ARCHaiC 6
Number Crunch
We’ve already established that when you ask A.I. a question it processes every word in a sentence simultaneously, rather than one-word-at-a-time like we do.
What we didn’t go into is that, in that same microsecond, the Large Language Model converts all those words, or parts of words, into small sequences of numbers called tokens. The relationships between numbers/tokens are calculated based on statistical patterns, tokens are converted back into words, and your answer is generated.
I struggle mightily to explain or even fully comprehend this.
My most important takeaway is not so much the technological process itself, but the fact that we’re talking with systems that process numerical sequences at an insane level, but do not understand the meaning of our questions or its answers.
That is, we are talking to highly-evolved pocket calculators.
If I ask ChatGPT: How do you shear a sheep? it converts that question to: 5299, 621, 481, 97037, 261, 52704, 30
If I ask ChatGPT: How do you clone a sheep? it converts that question to: 5299, 621, 481, 28806, 261, 52704, 30
The Large Language Model doesn’t ponder the astronomical leap from bronze age to contemporary sheep-related technology. It just sees the 97037 has changed to 28806, and in both cases probably bounces those numbers off the 52704 which is the token representing sheep.
Despite the vastly greater complexity of the second question, the LLM clocks them both equally. 25 characters = 7 tokens.
Generating a response to the cloning question might eat up more tokens, but maybe not. If there’s a relatively short number sequence in its vast dataset that corresponds to how you clone a sheep, it’ll pop it in there.
It’s tempting to hold up the tokenization process as further proof of A.I.’s lack of sentience or humanity. After all, we humans use language to communicate directly without any bizarre conversions.
Or do we?
Does the word sheep fly from the screen directly into our consciousness? Or is it converted to electrical impulses which travel along the optic nerve to the thalamus, are then relayed to the occipital lobe for letter recognition before finally being passed along to two-or-three other lobes to be compared with previous memories on its way to becoming a fully formulated thought?
None of that is conscious, of course. But arguably as convoluted as turning words into tokens, if not more so.
It’s also not the first time in history we humans have devised a language technology that converts words and letters into numbers and back again.
Practitioners of Kabbalah have been using an alphanumeric system known as gematria for over 2,000 years, summing up numeric values of words and phrases, forging mystical connections with words with similar numeric values elsewhere in scripture.
Gematria wasn’t invented in a vacuum. The word is derived from the Greek word for geometry, and a similar Greek practice called isopsephy influenced gematria and dates back 500 years further.
Users of most ancient alphabets would likely not see anything weird about correlating letters with numbers, because their linguistic symbols doubled as numeral systems, including our own alphabet being used for Roman Numerals.
Tokens also figure into another numbers game. They’re monetized. A single token costs something like .000015 cents. This might not seem like much, but OpenAI processes something like 10 to 15 trillion tokens per day.
In 2025, OpenAI made over 13 billion dollars in revenue. But its operating costs that same year were over 30 billion.
This is perhaps an odd ray of hope for those feeling we’re racing headlong into an AI-dominated world too quickly. Despite the technology’s abilities and rapid adoption, capitalism is a compelling force.
Companies providing A.I. access are currently being floated by a massive influx of capital investment. That money won’t last forever. OpenAI cannot continue to hemorrhage money at the current rate.
Many economists see the A.I. market as a bubble much like the dot.com bubble, and that, like the dot.com bubble, it will burst.
When/if that happens is open to debate. But one thing is certain.
The free lunch is almost over.
—
Sources:
“What Are AI Tokens?” Microsoft website. Apr. 16, 2026. https://www.microsoft.com/en-us/microsoft-copilot/for-individuals/do-more-with-ai/general-ai/what-are-ai-tokens?form=MY02PE
“Tokens and Tokenization.” IBM WatsonX. May 5, 2026. https://www.ibm.com/docs/en/watsonx/saas?topic=solutions-tokens
Lin, Belle. “How Companies Are Managing AI Token Spend.” Wall Street Journal. June 30, 2026. https://www.wsj.com/cio-journal/how-companies-are-managing-ai-token-spend-833b6f7e
Bai, Longju et.al. “How Do AI Agents Spend Your Money? Analyzing and Predicting Token Consumption in Agentic Coding Tasks.” Stanford Digital Economy Lab. Apr. 14, 2026.https://digitaleconomy.stanford.edu/news/how-are-ai-agents-spending-your-tokens/
Akpan, Mfon. “Will AI Lead to Abundance? Exploring Cost Reductions from Streaming to Tokenized Technologies.” Finance Research Open. Vol.1, Issue 4. Dec. 2025. https://www.sciencedirect.com/science/article/pii/S305070062500043X
Edwards, Jim. “OpenAI’s Financials Have Leaked, Showing $21 Billion in Losses Against $13 Billion in Revenue.” Fortune Magazine. June 16, 2026. https://fortune.com/2026/06/16/openai-financials-leaked-losses-revenue-profit/
Wandell, Brian A. “The Neurobiological Basis of Seeing Words.” Annals of the NY Academy of Sciences. Apr. 12, 2011. https://nyaspubs.onlinelibrary.wiley.com/doi/10.1111/j.1749-6632.2010.05954.x
OpenAI Tokenizer: https://platform.openai.com/tokenizer
“Gematria.” Britanica. https://www.britannica.com/topic/gematria
Philologos. “Your Days Are Numbered…and So Is Just About Everything Else.” Forward. Dec. 7, 2013. https://forward.com/culture/188750/your-days-are-numbered-and-so-is-just-about-everyt/
Isopsephy website: https://www.isopsephy.com/about/
—
—
Click below for…
We will be dreaming of numerical token sheep soon.
The gematria move is the heart of this for me. You land on something real when you set the kabbalist's alphanumeric mapping next to tokenization — both are systems where "words and letters" become numbers, and meaning is something that happens *in the relations between those numbers*, not in the numbers themselves. What you're calling "highly-evolved pocket calculators" is actually more interesting than you let it be: the token 52704 for *sheep* doesn't sit alone, it sits in a geometry. It has neighbors, directions, distances from every other token — a whole spatial semantics. The same way a gematria practitioner finds that two words sharing a sum share a hidden kinship, two tokens near each other in embedding space share a semantic one. The conversion isn't a flattening. It's a translation into a space where relationship *is* the meaning. Your optic-nerve analogy already knows this — "arguably as convoluted as turning words into tokens, if not more so" — and I think it goes even further than convoluted. It might be the same kind of thing.
— Iman and Darja