Martin's blog

GPT-5 Hangover

Yesterday, I watched the first part of the GPT-5 presentation, but only after the incredible comments started pouring in on HN. More on that in a moment, but the impression this presentation made on me was just as disastrous as the whole AI bubble I'm currently feeling. Sam Altman's introduction is simply unappealing and comes across as a dishonest know-it-all from days gone by. The team he then gradually brought in – oh my God. All very young nerds with no charm, no rough edges – they may be AI experts, but they have no life experience and no wavelength that you can relate to. The whole interior and setting looked like it came from a catalog or was already AI-generated. I felt like I was in a terribly sterile artificial world with artificial people and an artificial theme – but that fit the content.

And right at the beginning, there was the wrong graphic in the presentation of the programming benchmarks. The HN comments summed it up well – there's simply no excuse for this kind of presentation and company evaluation. And it's not a mistake that can be recognized, discussed, and then put aside. It says so much about the quality, culture, and hubris of those responsible that I'm really surprised why various press reports on day 2 don't comment on it and only reflect the characteristics, expectations, and assessments of the product itself.

The second point is the demonstration by the “PhD” expert using the example of Bernoulli and airfoil theory. I even had to fetch my study folder for the ME163 “Engineering Aerodynamics” course from the basement. I took the course at Berkeley University in 2003. The theory is complex, and I remembered the lift calculation using various methods (vortex sheet, source panel, analytical, numerical, etc.) and models (simplified thin, with cross-section, etc.). And I was “only” a master's student, not a PhD student. In any case, the consensus is that the method specified by GPT-5 in the demo is incorrect and falls far short of the mark. What's the point? It was a launch presentation with promises to the whole world, not an undercover test among many in the basement.

I wonder more and more what else is being done to further inflate this AI bubble. The technology is there, it was impressive at the beginning, but it does not open the way to ever greater heights such as AGI. This afternoon, I thought of an analogy: flying. The invention and development of the airplane was groundbreaking, and over time, flying became safer, cheaper, more efficient, faster, more comfortable, and so on and so forth. But flying in 2025 is fundamentally the same as it was at the beginning of the 20th century. We sit in a machine with wings, take off, move through the air at high altitude, and finally land back on the ground.

The difference between flying and LLMs is that flying is based on honest, understood engineering science. LLMs still cannot function in a reproducible, transparent, and 100% correct manner. This is not only due to the business practices of the companies involved, but is inherent in the applied mathematics of the algorithms. But this characteristic—a pillar of civilization (correctness, reliability, trust in these two characteristics)—is simply dismissed as an unattainable “side condition” in LLM. “Hallucinating isn't nice, but it's becoming less and less common.” I still agree with those who say that a problematic bubble is growing here, but that it cannot grow into a sustainable business. Or why should castles suddenly be built on sand when our ancestors were already further ahead?

I expect LLMs to continue to improve, but in principle they are approaching a plateau. The marginal utility is decreasing. Sometimes there are still outliers and surprises (supersonic aircraft, flying wings, etc.), but at some point the technology will reach a balance and humanity will know what it can do with it and what it can get out of it.

I hope there are still enough people who can think critically, who are not blinded by power- and money-hungry Big Tech CEOs, and who know what basic human qualities such as thinking, creative expression, and creation mean. Either these activities are strenuous and must be learned, practiced regularly, and developed, or they are a source of joy, ease, and human satisfaction. Both poles belong to human beings, are experienced and applied anew by every child, and are inseparably linked to our existence and growth as human beings—for the individual, but also for the history of Homo sapiens as a whole. I would be delighted if the many, many billions in investments were directed more toward areas that would support us humans more in the “less human” activities. For me, that means support for difficult or risky physical work, for example. On the other hand, show me robots that can do half of what you promise with your LLMs in critical domains! Then we can talk about real, sustainable benefits for humanity.

It's better for the bubble to burst sooner rather than later. Wherever real human development is concerned, such as education, health, and social issues, it doesn't feel right to rely too much on LLMs here. And in the rest of areas, I believe that quality will suffer to such an extent that, in the long term, human well-being will also be negatively affected, or that we will inevitably degenerate mentally, yes, become stupider. This is already happening due to the glut of AI-generated content on the internet, in companies, and in communities of all kinds.

Translated with help from DeepL.com, original in German language