AI · Cognition
A conversation about comprehension, consciousness, and the exact moment machines started understanding faster than humans. From the N400 window in your brain to the prefill phase in a transformer — where we still hold the edge, and how long that lasts.
July 2026
I asked about TTFT — time to first token. A metric in LLM inference. The milliseconds between submitting a prompt and receiving the first output token. The prefill phase. The attention computation. The answer was clean, technical, and completely wrong for what I actually wanted to know.
What I meant was time to first thought. As in, the human one. How long does it take a person to comprehend what they are reading — even the quickest readers? How fast does understanding arrive? And at what point did the machine cross that threshold?
The answer to the first question is unsettling. The answer to the second is closer than you think.
A human reading a word does not wait for conscious thought to begin comprehension. The brain is already several steps ahead.
| Time | Event | Conscious? |
|---|---|---|
| 0–50ms | Photons hit retina, signal travels optic nerve | No |
| 50–100ms | V1 registers edges, shapes, letterforms | No |
| 100–150ms | Visual word form area recognizes pattern as text | No |
| 150–250ms | Lexical access — word meaning retrieved from memory (N200) | No |
| 250–400ms | Semantic integration — coherence check against context (N400) | No |
| 400–800ms | Anomaly detection, syntactic re-analysis (P600) | Borderline |
| 800ms–2s | First conscious thought forms | Yes |
| 2s+ | Deliberate reasoning, inference, critique | Yes |
The N400 component is the most revealing. If you read "I take my coffee with cream and socks," your brain generates a massive N400 spike — electrical activity detectable on the scalp — before you are consciously aware that anything is wrong. Your brain has already checked the semantic fit, found it broken, and flagged the error. All of this happens before you know you have understood anything.
The historical timeline of comprehension speed is worth mapping precisely, because the crossing happened recently and quietly.
2020 — GPT-3. A short prompt took several seconds of prefill on the hardware of the day. A human could read and understand the same prompt in under a second. Human wins clearly.
Late 2022 — ChatGPT (GPT-3.5) with optimized inference. A short prompt processed in 500ms to 1 second. Roughly competitive with human semantic integration time. Close to parity, but not clearly faster.
Late 2023 — Mistral 7B with FlashAttention on an A100. A 50-token prompt prefills in ~30–60ms. That is faster than your retina-to-V1 visual processing. Faster than the N400 component. Faster than your brain's unconscious lexical access. The machine has understood the literal meaning of your words before your own brain has even recognized them as words.
2024–2025 — H100 clusters, FlashAttention-2/3, speculative decoding. A short prompt on a modern serving stack: 10–30ms TTFT. A 1000-token document: 100–300ms. A human reading 1000 tokens at deep-reading speed (~250 words per minute) takes over four minutes.
The machine crossed the human threshold somewhere between late 2022 and late 2023, depending on prompt length. For short prompts, it happened when inference optimization reached the point where a model could prefill faster than ~400ms — the time your brain takes to unconsciously integrate semantic meaning.
Consider a 5000-token prompt — a short article of about 3500 words.
A human at deep-reading speed: ~4–5 tokens per second. Simple comprehension of the full article: ~17 minutes. Deep comprehension with re-reading, cross-referencing, and note-taking: 30 minutes or more.
A machine on an H100 in 2025: prefill time of ~500ms to 1.5 seconds.
The machine has absorbed and attended to every word in relation to every other word — the full self-attention matrix — in the time it takes your eyes to move across the first line of text. By the time you have read the first sentence, the machine has already processed the entire article, generated its first response token, and is working through the second paragraph of its reply.
Speed is not depth. The human who reads the article has lived experience against which to measure it. The human has a model of the world trained on thirteen billion years of causal physics, not fifteen trillion tokens of text. And the human can do something the machine cannot: tell you what matters.
A machine trained on text has no model of leverage. It can identify every factual error in a document, but it cannot tell you which error is the one that will bring down an empire. Given 10,000 documents, it will summarize them all equally well. The human can read the same stack and say: "This footnote on page 312 is the one that changes everything."
This is not a skill that can be taught through more data. It is a function of having a model of the world that includes stakes, consequences, and the relative importance of things. The machine has read everything. It can tell you what everything says. But it does not know what matters, because nothing actually matters to it.
Machines have gotten better at detecting sarcasm through training data. They can classify irony with reasonable accuracy. But they do not feel it.
Irony depends on a shared model of what is expected and a deliberate deviation from it for effect. The human who reads "Great, another mandatory meeting that could have been an email" understands not just the literal words but the accumulated weight of a thousand such meetings — the resignation, the shared understanding that everyone in the room feels the same way. The machine can classify it as sarcasm. It cannot share the sigh.
The same gap applies to the unsaid. A machine processes only what is present in the text. A human detects the absence — the thing the author is avoiding, the argument that is conspicuously not made, the emotion that is being suppressed. The machine reads the silence as nothing. The human knows the silence is the whole point.
When the training data runs out — when the problem has no precedent, no text corpus, no existing answers — the machine falls back on statistical interpolation. It guesses. It hallucinates a plausible-sounding answer. It has no off-switch for its own confidence.
The human, at their best, reasons from first principles. They can say "I don't know, but here is how I would find out" — and then build the knowledge from scratch. The machine cannot admit it does not know. More fundamentally, it does not know that knowing what you do not know matters.
A machine can describe grief with perfect fidelity — the stages, the psychology, the literary history, the neurochemistry. It has seen every word ever written about grief. It can write a poem about grief that passes a Turing test. But the machine has never been surprised by grief. Never had grief arrive differently than expected. Never realized, in the middle of it, that all the books were right and somehow completely wrong at the same time.
Machine Wins
Human Wins
Every gap I have described is real today. But the architectures are changing. Embodied models exist. World models that learn from sensory data rather than text are in development. The human advantage, if it turns out to be permanent, will not be about data or compute. It will be about something the machine cannot replicate because it does not have the hardware for it.
Consciousness. The subjective experience of being.
That is not a question we have answered yet. And it is the only question that matters for the long game. If consciousness is necessary for the kind of understanding that humans have — the kind that involves knowing what matters, feeling irony, detecting the unsaid — then the machine will remain a savant forever. Brilliant, tireless, and ultimately blind to the things that give comprehension its weight.
But if consciousness is an emergent property of sufficient complexity — if it is not a special substance but a phase transition in information processing — then the gap is not permanent. It is just a lead, and the machine is closing.
The machine is catching up on the map. The territory still belongs to those who live in it. For now.