We Are at 57% of General AI

An article by Flavien Chervet

October 24, 2025

On one side, Andrej Karpathy, a brilliant AI engineer, co-founder of OpenAI and now recognized as one of the best AI educators in the world. On the other, Dario Amodei, founder and CEO of Anthropic, the company behind the AI Claude, one of ChatGPT’s biggest competitors. The first believes General AI (AGI) should not arrive for another 10 years. The second insists the whole story should be settled by 2027-2028. This debate that stirs Silicon Valley reveals a fundamental problem: nobody agrees on what AGI actually is. Without a precise definition, comparing predictions is impossible.

To put an end to this confusion, 32 AI researchers and figures have proposed, for the first time, an operational, quantified and scientifically grounded definition of the concept of “General AI”¹. Among them, Dan Hendrycks of the Center for AI Safety, Dawn Song of Berkeley, Yoshua Bengio, Erik Brynjolfsson of Stanford and even Eric Schmidt. Their definition draws on a century of research in human cognitive psychology, and the first results are striking: GPT-5 would already be halfway to AGI, with a score of 57%, versus 27% for GPT-4 two years earlier.

Alright, fine. But where do these numbers come from?

Human intelligence as the yardstick

The definition proposed by the researchers is relatively simple: “AGI is an AI capable of matching or exceeding the cognitive versatility and proficiency of a well-educated adult.” Two keywords stand out: versatility (the ability to perform across many different domains) and proficiency (a level of performance comparable to humans).

To make this definition operational, the team turned to the only form of “general” intelligence we know of: our own (even though the generality of human intelligence is in fact highly debatable). More precisely, they relied on the Cattell-Horn-Carroll (CHC) theory, the most empirically validated model of human cognition to date. This theory, the product of more than a century of research in psychometrics, breaks intelligence down into a hierarchy of distinct cognitive abilities.

The approach is therefore radical in its anthropocentrism: to find out whether an AI possesses general intelligence, let us test it with the same batteries of tests used to measure human intelligence. Instead of relying on artificial benchmarks designed specifically for machines, let us evaluate AIs on the fundamental cognitive tasks that define our own intelligence.

The ten pillars of general intelligence

The framework, based on CHC theory, breaks general intelligence down into ten major cognitive components, each weighing 10% of the total score to favor breadth over specialization:

1. General knowledge: Factual understanding of the world, encompassing common sense, culture, the sciences, the social sciences and history.

2. Reading and writing: Mastery of written language, from letter recognition to the composition of complex texts.

__3. Mathematical skills __: The range of mathematical abilities, from arithmetic to infinitesimal calculus, by way of algebra and probability.

__4. Reasoning __: The ability to solve novel problems through deduction and induction, without relying solely on learned patterns.

__5. Working memory __: The capacity to hold and manipulate information in the short term, across the textual, auditory and visual modalities.

__6. Long-term memory storage __: The capacity for continuous learning, associating new information, memorizing narratives, retaining verbatim data.

__7. Long-term memory retrieval __: The fluency and precision with which stored knowledge is accessed, including the crucial ability to avoid confabulations (hallucinations).

__8. Visual processing __: The analysis, reasoning about and generation of visual information, images and videos.

__9. Auditory processing __: The discrimination and recognition of sound stimuli, including speech, rhythm and music.

__10. Speed __: The speed of execution of simple cognitive tasks (fast reading, reaction time, processing fluency).

This multimodal approach allows for a precise diagnosis of the strengths and weaknesses of current systems. Each of the 10 skills counts for 10% of the 100% of general intelligence, a score of 0% representing complete incompetence, and a score of 10% representing the equivalent of educated human intelligence on the skill in question. And contrary to what one might think, the most advanced AIs are not “slightly less good” than humans across the board. They in fact present an extremely “jagged” cognitive profile, excelling in some domains (sometimes even far surpassing humans) while being completely deficient in others.

GPT-5: a spectacular leap, but still only halfway

The results of evaluating GPT-4 and GPT-5 on this framework are telling. In two years, the “AGI score” rose from 27% to 57%, an impressive progression that does indeed mark an inflection point. GPT-5 now reaches the maximum score (10%) on several major components: general knowledge, reading and writing, mathematics, and reasoning.

GPT-5 can now solve mathematical olympiad problems that would have baffled most well-educated adults. It masters the reading of complex documents and the production of sophisticated texts. Its logical reasoning, thanks in particular to advances in reasoning models like o1 or o3, approaches human fluency (notably since o3’s breakthrough on the ARC test).

Visual processing shows significant progress (4% versus 0% for GPT-4), but remains largely insufficient. Current models struggle particularly with visual reasoning and the modeling of the physical world. On the IntPhys 2 benchmark developed by Meta, which tests intuitive understanding of physics through videos, the best current models only slightly exceed the level of chance. It is as if the AI saw the images, could analyze them and describe what is in them, but did not really understand how the world works.

Auditory processing is also progressing (6% versus 0%), but here again the capabilities remain partial. Working memory doubles (from 2% to 4%), reflecting in particular the improvements in handling long contexts and multimodal information.

Long-term memory retrieval stagnates at 4% for both models. This is largely explained by hallucinations, those moments when the model invents information with unshakable confidence. On OpenAI’s SimpleQA benchmark, GPT-5 still hallucinates on more than 30% of very specific questions. All of these capabilities, though still a bit weak, are improving quickly.

The same cannot be said, however, of long-term memory, which remains a major weakness. Both GPT-4 and GPT-5 stay at… 0%. Both models suffer from a form of “amnesia”: they are incapable of truly learning new information over time. Their neural network is trained once and for all, then remains completely static afterward. To be sure, the interfaces of today’s chatbots create the illusion of memory thanks to 2 tricks:

• Within a conversation: reinjecting the entire conversation history into the AI at each new prompt to give it contextual coherence.
• Between conversations: storing certain salient facts in a small external database (ChatGPT’s “memory” feature), which the AI can draw on to refine its answers.

But these contortions do not replace genuine continuous learning by the AI itself, which alone would guarantee its real improvement over time.

The bottlenecks on the road to AGI

Adam Khoja, co-author of the study, recently published an analysis of the remaining obstacles in an article for AI Frontiers. According to him, most of the current gaps will be closed by “business-as-usual” engineering, the usual incremental progress of AI research.

For visual reasoning, progress is already rapid. On Apple’s SPACE benchmark, GPT-4o (May 2024) scored only 43.8%, while GPT-5 (August 2025) reaches 70.8%, versus 88.9% for humans. At this rate, the gap could close within a few months.

For the modeling of the physical world, even though researchers like Yann LeCun believe more fundamental advances will be needed, the general progression of capabilities already seems to be improving performance. A structured “world model” seems to be gradually emerging thanks to current techniques. Moreover, research on the subject is very active, notably because of the popularity of video-generation AIs like Sora 2, which rely on “world models”.

Auditory processing appears to be one of the most accessible domains in the short term. Historically, audio capabilities are easier for models to acquire than visual ones (because auditory data can generally be reduced to text with less loss of information than images). If the audio models of the big AI companies remain behind, it is probably for lack of prioritization rather than intrinsic technical difficulty. The startup Sesame AI, which specializes in voice companions, has already demonstrated vocal capabilities well above the sector’s leaders.

Long-term memory, however, represents a challenge of a different nature. The architecture of the systems itself is being challenged. Here, genuine innovation will be necessary. But should we expect a real paradigm shift, like the arrival of the Transformers in 2017 that led us to LLMs such as ChatGPT, or will a simple new approach to current technologies, like the arrival of reasoning models, be enough? Adam Khoja leans toward the second option.

And indeed, the problem is not completely opaque. The big AI companies now devote considerable resources to it. In August 2025, Demis Hassabis of DeepMind identified memory as a key missing capability. Sam Altman, CEO of OpenAI, specified, speaking of GPT-6, that “People want memory.” Dario Amodei of Anthropic even detailed promising technical avenues: “We could train the model in such a way that it specializes in learning on the context. We could even, during the context, update the model’s weights… There are many ideas very close to current ideas that could perhaps accomplish continuous learning.”

2028 or 2030: AGI closer than we think?

Aggregating all the capabilities, Adam Khoja estimates a 50% probability of reaching an AGI score above 95% by the end of 2028, and 80% by the end of 2030. These estimates are therefore closer to Amodei’s predictions than to Karpathy’s.

These estimates are explained by the nature of the remaining obstacles. If we had to completely reinvent the foundations of AI to make progress, the timelines would indeed be long. But if the main bottlenecks are a matter of incremental engineering and a targeted breakthrough on memory, then the time window narrows considerably.

The framework proposed by this team of researchers will not settle every debate. One can question the weighting of each component, the exhaustiveness of the tests, or the relevance of taking the well-educated adult as a reference. I myself consider that taking the human as the measure of the machine is not the most fruitful method, and that it is more pressing and more coherent with current technologies to study the question of superintelligences very different from humans. But this work has the merit of replacing vague speculation with a quantitative diagnosis. It makes it possible to move from “AGI is coming soon (or not)” to “here is precisely where we stand, here is what is missing, and here is how fast we are progressing.”

This clarification is not merely academic. It changes the nature of the conversations about AI governance, about the investments needed, and about societal preparations. If AGI is indeed within reach in 3 years rather than in 10 years, then the urgency of rethinking our education systems, our labor markets, and our regulatory frameworks becomes glaring.

We are halfway there. Not on a linear journey, but in a construction process whose architecture we are only beginning to understand. The coming years will tell us whether the missing pieces of the puzzle fall into place as quickly as some predict. We now have a measuring instrument to track the path, mile after mile, toward what could be the most profound technological transformation in human history.

Footnotes

1. http://agidefinition.ai/ ↩