases
I have some friends who use AI from time to time to create funny images and short videos. They all praise the improving quality of the results. Others use AI to prepare business presentations, summaries and the analysis of data ahead of management decisions. Some younger friends create SW-code with the help of AI – for products and customer related projects. Many friends instead turn infrequently to some AI-service simply to get answers to questions that pop up at their working place or in private life.
In all these cases I regard some exchange of information or fees as a price for an IT-service as justified. AI in theses cases supports projects from which mainly the project owner profits in one way or another. In the case of business oriented usage capitalistic profit interests are supported – which is OK. (I leave political influencing or manipulation out of my considerations in this post).
In most of the named cases no direct general improvement of a society’s welfare is intended. The private user does not share special techniques to create the images/videos with others, at least not under typical circumstances. A product and a business decision may prove to be good for customers and a society on some longer timescale – but the costs of AI in this case hopefully are more than matched by the profit of the AI-supported business. Gathered knowledge is not shared – such a sharing of know-how may even become a threat to continued profit.
AI itself has become an example of closed knowledge accumulation: After the release of GPT 2 and with rising competition, the weights of a trained GPT-model were no longer published by OpenAI. With the exception of a few companies and models, AI development has become a closed source business. Little knowledge sharing, but ever growing profit interests. OpenAI clearly is on this path.
While we are at it: The success of American AI companies depends heavily on investments of the public and administrative sector – including the military, of course. So, regarding sales, the tax payers in the US and in several ally states already contribute heavily to the growth of AI tech companies.
I do not want to go into a discussion of capitalism driving the welfare of western societies or not. Capitalism had and has a role in innovation and during the establishment of new technologies via mass products. Neither do I want to discuss the impact of the military on the AI development or even the rights of a government to maximize its military power by exploiting AI. (At least not in this post).
Education and science as public domains
However, there are good reasons why advanced societies and democracies regard education and science as sectors which must be supported and financed by a state’s administration. Based on taxes and not on investments in private enterprises. At least in Europe.
Education is a society’s investment in its future. With a major obligation of democratic states and administration to ensure equal access to accumulated knowledge, independent of a citizen’s financial or social background. In addition, education and science are based and driven by the sharing of knowledge. Publicly within a state and even across the border of nations. If we leave this path, we harm democracy and betray the very principles of enlightenment.
The problem these days is that AI has more and more become a relevant tool of knowledge development and knowledge accumulation. On very short timescales … But, predominantly for customers with money. The other problem is the on-the-fly accumulation of knowledge, data and AI-centered research within the big tech companies like Alphabet, OpenAI or Anthropic.
But let us turn to people who are in the phase of acquiring knowledge, i.e. people who study …
AI and the education of students and freelance researchers
My personal level of math has unfortunately become medium or lower – measured on university levels. After having got an education as a theoretical physicist up to a PhD degree many decades ago, I have worked in IT up to retirement. Frankly, during the decades in IT I never needed any math at university level. One gets a bit out of touch and practic. So, talking in my remaining years with an AI about math is exciting – and during some sessions, I have quickly refreshed stuff again which students in physics and math learn during the first six semesters.
This personal experience has told me that for committed students a capable and adaptive AI on today’s level and beyond can be a really helpful tool and sparring partner during their studies, their bachelor and master thesis work and beyond. Without any doubt. As long as they keep on thinking themselves first before turning to AI on advanced challenges.
The question is: Who gets access to such a capability boosting tool?
Another group of people may be the elderly ones – well educated and still having some interest in scientific questions – but no longer associated with a research institution. Willing to share their gathered knowledge or AI supported thoughts on scientific problems. In my case via planned Opensource books.
The role of money …
In my last post, I have discussed that a free plan for using GPT has very narrow limits regarding resource usage. Serious work with GPT on advanced math or physics projects requires a real investment, i.e. money.
More importantly: The gathered knowledge, proofs and results as well as summaries (e.g. in LaTex) must be available at any time during a longer project (like writing a math book). Regarding ChatGPT, this is only given when the user transfers money to OpenAI. The same is probably true for Anthropic. (Write me a mail, if you have different experiences.)
Considering enormous amounts of net loss of some AI companies in the first half of 2026, a financial analyst might say: This transfer of money is absolutely necessary to keep the long term business model of AI alive for private investors. And to avoid an AI bubble from bursting.
Well, let me contrast this development with the result of a recent study: A large percentage of German students (40%) live in poverty. These guys do not have the money for the requested AI price plans available. And: OpenAI presently (Oct. 2026) offers a temporary 4 months free plan including Pro-features only for students at US universities. Not funny.
A (maybe intended?) side effect is: Students with relatively rich parents studying at US universities get a huge advantage during their studies – if the world worked according to ideas of OpenAI (and other US companies).
Thereby the present AI development increases the weight and importance of a general political question:
- Do we want to create societies where only people with money get the best AI tools and thus also the best education?
My answer as an European is: No, we do not want to support such a development. Democracy must be build at least on the promise that the chances for every citizen to get the best education possible are equal and independent of the income of their parents or relatives.
Does OpenAI (or any of the other leading AI companies) offer a permanent free or low cost plan for students and scientific studies? No, as far as I know, this is not the case.
Does OpenAI (or any of the other leading AI companies) offer a permanent free or low cost plan for scientific work in general? Maybe under the obligation to cite the contribution of GPT? No, it does not.
You could say that a purpose related use of AI resources would be difficult to control by the AI provider. Counter question: What is their AI good for? It could of course check the contents exchanged with the customer. Well, and other companies in the scientific publishing realm do it – e.g. companies where the access to articles is based on whether the customer him-/herself has already published in the science domain.
By the way: My experience is that your data during longer math sessions are used for future training if they are interesting enough. If so, it is simply a deal: With every interaction the AI company may collect and centralize more knowledge to a pool, hopefully available to many others in the future. Well, for real front-research there is the question of intellectual property and respective rights, which may require special contracts. But this point makes the point of payment for AI usage for scientific tasks even more urgent and questionable.
In my opinion, with the growing competition among AI companies the whole development more and more undermines the basic requirements of a modern, democratic society: Equal chances regarding education and the access to respective tools.
There would be a simple solution:
The state limits the access of AI-companies and their data centers to land, energy and water resources and charges significant taxes for the use of these resources. Tax reduction is granted when the companies offer students and scientists special free or low cost plans for the access to their models. Oh, my god, isn’t this a kind of socialism? (At least, regarding the strange ideas some rich Americans may have about socialism). My answer is: No, some taxes for the general welfare – and in particular taxes to support for education and science – are a question of civilization.
Alternatives and the market
Is the free plan of OpenAI without alternatives – regarding longer projects and the saving of achieved results? No, it is not.
Noteworthy, I did not yet experience such strict limits as with OpenAI when using the Chinese DeepSeek4. Answering speed sometimes gets a bit low. But achieved results are not deleted (without warning). The quality on my requested math level (mostly undergraduate, up to 4th semester, and partially above) was convincing – though I sometimes wondered a bit about the “style” of attacking a math problem. Nevertheless, I found DeepSeek4 rather convincing. A real problem was the missing export via Latex code.
The other alternative is the French Mistral Vibe. (I will write in more detail about my experiences with Mistral in another post.) The essence is: For my purposes considering math, Mistral is an almost ideal companion. In particular due to its directly offered method of testing results of mathematical derivations via numerically evaluated examples and respective programs created on-the-fly in the background. This has in some (seldom) cases led to a convincing self-correction of a mistake during reasoning loops – which the AI frankly explained to me afterward. Export of compile-able Latex-code is working perfectly.
I also like Vibe’s continuous maintenance of a list of errors and a repository of major results and tricks to learn from for further tasks in the same math domain. The somewhat “sarcastic” reaction to pitfalls in some tests I have gathered over time and confronted Mistral with amuses me. At least I regard detecting a pitfall and calling it a “double diabolic trap” – as Mistral sometimes does – a kind of sarcasm.
So, for studying math problems there are alternatives – outside the closed source and capital-only driven AI development in the US.
For me as a liberal European, the choice between an US AI with significant costs or a Chinese AI for free amounted to something like a choice between pest and cholera. In the case of DeepSeek predominantly for political reasons.
Pareto
Some readers may now point to ranking lists of AI models for math. You may find DeepSeek V4 not in the top positions there and Mistral far below. You know what? I do not care any more about such lists.
In the relatively wide range of my problems (multivariate statistics, hypersurfaces and multivariate calculus, Linear Algebra) GPT Luna has so far made more mistakes than Mistral – and in parts really stupid ones. This is one of the major lessons I have learned in 2026:
Test AI models intensively for your class of problems. It does not really help you that a model can solve frontier math problems when it – at the same time – stumbles into elementary traps in much simpler tasks. Such an experience will only make you angry if you had paid for it and find that even the best AI models still make silly mistakes. An example: I have got 70 pages of a really intelligently planned proof from GPT Luna which contained a blatantly stupid mistake on page 32.
Having reached this point, the Pareto principle comes into the game:
If and when I can solve my class of math problems with a sufficiently intelligent reasoning engine of a Chinese or European AI and when I can build up a kind of knowledge gathering system from which also the AI itself learns during the continuous interaction – and all this for free or significantly less costs than I would have to pay for a US based provider – why should I stay with a US system? In particular when the US system also makes mistakes?
I do not say this lightly or because I do not like Big Tech companies and their cooperation with the present anti-european US administration. I say it because my own example can be regarded as a general warning for those who invest in such companies:
- The west and in particular the US could loose the AI war with China due to the Pareto principle. In the end it will be the market and a reasonable balance between costs and service that decides. Frontier models require enormous investments. But it is more than probable that international customers will not pay major amounts of money, if or as soon as these customers can get 80% of a service on a reliable basis for much less money or in the case of math topics even for free.
The point of costs is certainly relevant for students who do not have rich parents, but are willing to learn. Together with an AI model – even if it is not at the top of some dubious chart lists. Due to present price policies we may drive our young students to use Chinese AI products – and this is something we should absolutely avoid.
Which brings me the point of what European governments should do regarding access to the only relevant European AI, namely Mistral Vibe. But this is stuff for another post.