AGI Is Not Around the Corner: Why Today’s LLMs Aren’t True Intelligence

By Christian Holmgreen, Founder of Epium.

The emergence of large language models (LLMs) like GPT-4, Anthropic’s Claude, and Google’s Gemini has sparked a wave of hype that artificial general intelligence (AGI) is just about to dawn. Some observers claim these models are already passing Turing Tests, that we’re only 2–3 years away from human-level AI, or that simply scaling up current systems will magically yield true general intelligence. Such claims make for sensational headlines – but they don’t hold up under a sober examination of the technology. In reality, today’s AI systems are nowhere near AGI, and their incremental improvements don’t indicate exponential progress toward it. This post takes a rationally skeptical look at the state of AI, explaining clearly why current LLMs are not AGI, what critical ingredients are missing, and why doomsday fears of an imminent superintelligent takeover (a la SkyNet) are premature and unsupported by the facts.

What Is AGI, and Why LLMs Don’t Qualify

To start, it’s important to clarify what we mean by artificial general intelligence. AGI refers to an AI with human-level cognitive abilities across a broad range of tasks – not just chatting or writing code, but understanding the world, learning new skills on the fly, devising plans to achieve goals, and adapting to unforeseen changes. In other words, an AGI would demonstrate autonomous reasoning, deep understanding, and agency in a manner comparable to a human mind.

Current LLMs, by contrast, are specialized pattern recognizers. They are brilliant at one narrow trick: given a prompt, they predict likely sequences of text based on billions of examples. This yields impressively human-like prose and answers. But under the hood, an LLM is essentially “a system for haphazardly stitching together sequences of linguistic forms it has observed in its vast training data, according to probabilistic information about how they combine, but without any reference to meaning: a stochastic parrot” (quoteinvestigator.com). In plainer terms, these models mimic the form of human language without truly understanding the content. They lack the grounded comprehension of concepts and context that humans (and any true AGI) possess.

Modern LLMs are not conscious, sentient, or agentic. They do not formulate goals of their own or pursue objectives over time. They only respond to prompts given by users or environments. They have no internal drive or intentions – no more than a calculator “wants” to solve equations. As one observer bluntly put it on Hacker News: “There is no mechanism by which LLMs have agency. They have no internal desires, drives, [or] motivations.” The apparent cleverness and conversational ability of models like GPT-4 can easily be mistaken for general intelligence, but it is surface-level competence. Indeed, researchers note that the cognitive abilities of state-of-the-art LLMs are still “superficial and brittle”, and generic LLMs remain “severely limited in their generalist capabilities.” Fundamental prerequisites like embodiment, real-world grounding, causality and memory are “required to be addressed for LLMs to attain human-level general intelligence.” (arxiv.org) In short, current models are powerful tools, not thinking entities.

LLMs Imitate Intelligence – They Don’t Truly Understand

One common point of hype is: “Well, these AI models are already passing the Turing Test, so haven’t they essentially achieved human-like intelligence?” It’s true that in superficial interactions, advanced chatbots can fool people. For example, in one recent study, participants chatted with either a human, the old ELIZA program, GPT-3.5, or GPT-4, and then guessed which was human. GPT-4 was judged to be human 54% of the time (just over the 50% threshold for a classic Turing Test pass), whereas the actual human was identified correctly 67% of the time (livescience.com). On the surface, this sounds like a milestone – GPT-4 “beat” the Turing Test. But what does that really indicate?

First, the Turing Test itself is a limited and arguably outdated benchmark. Even the researchers in that study cautioned that the test is “too simplistic” and that “stylistic and socio-emotional factors play a larger role in passing the Turing test than traditional notions of intelligence.” (livescience.com) In other words, an AI can fool people by mimicking conversational style and emotional cues, without possessing any deep understanding or general reasoning ability. Passing a short chat-based imitation game is not equivalent to possessing human-level intellect. As AI expert Gary Marcus quipped, the Turing Test is more a measure of human gullibility than of machine intelligence.

Indeed, GPT-4’s conversational eloquence belies significant gaps in comprehension. LLMs do not know what they are talking about – they lack a grounded model of the world that gives meaning to the words. They often make errors that no knowledgeable human would, precisely because they have no real understanding. A striking example is the phenomenon of hallucinations, where the model confidently fabricates non-existent facts, citations, or steps in reasoning. ChatGPT or Claude might tell you a very plausible-sounding but completely false biography of a person, or assert that 2+2=5 if prompted in a tricky way, simply because it’s statistically permissible in its training distribution. These failures underscore that LLMs have zero concept of “truth” or “reality” beyond patterns of text. By contrast, a human (or a true AGI) builds an internal model of the world through perception and experience, which keeps our reasoning tethered (mostly) to reality. Current AIs have no perception – they’re trapped in the world of words.

In essence, today’s models excel at imitation – they can sound like an expert, but they are not actually experts. As another analysis succinctly summarized: “These are four essential characteristics of human intelligence – reasoning, planning, persistent memory, and understanding the physical world – that current AI systems can’t do.” (medium.com) They can simulate these abilities in limited contexts, but they do not genuinely possess them. No matter how fluent or knowledgeable an LLM seems, its intelligence is narrow and shallow. It cannot reliably reason through novel multi-step problems that require understanding causality or physical dynamics. It cannot plan a sequence of real-world actions to achieve a goal. It cannot remember new information over time in the way a person can learn continually. And it certainly has no common-sense grasp of the physical environment – something as simple as knowing that putting a turkey in a refrigerator overnight (not for five minutes) is how you thaw it safely, or that if you drop a glass it will likely shatter. The “common sense” databases of humans are built from lifetime experience; LLMs have none of that.

To illustrate, NYU professor Yann LeCun points out that even young children display a kind of understanding that eludes AI. A 10-year-old can learn how to clear a dinner table after seeing it done once, and a teenager can learn to drive a car after a few hours of practice. Meanwhile, “even the world’s most advanced AI systems today, built on thousands or millions of hours of data, can’t reliably operate in the physical world.” (techcrunch.com) The AI can describe driving or answer questions about it, but you wouldn’t trust GPT-4 to take the wheel or even load a dishwasher. Why? Because these models lack a “world model” – a rich, internal representation of how the physical world works that would let them simulate outcomes and transfer knowledge between domains. LeCun emphasizes that current language models “don’t really understand the three-dimensional world” at all (techcrunch.com). They are stuck predicting text, whereas human-like intelligence requires perceiving, modeling, and interacting with a 3D environment.

The bottom line is that imitating conversational intelligence is not the same as possessing general intelligence. LLMs pass shallow tests by aping the style of human responses, not by replicating the underlying cognitive machinery that produces those responses. We must be careful not to confuse a clever illusion for the real thing.

Incremental Advances ≠ Imminent AGI

Another popular refrain from the hype machine is: “We’re just a couple years away from AGI. Look at how much smarter each new model is – AI is improving at an exponential rate!” It’s true that AI capabilities, especially in language, have advanced rapidly in recent years. But there is little evidence that we are on the cusp of general intelligence, and plenty of reasons to think it’s still far off. The year-over-year gains are real but modest – more evolutionary than revolutionary – and some metrics even show signs of leveling off.

It’s worth noting that expert opinions on AGI timelines vary widely, but most are not as short-term as the hype suggests. Recent surveys of AI researchers generally put the 50% likelihood of achieving AGI around 2040-2050, if at all (research.aimultiple.com) (research.aimultiple.com). For example, an October 2023 survey of 2,700+ AI experts gave a median estimate of 2040 for “high-level machine intelligence” (research.aimultiple.com). Importantly, a 2024 AAAI expert panel found that 76% of respondents believe that simply scaling up current AI approaches is unlikely to lead to AGI (research.aimultiple.com). In other words, most experts don’t think we’ll get to human-level AI just by making today’s models bigger or training on more data. Yet the AI hype cycle tends to gloss over this consensus, with bold predictions that “AGI will be here by 2028” or that GPT-5 or GPT-6 will spontaneously become self-aware.

If you look at the actual progress from GPT-3 to GPT-4, or from early Claude to Claude 2, the improvements – while impressive – do not represent a fundamentally new kind of intelligence. They are mostly quantitative (more parameters, more data, better fine-tuning) with some incremental qualitative advances (longer memory window, better alignment to instructions). We are not seeing the emergence of a unified reasoning architecture or a stable of AI that can set its own goals. Each new model is still an auto-complete savant, just with more knowledge and some refinement. As one detailed analysis put it, the gains from scaling are slowing down: LLMs have “indeed reached a point of diminishing returns.” Simply put, bigger models are yielding smaller improvements now, not exponential leaps.

A good case study is the pursuit of ever-larger context windows (the amount of text the model can consider at once). Some startups have argued that if we can just extend an LLM’s context to, say, millions of tokens (so it can “read” and “remember” book-length or even Wikipedia-scale knowledge in one go), then boom, we’d have an AGI. Unfortunately, research shows this to be overly optimistic. For one, effective context length is much smaller than the raw window size – models don’t actually utilize super-long contexts efficiently, often forgetting or ignoring most of the distant information (arxiv.org). And crucially, experiments find that “simultaneously finding relevant information in a long context and conducting reasoning is nearly impossible” for current LLMs (arxiv.org). In other words, even if you give GPT-4 a 500-page novel as input, it struggles to both pick out the key bits and reason about them coherently. Long context alone doesn’t fix the fundamental limitations in reasoning. More data or memory can help up to a point, but it doesn’t automatically create deep understanding or planning ability.

Similarly, the idea that just scaling model size will inevitably produce AGI – sometimes called the “scaling hypothesis” – is on shaky ground. Yes, larger models have shown emergent abilities (GPT-4 can do things GPT-2 couldn’t dream of). However, we have not seen emergence of a reliable reasoning module, or a jump from narrow to general competency. If anything, the field is grappling with diminishing returns and the economic and energy feasibility of scaling further. Training GPT-4 reportedly cost over $100 million – and yielded an AI that still makes dumb mistakes and has no genuine self-directed learning. It’s not clear that GPT-5 (if it’s just bigger) would be worth, say, $1 billion to train for a marginal improvement. As a majority of surveyed experts agreed, something more than scaling is needed to break through to AGI (research.aimultiple.com).

In short, there is no clear exponential trend toward AGI right now. What we see instead are gradual improvements and a lot of over-interpretation of those improvements as “signs of AGI.” The hype cycle has a way of compressing timelines – four years ago, some predicted AGI in 2050; now suddenly the same folks say 2030 or sooner, simply because GPT-4 is cool. But cool is not the same as ground-breaking. It’s worth remembering that AI history is littered with premature proclamations (the 1970s saw people claiming general AI was 10 years away, then came the AI winters). Caution and humility are warranted. As LeCun said in late 2024, humanity could be “years to decades away” from human-level AI (techcrunch.com) – we simply don’t know, but there’s no strong evidence it’s just 2–3 years out.

What’s Missing: The Path to True AGI

If current LLMs aren’t AGI and won’t magically become so just by getting larger, what would it take to reach true general intelligence? Researchers and skeptics alike point to several key ingredients that are missing from today’s AI. Attaining AGI will likely require significant architectural breakthroughs and new ideas, not just more of the same. Here are some of the critical capabilities and research directions that could pave the way to AGI:

• Integrated, Long-Term Memory: Humans don’t forget a conversation as soon as it ends; we build knowledge over our lifetime. By contrast, an LLM has a short memory (limited context window) and no persistent internal storage of new knowledge. To become generally intelligent, an AI needs a form of structured memory – a way to store, organize, and recall information it has learned across time. This might involve neural architectures that can accumulate information (beyond just compressing it into billions of weights), or hybrid systems that connect LLMs to external knowledge bases in a deeply integrated way. Recent work on “AI-native memory” envisions systems where an LLM is the core, surrounded by a memory store of facts and conclusions derived from reasoning, which can be updated continually (arxiv.org). Such a memory would let the AI learn from experience rather than being frozen at the time of its training. Some prototypes (like retrieval-augmented generation) bolt on databases to LLMs, but these are early steps. True AGI might require memory that is as richly structured and efficiently accessible as human episodic and semantic memory.

• Grounded World Modeling: As discussed, current AIs lack a world model – they do not truly grasp how the world works because they only ingest text (or images) but don’t experience or simulate the physical environment. A crucial research frontier is giving AI models a form of embodied understanding or at least a detailed simulation of the world’s physics and causality. One approach is the development of world models in AI – systems that can predict the consequences of actions in an environment. LeCun and others argue that we need AIs that can imagine and plan in a 3D space, not just predict words (techcrunch.com). For example, an AI with a world model could look at a messy room and plan a sequence of actions to clean it up – all in simulation – before acting (techcrunch.com). This capability is far beyond current AI, but it might be essential for AGI. It implies a combination of perception, physics understanding, and goal-driven simulation. We see early attempts in robotics and game AI, but no current language model has such a built-in world model. Bridging this gap may involve multi-modal learning (combining vision, proprioception, etc. with language) so the AI can connect words to things and events.

• Reasoning and Planning Abilities: Human intelligence isn’t just knowledge; it’s the ability to manipulate that knowledge through reasoning. LLMs can do a form of reasoning by generating step-by-step “chain-of-thought” text, but they lack a reliable, general reasoning faculty. They don’t have an internal scratch pad for logic, or an ability to plan multi-step tasks under uncertainty. Many experts believe we need new architectures (or hybrids with classical AI) to handle symbolic reasoning, long-range planning, and complex problem solving. For instance, one might integrate a reasoning module that carries out logical inferences or mathematical computations alongside the neural network’s predictions. Some research is exploring neuro-symbolic methods or the incorporation of explicit planning algorithms guided by language models. To reach AGI, an AI should be able to formulate subgoals, make decisions, and adjust plans when circumstances change – essentially, perform the kind of mental simulation and reasoning that humans do when solving novel problems. Current LLMs do this weakly at best. As a result, they struggle with tasks like puzzles that require consistent application of rules or multi-stage derivations. Architectural innovation (not just scale) is likely needed to imbue AI with robust reasoning skills.

• Causal Inference and Understanding of Reality: Related to reasoning is the concept of causality – knowing why things happen, not just what tends to co-occur. A true AGI must learn not just patterns (correlation) but principles (cause and effect). For example, it’s one thing to read millions of sentences about illnesses and symptoms; it’s another to deduce which symptom is caused by which illness. Humans learn causal relationships through experimentation and observation. For AI, acquiring this might require new training paradigms (like actively querying the world, running experiments in a simulated environment, etc.). Current AI largely learns correlations from static datasets. Moving toward AGI might involve techniques for causal learning, so the AI can form hypotheses (e.g., “X causes Y”) and test them, or at least recognize the difference between correlation and causation. This ties into the need for interaction – an AI that can only passively read data may never untangle causal structures fully. Some researchers suggest that embodied AI (like robots or agents in simulated worlds) could gain causal understanding by doing things and seeing outcomes, much as children do.

• Autonomy and Intentionality: Finally, one of the hallmarks of general intelligence is agency – the ability to set goals, pursue them proactively, and adapt strategies to achieve them. Present-day

•  LLMs are reactive. They do what you ask but otherwise sit idle. They have no intrinsic goals or persistence. An AGI, by contrast, would be able to initiate behaviors on its own to satisfy its objectives (which might be given or learned). This doesn’t mean it has to have human-like “desires,” but it does mean operating more like an agent than a tool. Researchers are experimenting with agentic wrappers around LLMs (e.g. AutoGPT, BabyAGI) that loop the model’s outputs to create a form of goal-directed behavior. However, those early attempts have shown just how far there is to go – Auto-GPT, for instance, often gets stuck in loops, makes silly mistakes, or fails to accomplish modest tasks due to the model’s limitations in planning and memory. An analysis of Auto-GPT noted that its “limitations in reasoning capabilities…and the early-stage development of agent mechanisms reveal that it is far from being a practical solution.” (jina.ai) It was a proof-of-concept, not a proto-AGI. True agency in AI will require more stable and reliable cognitive loops that keep the AI on track toward a goal, allow it to break down tasks, learn from intermediate results, and avoid getting confused. Achieving this might involve combining learning-based approaches with classic AI planning algorithms, and ensuring the AI can learn from its experiences (memory) as it operates. Until AIs can robustly pursue goals over long periods, they remain well short of general intelligence.

These are just some of the research frontiers that many believe are essential for AGI. The overarching theme is that new ideas are needed – simply increasing token windows or parameter counts on the current transformer models is not likely to spontaneously generate these capabilities. We may need hybrid architectures (for example, an LLM + a database + a logic engine + a reinforcement learning module, all integrated), or something entirely new that breaks the paradigm of today’s neural networks. The human brain is still far more complex and dynamic than our AI models; it integrates memory, perception, action, and learning in a unified system with extraordinary efficiency. Our AI models are modest by comparison, and we shouldn’t be surprised that they can’t do all the brain can do. To get to AGI by the 2035–2045 timeframe (a reasonable guess by many experts), significant scientific breakthroughs will have to occur. It’s not just an engineering problem of “more GPUs!” but a scientific problem of “how do we make machines that learn and think like humans (or even animals)?”

Don’t Fear SkyNet: Why AI Doomerism Is Premature

With the hype about imminent AGI often comes the counterpart fear: if AGI is just around the corner, does that mean a superintelligent AI will soon pose an existential threat to humanity? This is the narrative of countless science fiction plots and the alarm of some high-profile doomsayers who invoke images of SkyNet or a rogue AI turning against us. It’s important to address this, because fear-mongering about AI can be just as misguided as overhyping its capabilities.

The reality is that today’s AI is nowhere near posing an existential risk. You’ve seen why current systems are not general intelligences. They are also fully dependent on human operators – a GPT-4 cannot do anything in the world unless a person or program uses its output to take actions. It has no will of its own. Even the most “agentic” AI systems today (like experimental autonomous agents) are brittle and easily confused. They don’t suddenly gain survival instincts or a lust for power. The nightmare scenario of an AI that “decides” to harm humans presupposes an AI with a high degree of independent goal-seeking, strategic planning, and self-preservation instincts. No such AI exists, even in rudimentary form. And as argued above, we’re likely years or decades away from any system that could qualify as generally intelligent, let alone superintelligent and conniving.

Prominent AI scientists have noted that these apocalyptic fears are distracting and often purely speculative. Andrew Ng famously quipped, “Worrying about evil AI killer robots today is a little bit like worrying about overpopulation on Mars.” (deloitte.com) In other words, it’s a problem to consider in the abstract for the future, perhaps, but it’s not a real or present danger. Likewise, discussions of “AI uprising” or comparisons to Terminator scenarios “often possess little to no basis in truth.” (deloitte.com) Yes, it’s good to be aware of long-term risks and to build AI responsibly with safety in mind. But there is a big difference between acknowledging theoretical future risks and claiming that we’re on the brink of an AI-induced apocalypse in the next year or two. The latter is not grounded in technological reality. It tends to be based on assuming far more capability than AIs actually have.

In fact, an excessive focus on doomsday scenarios can be counterproductive. It can lead to public panic, misguided regulation, or even a kind of fatalism that “AGI will inevitably destroy us.” Instead, our stance should be one of proactive but rational management: yes, let’s invest in AI safety research, work on alignment (making sure future AIs follow human-aligned goals), and think about how to contain or collaborate with an AGI if one is created. But let’s also keep in mind that we have time to get this right. The first true AGI is not going to spontaneously appear overnight from a chatbot; it will be the result of a long research process, which gives us the opportunity to shape its development. As of 2025, the sky is not falling – no AI has volition or the extreme capabilities required to pose an existential threat. So any talk of near-term “AI extinction risk” is, frankly, science fiction at this point.

This isn’t to say AI can’t do any harm today – it certainly can, but of a different kind (bias, misinformation, cyber attacks by automated systems, etc.). Those are real issues to tackle without invoking AGI doomsday. We should separate the far-fetched fears from the practical challenges. When someone says “What if the AI becomes evil and kills us all next year?”, the best answer is: that reflects a misunderstanding of the technology. Before an AI could pose that kind of threat, it would have to achieve a level of competence and independence that is far beyond the current state of the art. And if we ever do get close to such powerful AI, it won’t be a surprise – we’ll see it coming through the gradual progress and we can engineer safety measures alongside it.

The hyperbolic fears and “AI doomerism” also ignore an important reality: we are not powerless creators. Humans design these systems. An AGI isn’t going to magically appear and slip out of our control without a series of human decisions enabling that. So focusing on making those decisions wisely (such as implementing proper fail-safes, oversight, and global cooperation on AI safety) is far more productive than scaring ourselves with sci-fi tales. As one tech commentator noted, no matter how much an AI advances, “its AI system cannot suddenly […] decide to kill you or a specific bystander” on its own (deloitte.com). There is always a chain of cause and effect that we can monitor. Thus, while it’s good to be mindful of long-term possibilities, there’s no need to panic or halt all AI research out of fear. We’re still in the phase of making AI actually work in a general sense; world-dominating superintelligence is a problem for another day (if ever).

Conclusion: Rational Optimism for the Long Road Ahead

It’s an exciting time in AI – systems like GPT-4 and its peers are dazzling in many ways, and progress is steady. But excitement should be tempered with clear-eyed realism. We do not have AGI, and we’re not on the brink of it with current technology. Today’s models are narrow savants with glaring weaknesses in memory, reasoning, planning, and real-world interaction. They are growing smarter in small steps, not leaping into omniscience. As such, we should push back against the hype that every marginal improvement is a sign of an impending singularity. Believing that “AGI is just a few years away” can lead to disappointment or poor choices (whether it’s misallocating investments or prematurely worrying about sci-fi scenarios).

Instead, the evidence suggests a longer timeline and the need for new innovations. Perhaps AGI will grace us in the 2035–2045 period, as some optimistic experts predict, but achieving that will require solving hard scientific problems and inventing AI systems with fundamentally new capabilities. It might involve rethinking the architecture of AI from the ground up – incorporating memory, world knowledge, reasoning modules, and more, in ways we haven’t yet discovered. The journey to AGI is likely to be a marathon, not a sprint.

We should be optimistic but realistic: optimistic that humans can eventually create something as intelligent as ourselves (there’s no physical law against it, as far as we know), but realistic that it’s a long-term endeavor requiring careful research, not just throwing more data at a giant black box. In the meantime, managing our expectations and fears is key. Hype and fear both thrive on misunderstanding. The antidote is understanding: recognizing what current AI can and cannot do. When we appreciate the true limitations of LLMs, we can both be impressed by their achievements and cognizant of their shortcomings. This balanced view will help us direct our efforts where they’re needed – toward genuine breakthroughs – and approach the future of AI with rational confidence rather than misplaced hype or dread.

In summary, AGI is not imminent, but neither is it impossible. It remains a grand challenge for the coming decades. By cutting through the noise of hype, we can see the work that lies ahead and pursue it with clear vision. The road to human-level AI may be long, but understanding that prevents us from getting lost in mirages along the way. As the saying goes, “reports of the birth of AGI are greatly exaggerated” – our job now is to turn down the noise, get to work, and make progress one step at a time toward that distant goal, while using the amazing (but not yet general) AI we have in responsible and productive ways. The future of AI is bright – just not as immediate or as dire as some would have you believe.