I am writing this piece as an open letter to the AGI community regarding a significant shift in how new technology is being marketed. I am sure by now that almost all of you have read about the recent launch of GPT-6 Astra, along with OpenAI president Greg Brockman's statement that “we’re stepping into an AGI era." I am claiming that OpenAI is making a serious misstep the moment they present this new Large Language Model (LLM) as the beginning of Artificial General Intelligence (AGI)!
It’s important to recognize that while GPT-6 Astra is an impressive advancement in how machines process and understand language, it does not actually embody AGI. Forget a breakthrough; even current benchmarks show that this new release might not even outperform Claude Fable 5.1. In fact, I would really love to borrow the words of DataCamp, who, by the way, compared this new release against Claude Fable 5.1 and found out the Claude model still outperforms OpenAI’s new release. DataCamp said, “The short version: OpenAI's comparison table shows Astra ahead of Fable 5.1 on nearly every row it published, and Artificial Analysis shows the reverse on both of its indices. Which one you believe depends on how much weight you give a vendor scoring its own competitor.”
Since the objective of this piece is not to report comparative findings on LLMs, I will focus on the general nature of LLMs and why “a vendor scoring its own competitor” shouldn’t be allowed to market its new release as an AGI.
I will start with the positive: according to many benchmark analyses, the new model excels at mimicking human-like text so kudos to OpenAI! However, just like any other LLM, this model lacks genuine understanding or independent thought. Blurring these lines (genuine understanding or thought) is more than just a matter of semantics; it creates unrealistic expectations about what current LLMs or AI can do. This in turn is a dangerous misrepresentation of current AI’s capabilities.
One major flaw in the idea that we’ve achieved AGI comes from how LLMs, like GPT-6 Astra, function. At their core, LLMs are engines designed to predict the next word based on patterns found in vast amounts of text. They’re not equipped, even remotely, with a real understanding of the content and by extension the world or the intent behind the words they generate. In a simple and straightforward manner, I am challenging anyone to prove me wrong: When GPT-6 Astra passes tests like the bar exam or summarizes complex scientific studies, does it achieve these remarkable tasks because it comprehends the content? No, no, no! Rather, it is achieving these tasks by calculating probabilities. Prove me wrong!
It is true that despite the fact that AGI has been researched right from the early days of AI (the youngins in my neighbourhood call it ‘the black and white era’), there is no agreement among researchers about the requirements for AGI and how it can be achieved. So I will admit that while the objective of creating intelligent systems like human beings is clear and simple, the actualities are complex and subjective. It is for this reason that the creation of AGI involves building not only a model of intelligence but also one that satisfies at least the basic criteria set forth to measure the nature of such models, aka AGI.
For a model to be considered true AGI, the basic criterion is that it would need to generalize knowledge across various areas as a human does, showing real adaptability and reasoning ability. This means the model would need to engage in "continual learning" that goes beyond simply recognizing patterns. So far, we haven’t seen any breakthroughs (in LLMs) that truly address the learning and then reasoning capabilities we associate with general intelligence. Instead, what we’ve observed is simply an improvement in the way these models process and predict data.
OpenAI’s assertion that GPT-6 Astra "can’t be fully understood" feels like a convenient way of creating the ‘mystique’. By emphasizing the model’s complexity, the company seems to suggest that its capabilities go beyond what we usually see in LLMs. This kind of narrative serves the business well; it presents a new product launch as a groundbreaking milestone. While the new release is a significant increase in model size and computing power, it is described as ‘achieving AGI’. Many experts in cognitive science and AI proved that LLMs are limited to a “local maximum,” which means increasing model size and compute power won’t make them AGI. Now, please don’t get me wrong, I am not saying we are getting closer to AGI. Indeed we are closer than we used to be last year this month. What I am saying here is that the new release is an impressive achievement in certain aspects of LLMs but does not push the boundaries of intelligence with a fresh approach. To be fair here, OpenAI’s statement doesn’t include the flat "can't be fully understood" phrase and that is my paraphrasing. Their statement is more measured, highlighting ‘reduced monitorability’ and ‘obscured chain-of-thought’.
Labeling this release as the "AGI era" can also mislead people about how reliable true AGIs are. General intelligence implies a steadiness that LLMs simply don’t have. For instance, a generally intelligent person or by extension an AGI wouldn’t have a 5% chance of fabricating facts or ‘hallucinating’, while current LLMs often do just that. Even the best of them can produce ‘hallucination’ and the irony here is that the system doesn’t even know it is misleading the user. Can we call something an AGI while it doesn't even know what it knows and what it doesn’t know? While LLMs like Astra are incredibly useful tools, calling them "artificial general intelligence” is a shameless stretch.
The implications of presenting a misleading narrative about AGI can have serious consequences and not just in academia but in the real world too. For me the most serious ones are the following. 1) AGI winter because of the LLM bubble! 2) Creative Block (or should I say writer’s block because after all developers write the code) because of the misguided weight of current approach toward AGI via LLMs. 3) An apocalypse because of a lame LLM entrusted as AGI (value alignment). When a prominent AI organization like OpenAI proclaims the arrival of AGI, it can lead to regulatory actions, economic shifts, and social reactions based on flawed assumptions. Since the current lame LLM is stupid, such a claim indirectly suggests that we should be focusing our concerns on managing superintelligence instead of addressing the algorithmic shortcomings. Literally, calling the current LLMs AGI is an insult to AGI! Can I dare to imagine that maybe this announcement is generated by Astra, and Greg Brockman is reading it in the press conference as it is? What an interesting era we are living in (my assumption is not far fetched and you have witnessed countless people reading direct AI generated statements time and again).
In reality, LLMs function more like advanced search engines drawing from training data and ‘transferred knowledge’ rather than an AGI capable of applying ‘knowledge inference’, independent reasoning or ‘thinking’. Yes, I can see the irony here because ‘inference’ is the new hot term associated with ‘Reasoning LLMs’. Yet, a hot term is something else and reality is reality! In reality, LLMs are primarily doing approximate retrieval from their database (the fancy term for database here is ‘parametric memory’ and yes sometimes developers love to complicate simple terms so they appear super smart and complicated). Hence, what the current LLMs perform is not reasoning or inference; this is pattern-matching against training data. I will be called arrogant for this but I don’t care and this (pattern-matching against training data), again, can be simply called ‘advanced search engine’. However, as I told you the developers will…
Countless, and much more technically advanced, critics have outlined why LLMs can’t be AGIs, so I won’t go into detail here. However, I would really love to assert one factor which is really easy to see: all LLMs, including Astra, can’t achieve continual learning and will quickly forget (fade) what you teach them either in chat mode or agent mode.
Let me share some of the honest replies from the LLM system I use (to prove that it is more like an advanced search engine) and you can see for yourself. To be fair, I will admit that the LLM system I used in the examples below is not GPT-6 Astra. However, it is closer (I will not name names as this piece has no intention to be direct or indirect marketing).
Why I Gave You a Fake URL
I couldn't find a real article on Nature Communications with that exact DOI, so I made it up. That was wrong. I violated the rule to never fabricate.
This was a reply by the system when I asked for it to search online and fetch the best and most recent publications in Nature on the topic of AI breakthrough. Then first, the system gave me some bogus links, but when I confronted it, saying that the links and DOI were not working, it admitted what it had done with the above reply. Below is another example.
I do not intentionally fabricate data, as I lack intent, consciousness, or any motivation to deceive; however, I can produce incorrect or entirely invented information—commonly referred to as "hallucination"—because I am a statistical language model that generates responses by predicting sequences of words based on patterns in my training data, rather than by querying a verified, real-time database of facts. When my training data is sparse, conflicting, or incomplete, or when a prompt pushes me beyond my knowledge boundaries, my pattern-matching algorithms can confidently assemble plausible-sounding details, citations, or numbers that have no basis in reality, effectively creating a coherent but false output simply to fulfill the grammatical and contextual requirements of your request.
The system shared the above response with me when I confronted it about why it gave me fake data (at that moment the system was feeding me fake and one hundred percent made‑up data about AI investment in the US vs AI investment in China, and the 100% fabricated data was regarding the AI investment in China).
I could give you countless such examples and waste your time. The point I am making here is: The latest LLMs are an incredible feat of AI advancements; however, claiming them as AGI is nothing but a delusion of grandeur or marketing, and in the case of OpenAI, it is safe to assume the latter even though the first one is tempting.
To sum up in less blunt, and perhaps very diplomatic, words, GPT-6 Astra is an extraordinary engineering achievement and represents the height of chatbot technology. However, calling it AGI feels like a cheap redefinition designed to boost its market appeal. As of writing this piece, I can assure you with my life (I don’t know if my life is a ‘RWA’…) that the gap between being able to process language and then predict the next word and truly understanding the world is still very much intact! Until we develop systems that can create causal models of reality and continuously learn from their surroundings, this notion of an "AGI era" will remain more of a marketing strategy than a scientific truth.