There is nothing wrong with GPT-5, we are the problem

The false prophecy

Many industry leaders prophesied that AGI would arrive in 2025. Sam Altman proclaimed in his blog: “We are now confident we know how to build AGI as we have traditionally understood it. We believe that, in 2025, we may see the first AI agents join the workforce and materially change the output of companies.” [1] I agree with the second part. AI agents will join the workforce in 2025. It’s the first part that I find problematic.
 
The prophecy of near-term AGI is based on LLM scaling laws, empirical observations that model performance scales logarithmically with compute, data, and other resources needed to train these models. These laws have been shown to hold over many orders of magnitude and are seen by many as a sign that all you need to reach AGI is to throw more money, electricity, and data into the mix.
 
Altman himself wrote recently: “It appears that you can spend arbitrary amounts of money and get continuous and predictable gains [in intelligence]; the scaling laws that predict this are accurate over many orders of magnitude.”[2], a profoundly flawed and troubling statement, yet one which almost perfectly predicts what we should all have expected from GPT-5.

Let me explain.

While it is true that logarithmic growth (in this case, “intelligence” or model performance) continuously increases with the increase in input resources, there are two crucial points that Altman and others, consciously or unconsciously, ignore.

First, there are no infinities in the real world. At best, a logarithmic model of performance increase is an effective model that allows you to predict improvements “for a while,” but ultimately must fail. This usually happens because, if nothing else, you will eventually run out of the resources you need for the improvement (e.g., money, compute, electricity, valuable data, GPUs …). These limitations are not considered in the LLM scaling laws. The assumption is simply that we can keep pumping resources into model development to the point of exhaustion.
 
We already have evidence that resources are becoming strained. Epoch AI predicts that by 2030, models may require as much as 10-15 GW of power[3], equivalent to one and a half to two of the most powerful nuclear power plants in existence. The CEO of Anthropic recently asked the US government to provide 50 GW of additional power production by 2027, an equivalent of 2 and a half Three Gorges Dams, the largest power plant ever built.

In addition, model training costs have already crossed the 1-billion-euro mark, and we have effectively exhausted the valid training data of the entire internet. Synthetic data won’t help either, as it doesn’t add new information to the training datasets. All these limits are putting a strain not on the logarithmic scaling, but on how far into the logarithmic tail you can go.
 
Second, while logarithms continuously grow, they grow in such a way that for each additional percent of improvement, you need an exponentially higher amount of input resources. This is an especially prominent problem in the “tail” of the logarithm. You don’t believe me? A log (base 10) of 10 to the power of 25 (1 with 25 0s behind it) is 25. The same log of 10 to the power of 30 is 30. An increase of 20% in this log region, hence, requires an increase of the inputs to the log by a factor of 100.000! That’s how logarithms work.
 
There is sufficient evidence to suggest that we are living in the tail of the scaling law logarithms. The relative performance improvements between GPT-4, GPT-4o, GPT-5, etc. suggest order (5-10%) effects, while the resource inputs increased by a factor of 10 on average with each iteration. This means that, precisely as the scaling laws predict, throwing more money, compute, power, and data into training the next GPT will yield diminishing returns.
 

The question then is not why GPT-5 is not a quantum leap to AGI compared to GPT-4.
The question is: Why would you expect this to happen?!

The right way

The problem is not GPT-5. The observed performance of this model is well in line with everything we know about how the model performance should scale. The problem is that too many of us choose to ignore the facts and instead drink the Kool-Aid made of false prophecies made by a handful of influential individuals.
 
The sad fact is that everyone is focused on the disappointment with GPT-5, where in fact, GPT-5 is a state-of-the-art model with incredible capabilities. And it’s not the only one. The recently open-sourced GPT-Oss class of models, as well as OpenAI's proprietary competitors Claude and Gemini, are compelling models. Why are we obsessed with reaching AGI tomorrow when we already have models of exceptional capabilities at hand, here, today?!
 
Instead of investing hundreds of billions of euros in monolithic AI, which will yield diminishing returns, we should focus our resources on leveraging the powerful technology we have today, combined with our intellectual abilities.
 
Small language models are making a comeback as more specialized yet performant executors of roles and tasks. Multi-agent systems promise to leverage the best of large/small language models and humans, collaborating in a way that works better than the sum of the parts. Human/AI hybrid teams are a way to improve the performance of AI systems by leveraging models like GPT-5 in areas where AI excels and leveraging human intelligence in other places. This is what we do with human-only teams already: we leverage collaborative intelligence to solve complex problems.

The only difference is that AI is now a team member as well. Such systems have a host of advantages over monolithic AI in addressing practical challenges with confidentiality, data governance, sustainability of AI, etc. We should at least pay the same attention to them as to the development of ever larger LLMs because there is already enough value to unlock from both human and artificial intelligence we have today.
 
Overall, I bet that in the near to medium term, we will see many more improvements in AI performance coming from the multi-agentic ecosystem built around LLMs than from the upgrades in LLMs themselves

Can we reach AGI?

To be honest, I don’t know, because I struggle to even define what AGI is. And I’m ok with that. What I do know is that it will take more than belief in false infinities of the scaling laws to beat humans at all intellectual abilities we display (maybe we call this AGI?).
 
At their core, all modern LLMs are compression algorithms. They are also exceptionally efficient compression algorithms. With LLMs such as GPT-4 we managed to compress almost the entire internet into about 1 trillion floating-point numbers. If that doesn’t make your head explode, I don’t know what will.  By compressing such a vast amount of textual data, LLMs have “learned” two crucial things: the structure of languages and how to understand textual context. A combination of these two features is what makes LLMs so valuable.

More recently, people have also been able to use these capabilities to allow LLMs to use digital tools, giving rise to the era of agentic AI, perhaps the most significant recent advancement in AI.
 
As it turns out, simply understanding the meaning and context of text is sufficient to automate a majority of what we call intellectual work today. Copying forms into a computer application, summarizing text, and even writing computer code are now things that (agentic) LLMs can do.
 
However, this is not AGI.
 
True AGI will not come only from compression. True AGI will likely require deep architectural and other advancements in AI models, and not just more of the same. For AGI, we will need not only logic and cognition, but creativity and intuition. True AGI will also require access to the real world. The Internet is a vast source of information, but to “learn” concepts such as physics, AI will need to observe the actual physical world we live in.

A reality check

None of this means we should give up on AGI. But we should approach it with humility. Scaling alone will not get us there, and believing otherwise leads to disillusionment. Today’s AI is simultaneously exciting and scary, super-smart and stupid, astonishing and clumsy. The rational response is neither blind worship, nor despair, but balance: using AI wisely while acknowledging its limits.

Rather than obsessing over false prophecies of near-term AGI, we should focus on the immense value current AI systems already provide. GPT-5 is not a failed step on the road to AGI. It is proof that we need to rethink our expectations and redirect our resources. If we do, the real breakthroughs may come not from chasing infinities, but from combining AI capabilities we already have with the ingenuity still unique to us.

[1] https://blog.samaltman.com/reflections
[2] https://blog.samaltman.com/three-observations
[3] https://epoch.ai/blog/power-demands-of-frontier-ai-training

Enjoyed this insight?
Share it to your network.