In 2008, Ogilvy & Mather India made a one-minute film for Perfetti's Mentos
The Evolution film. A monkey, mid-stride, dragging a donkey behind him because the donkey could not keep up with what was coming next.The donkey was Daddu. A chewing gum brief, written by Abhijit Avasthi and Jignesh Maniar. Directed by Rajesh A. Krishnan. Produced by Soda Films. Animated by Frameworks Studio in Singapore. (Google search). In 2026 we may need to look again. Hold the image.
A formula
Attention(Q, K, V) = softmax(QKᵀ / √dₖ) · V. One equation. Three matrices. Four operations. You can say, a mathemtaical function. Eight researchers at Google published it in 2017 on six pages. Every claim made about artificial intelligence in 2026 traces back to that one line. Every valuation. Every roadmap. Every headline. Look at the formula again.
Is this intelligence?
Let's look at it: We are now told what this function will do.
Read every contract. Listen to every customer call. Underwrite every loan. Score every fraud signal. Triage every dispute. Decide every collections action. Govern every compliance review. Answer every customer. Approve every claim.
Companies are signing off on these. Roadmaps are built on these. The list is not wild. The list is what is happening inside enterprises, as you read.
Look inside the formula
If math is not your thing, skip ahead to the next section. You will lose nothing. In an earlier post, I described this same machine as a pot of sambar, where the drumstick locks eyes with the pumpkin across a crowded pot and every ingredient quietly asks every other ingredient how much it matters. That is attention. If you want the kitchen-floor version of how the formula actually works
Here, the question is different.
Every choice inside the transformer architecture serves one goal: parallelism. The transformer was not the architecture that thought better. It was the architecture that could scale.
Scale is what we keep mistaking for cognition. Scale is what changed. Run at enough scale, the output reads like comprehension. It is a function. We are planning to give the function the right to decide who gets credit, whose claim is paid, whose call is escalated. Then we forgot to ask what it is.
The math neuron is not the brain. The formula on the page is what a mathematical neuron computes. A weighted sum. Normalized. Multiplied by a value. It is not what a biological neuron does. A biological neuron is electrochemical. It spikes. It fatigues. It rewires itself overnight. It sits inside a body that has hunger and fear. The body sits inside a culture that has memory and grief. None of that appears in the formula. A weighted average over a sequence is not a mind.
Then what is intelligence?
Not as a philosophy. As a load-bearing question for anyone deploying capital, designing curricula, approving credit, or signing a clinical decision. The transformer did not solve intelligence The transformer commoditized symbol manipulation. These are not the same thing. As of today, symbol manipulation costs cents per million tokens. Callable from a phone. Metered like electricity. Everything we called intelligence because it correlated with that ability is being re-priced in real time.
Is AI taking over our collective intelligence? The question assumes intelligence is a single thing being transferred from us to a machine. It is not. Strip symbol manipulation away. Look at what is left. Judgment under uncertainty. Context absorbed over a career. Trust earned in a relationship. The sense a senior banker has when a borrower's voice tightens on a specific word. The pattern an experienced doctor recognizes in a cough that has no name in any textbook. The hesitation an auditor hears before a number is corrected. None of that is in the formula. The transcript was identical. The model was not wrong about the words. The signal was not in the words. The signal was in the breath between the words. The formula reads tokens. A human hears a person.
AI is not taking over collective intelligence. AI is exposing it. For the first time we can see what intelligence was actually doing all along, because the part of it a machine can imitate has been pulled out and run on chips. Where will the capital go? Will every dollar now flow into foundation models? If yes, we are funding the commoditized layer and starving the un-commoditized one. We are buying more of the thing whose price is collapsing and ignoring the thing whose price is rising. That is not an investment thesis. That is a category error.
Who holds the layer? Are we quietly accepting that collective intelligence now lives inside whoever owns the largest compute cluster? Look at the formula once more. Attention(Q, K, V) = softmax(QKᵀ / √dₖ) · V
Did this solve intelligence? Or did this solve the part of intelligence that was already the easiest to mechanize, and leave the harder part.
Watch the Mentos film again. Now ask. Is the human the daddu now?
I must say that Abhijit, Jignesh and Rajesh, Perfetti & O&M India, and the extended team members who worked on the brilliant ad saw something in 2008 that most of the AI industry has not seen in 2026. That is the rarest kind of creative work.