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 insufficient for AGI under any autonomy‑based definition, 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. AGI here means a system that can (1) form non‑trivial goals from observation, (2) plan and execute over days/weeks, (3) act via tools without stepwise prompting, and (4) update beliefs/strategies from outcomes.

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 probabilistic next‑token generator. The critique popularized by Bender et al. (2021) described such systems as “stochastic parrots” that stitch together linguistic forms according to statistics rather than grounded meaning (FAccT 2021). 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. 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?” Serious evaluations say otherwise. In a large public online Turing‑test study, the best GPT‑4 prompting was judged human in 49.7% of games, while humans scored 66%. Participants’ judgments leaned heavily on stylistic and socio‑emotional cues rather than deep reasoning, which is precisely why imitation ≠ intelligence (NAACL 2024, pdf).

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 hallucination, where the model confidently fabricates non‑existent facts, citations, or steps in reasoning. 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 reasoning tethered (mostly) to reality.

On long inputs, models frequently underuse or ignore distant information; performance is best when relevant information appears at the beginning or end, and degrades in the middle (“Lost in the Middle”). Extending context windows helps but does not resolve reasoning. Experiments show that “simultaneously finding relevant information in a long context and conducting reasoning is nearly impossible” for current LLMs; effective context is far smaller than the raw window size (AI‑native Memory, 2024).

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. The 2023 AI Impacts survey aggregates a median 50% estimate for high‑level machine intelligence around 2047, with substantial dispersion (AI Impacts 2023, report pdf). Meanwhile, many researchers highlight architectural gaps beyond today’s transformers for long‑horizon, interactive intelligence.

A good case study is the pursuit of ever‑larger context windows. Some argue that if we can extend context to millions of tokens, we get AGI. Reality: effective context length is much smaller; long‑context retrieval and reasoning together remain fragile (arXiv 2406.18312). In short, more tokens alone don’t create deep understanding or planning ability.

Similarly, the idea that just scaling model size will inevitably produce AGI is on shaky ground. Scaling has yielded impressive competence but hasn’t produced a reliable, general reasoning module or self‑directed goal formation. Calls from leading researchers emphasize the need for new architectures with world models, planning, memory, and intrinsic objectives rather than ever‑bigger next‑token predictors (LeCun, “A Path Towards Autonomous Machine Intelligence”).

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 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. Provable deletion of external memories is feasible; erasing information from a model’s weights remains an open, imperfect “targeted unlearning” problem.
• Grounded World Modeling: 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 internal predictive models of dynamics to plan and simulate outcomes. Position papers outline architectures with predictive world models, hierarchical planning, and intrinsic motivation (LeCun 2022).
• Reasoning and Planning Abilities: Human intelligence isn’t just knowledge; it’s the ability to manipulate that knowledge through reasoning. Evidence to date shows LLMs on their own are unreliable planners; better results come from LLM‑modulo setups where symbolic planners/verifiers are in the loop (Valmeekam et al., NeurIPS 2023; Kambhampati et al., 2024 (position)).
• Causal Inference and Understanding of Reality: General intelligence requires learning causes, not just correlations. That implies active experimentation (in sim or real), multimodal grounding, and robust model‑based planning – areas still immature in mainstream LLMs.
• Autonomy and Intentionality: Tool‑using agent wrappers increase usefulness but don’t create intrinsic goals. True agency needs stable cognitive loops for setting subgoals, monitoring progress, and adapting strategies over long horizons, not just next‑prompt continuation (see planning critiques above: NeurIPS 2023; 2024 position).

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.” (Stanford GSB). 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. 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, highly speculative relative to today’s capabilities.

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.

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 powerful probabilistic programs with broad competence but no self‑directed agency. 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.

This perspective is not intended to diminish the significance of today’s AI. The current generation of models and the surrounding ecosystem of tools are extraordinary achievements in their own right. Over the next 10-20 years we can expect continued, steady progress: better memory systems, more capable world models, tighter multimodal integration, and improved reasoning frameworks. These advances will make AI more useful, reliable, and deeply embedded in daily life well before anything resembling AGI arrives. Recognizing this trajectory allows us to celebrate what exists today while keeping expectations about timelines realistic.

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.

Related post: AGI Won’t Explode – It’ll Crawl (And We’ll Still Have Time to Pull the Plug)