Some days ago, I had my first longer “conversation” with GPT-5. Below, I just summarize some general impressions, not hard and evaluated facts. I have no time to perform thorough tests. Better founded pro/contra arguments will come with future use.
In my opinion, fundamental reasons speak against an AGI based on pure scaling of networks and the addition of some loops which check statistically found argumentation elements step-wise against elementary logic. So, I had and have not taken the announcements of Sam Altman and any speculation about a major achievement in the direction of an AGI as serious information. It appeared to be an advertisement campaign – and my first impressions of GPT-5 confirmed this suspicion.
GPT just is and remains an interesting and valuable tool, which has to be used carefully and with some guidance, for the time being. Results of your interaction with it still have to be questioned, checked and verified – whatever kind of topic.
Having said this, I want to state that, based on my preliminary impressions, GPT-5 represents a major step regarding its capabilities as an information tool – at least in comparison to the “free” version of GPT-4. Not more, not less. Do not even think about it in terms of something like a general AI or, even more stupidly, as a personal companion.
Tech talks
In my case I had a discussion about a Machine Learning project during which I want to use an Autoencoder based on ResNets. We also exchanged information and facts about the relevance of multivariate normal distributions for the representation of basic properties of human faces and possible limitations – in particular for the generative recreation of interesting faces, which may not fulfill some standard norm, by a ML-algorithm. The discussion meandered around well known arguments regarding the use of GANs vs. deep networks based on convolutional layers for image analysis and a generative production of imaged objects from gathered latent space information. We also passed areas of hard mathematical facts about MVD-analysis, statistics in general, latent space analysis via PCA in combination with MVD-assumptions, the convergence of GANs and so on. Mostly with knowledge clearly reproduced by GPT-5 from published papers.
But, I also got some clear (and hopefully helpful) hints regarding the choice of certain parameters in my project – based on some mathematical considerations. All in all, I have to admit that I, at least on first sight, got very clear information within a clear line of thought and argumentation. But see the addendum below.
I have had similar discussions with the free GPT-4 version on some Machine Learning topics and related math before. I can say that the depth and logic of the discussion with GPT-5 felt much more convincing and concise than with GPT-4.
It was also interesting to note that most of my questions triggered the “deeper thinking and reasoning mode” of GPT-5. And this improved the results substantially in comparison with what I got from GPT-4.
In addition the permanent and useless flattering of the user has almost disappeared. What a relief …
I also want to remark that GPT-5 followed two switches to other languages than English during the conversation effortlessly.
In a second shorter conversation, I asked some questions regarding physics and math for which I previously had got wrong answers by GPT-4. This did not happen with GPT-5. The reasoning mode now presented correct answers.
These were conversations about technical and mathematical topics. And here I actually do see many advantages of GPT-5 in comparison to GPT-4.
Frustrating Addendum, 08/12/25-afternoon:
Hmmm, first sight impressions can be incorrect.
I just sat down and analyzed the suggestions of GPT-5 regarding the structure of a ResNet V2 based encoder and decoder block (which others call a Residual Unit comprising so called Residual layers). Unfortunately, the suggested block/structure was incomplete regarding the bottle-neck structure of a typical ResNet V2-block – according to the original publications of K. He, X. Zhang, S. Ren , J. Sun in 2015. According to the advice of GPT-5, a ResNet V5-block just differs from ResNet V1 by a changed order of activation and convolution – with a starting pre-activation ahead of a Conv-layer. GPT-5 totally ignored the bottle-neck structure and the kernel-size variation at the middle residual layer of a typical ResNet V2 block – both for encoder and decoder.
What do we learn from this?
Do not trust anything what an AI suggests to you – even if it switched on a thinking mode.
A talk about gender-pay-gap
Then I followed a suggestion of some friends in a private chat-group and asked GPT-5 about the existing gender-pay-gap in western countries. And here the quality dropped somewhat in comparison to tech topics.
GPT not only gave me numbers for statistically evaluated pay-gaps in different so called “western” countries, but also offered an analysis of the resulting disadvantages of this gap and other general gender related in-equalities. In particular, it offered an analysis of the structural reasons for the pay-gap plus methods to overcome these structural obstacles. Okay, …
While there was nothing wrong or improper about the suggestions coming from GPT-5, I got the impression that the almost power-point like presented arguments reproduced basics. Standard points, nothing on a PhD-level as Sam Altman has promised. More on the level of an average structured newspaper article on the topic. Any advanced high-school or German gymnasium student could have written the proposed analysis and counter-measure agenda. From a PhD-level, I had expected some more original, something scientifically founded and new ideas – but no sign of this. Nevertheless, my friends, which had used GPT-4 for some time, too, were surprised about the length and structure of the answer. But, frankly, what I got is the minimum I expect from an analytic information tool.
Something that I totally missed during the conversation about this topic were suppositions and requirements a modern society and its labor organization as well as companies would have to fulfill to make the proposed suggestions really effective with respect to structural changes. I also missed references to some scientific work regarding the development of modern societies based on an organized division of labor – and any reference to the basic orientation of a state regarding relatively free capitalistic competition or competition organized by the state.
One reason: Not all western states can be regarded as free, liberal societies any more. I criticized GPT for this. And asked how the agenda of possible actions had to be changed and adapted to an autocratic state, which we see developing under Trump. Now, it got funny for a liberal European. Really – and at the same time, so sad …
GPT-5 took the question absolutely seriously. And clearly indicated that one had to “frame” the topics to achieve something positive regarding the objective. I have added a respective longer part of the answer as a PDF.
PDF containing GPT-5 answer to “frame” measures for overcoming a gender-pay-gap in Trump’s USA
I left out the answer of GPT-5 to the added question what the requirement of framing actually means regarding the level of maturity of a modern state and of its leaders. It may not have pleased some readers in the US …
I did not go deeper. But my impression was that GPT clearly needs guidance regarding a really deep analysis of a non-technical social or political topic. You have to insist on references to political and economical and sociological sciences and papers in these fields. Otherwise the discussion remains superficial – though (maybe?) not wrong.
The question also remains whether GPT-5 adapted to me as a German conversation partner. And what it would have answered a US citizen. I liked the answers I got regarding the framing for a “Trump-era political climate”, but would the answer have been “framed” by GPT for an US citizen? As GPT5 obviously is aware of this option …
Find out for yourself …
Other criticism
While GPT-5 is no longer flattering the user – at least not me – it still is chatty and sometimes diverting from a topic with its suggestions. With chattiness I mean the offer of non-requested information about side-aspects of a topic within the answers. This chattiness has advantages and disadvantages. In some cases it may give you valuable information. Sometimes, it may just reflect a reaction to a too open question. In this case the triggered thinking and reasoning mode may react over-carefully – to not to forget anything. I will observe this phenomenon more closely throughout further tech-talks.
Major disadvantage
The major disadvantage of GPT-5 is: It is not really free.
To be able to use the thinking mode – and this is the real advantage in comparison with the free GPT-4 – you need to have an account at OpenAI. To get it you at least pay with information – not only regarding the topics and details of your conversations – but also with a personality profile. You have to accept that this information can be used by OpenAI for further trainings – and maybe for other things as well.
Why do I stress the point about a “profile”? Because, only after 2 conversations, GPT-5 offered me a personality analysis based on my interaction with it. Not funny, not at all …. for a whole bunch of different reasons.
General advice
One thing I got aware of after the conversations was how smooth and effortless they went – in contrast to GPT-4.
So, if OpenAI has succeed with something, it would be: The illusion to talk to a human is better supported with the new version.
However, this is, in my opinion, a completely wrong and dangerous development – and it clearly has an economic side. OpenAI and others – as Meta – openly move into the direction of offering consumers so called “personal companions“. Things we trust in and share our thoughts with. This does not only open a door to creating dependencies and to earn billions of dollars, but it also opens the door to mass-surveillance and control very wide.
If they really go through this door, everything that we feared and still fear of personal information exploiting monopolists like Google and Meta, would only be remembered as a skinny shadow of what is going to come. If we fail to get some political control over these ideas openly pressed forward by the capitalistic leaders of a growing tech-oligarchy in the US …
So, folks, listen to a relatively old man who has seen enough of hidden and open manipulation:
GPT-5 is still an algorithm, nothing more. Do not take this as a simple statement – think the meaning of it and what “algorithmic” means through to the very bottom and carefully evaluate the logical and philosophical consequences. Accept that “algorithmic” under most practically relevant conditions also means “deterministic”. It is almost sure that our brain – even if it sometimes works algorithmically – works differently than today’s artificial neural networks. And if you really think that our world (including our brains) is purely deterministic, anyway – read a book about modern physics.
Aside of this aspect: GPT-5 is not yet based on real experiences with the real world. It has learned from knowledge and prejudices we humans have published somewhere – and most of what mankind has published on the Internet in recent decades is unverified and totally superfluous nonsense and in the best case nothing more than a bunch of prejudices. GPT is an information evaluating tool within strong limits and still based on information containing a lot of stupid human utterances. Garbage in, garbage out – whatever OpenAI may suggest its algorithm now is.
It may follow guide lines imposed by “reinforced learning” – but as any human guideline these guide lines are imperfect – and in addition driven by economic interests these days. This is no secret.
Therefore: GPT may adapt to a user’s way of conversation – but a human enemy would do so, too.
So, do not get emotionally involved and keep your intellectual distance during “conversations” with an AI. Do not use it for psychological analysis or profile creation of your or another person’s personality. In the end it is just a stupid machine – though maybe somewhat less stupid regarding some (!) logical problems than before.
But would you let a child that starts to think a bit logically sometimes, let you guide you through important decisions in your life? No, you would not.