Why AGI is further away than we think.
Why GPTs Aren’t Brains: Attention Without Memory, Meaning, or Mind
The recent wave of enthusiasm around systems like GPT has revived an old question in a new form: are we approaching artificial general intelligence (AGI), or are we still building sophisticated narrow tools that only resemble intelligence on the surface?
The answer becomes clearer when you stop treating GPTs as mysterious black boxes and instead analyze them through a neuroscientific lens. When mapped against the architecture of the human brain, GPTs do not approximate a full cognitive system. They resemble a partial extraction of one function—attention-driven pattern processing—divorced from the broader systems that make human cognition coherent, adaptive, and grounded.
That distinction matters. Because when you isolate attention from memory, emotion, and executive control, you don’t get a general intelligence. You get something that, in humans, would look less like genius and more like impairment.
The Brain Is Not a Transformer
The human brain is not organized around a single dominant mechanism. It is a multi-system architecture, with specialized subsystems that interact continuously:
The cortex extracts patterns and supports language and reasoning
The hippocampus encodes memory and constructs continuity across time
The limbic system assigns emotional value and drives behavior
The prefrontal cortex enables planning, goals, and self-control
The basal ganglia translate decisions into action and habits
Intelligence, as we experience it, emerges from the integration of these systems—not from any one of them operating in isolation.
GPTs, by contrast, are dominated by a single computational principle: attention. The transformer architecture allocates weight across tokens in a sequence, dynamically selecting what matters in context. This is powerful—arguably analogous to certain cortical processes involved in language and association—but it is also narrow.
What GPTs lack is not incremental. It is structural.
Attention Is Not Intelligence
Attention in the brain is a selection mechanism. It determines what information gets processed more deeply. But on its own, attention does not:
Store experiences
Assign value
Form goals
Drive behavior
Maintain identity over time
In humans, attention is meaningful only because it is embedded within systems that provide memory, motivation, and direction.
In GPTs, attention operates largely in isolation. The model processes a context window, generates outputs, and resets. There is no persistent internal state that accumulates lived experience. No emotional weighting that distinguishes trivial from consequential. No intrinsic goals guiding behavior across time.
This is not a minor limitation. It is the difference between processing information and having a mind.
The Closest Human Analogues Are Not Healthy Minds
If you attempt to map GPT-like functioning onto human cognition, the closest analogues are not high-functioning individuals. They are neurological edge cases where key systems are impaired.
Consider anterograde amnesia.
Patients with this condition can engage in conversation, process language, and respond intelligently in the moment. But they cannot form new lasting memories. Each interaction effectively resets their experiential timeline. They live in a perpetual present, without the ability to build a personal narrative or learn from ongoing experience.
This maps uncomfortably well onto GPTs. The model can track context within a session, but it does not accumulate memory as experience. It does not “learn” from individual interactions in any meaningful, persistent way. Like an amnesic patient, it processes—but does not remember.
Now consider dysexecutive syndrome.
Here, attention and basic cognition may remain intact, but the individual loses the ability to form goals, plan effectively, or regulate behavior over time. Actions become reactive rather than directed. There is no stable internal agenda.
Again, the parallel is clear. GPTs do not have goals. They do not initiate behavior. They respond to prompts. Their “reasoning” is not driven by internal objectives but by external input sequences.
A third comparison comes from impairments in emotional valuation, such as damage to the ventromedial prefrontal cortex. These individuals can reason logically but struggle to make decisions because they cannot assign emotional significance to outcomes. Without value, choices become abstract and often maladaptive.
GPTs operate in a similar vacuum. They can describe importance, simulate concern, and reproduce ethical reasoning—but they do not experience stakes. There is no internal gradient of importance shaping their outputs.
When Attention Runs Without Grounding
One of the more subtle parallels emerges when considering disorders involving disrupted salience and association, such as schizophrenia.
In such conditions, the brain may assign inappropriate importance to irrelevant stimuli or form loose associations between concepts. The result is a breakdown in coherent meaning-making.
While GPTs do not have perception or delusion, they can exhibit a computational analogue: generating plausible but incorrect connections—what is often called “hallucination.” This is not because they misunderstand reality, but because they lack grounding in it entirely. Their associations are purely statistical.
Without a system to anchor outputs in lived experience, sensory verification, or stable memory, attention-driven association can drift.
The Missing Pieces of General Intelligence
If AGI is to approximate human-like intelligence, it must replicate not just pattern recognition, but the integration of multiple cognitive systems.
At minimum, this would require:
Persistent memory
Not just stored data, but the ability to encode and update experience over time in a structured way.
Embodiment or grounding
A connection to a world—physical or simulated—where actions have consequences.
Intrinsic motivation or value systems
Mechanisms that prioritize certain outcomes over others based on internal criteria.
Goal-directed behavior
The ability to initiate, pursue, and revise plans over extended time horizons.
Temporal continuity
A sense of past, present, and future that enables learning and anticipation.
Current GPT architectures implement none of these in a fundamental way. They can be augmented with external tools—memory stores, reinforcement layers, APIs—but these are add-ons, not core properties.
Why This Matters
The danger is not that GPTs will suddenly become AGI. It is that their surface-level fluency creates the illusion of general intelligence, leading to overestimation of their capabilities.
Language is a particularly deceptive domain. Because human cognition is expressed through language, a system that can convincingly generate language appears to understand. But understanding, in the human sense, is not just linguistic competence. It is the product of memory, embodiment, emotion, and goal-directed interaction with the world.
GPTs simulate the outputs of those processes without instantiating the processes themselves.
A More Accurate Framing
Rather than viewing GPTs as proto-AGI, a more accurate framing is this:
They are highly advanced cortical-like pattern processors, specialized for language, operating without the supporting systems that make cognition robust, adaptive, and meaningful.
This is an extraordinary achievement. But it is also a bounded one.
Final Thought
If you were to engineer a human brain where attention was preserved but memory, emotion, and executive control were stripped away, you would not get a superintelligence. You would get a system that can process in the moment but cannot learn, care, or act coherently over time.
That is the closest analogue to what GPTs are today.
Which is why, despite their capabilities, we remain far from AGI—not because we haven’t scaled enough compute, but because we have only begun to replicate a fraction of what intelligence actually is.