The Ghost in the Machine is Just T9 on Steroids

Illustration: chili pepper mascot illustrating The Ghost in the Machine is Just T9 on Steroids

Remember the good old days? Back when your phone was a brick, Snake was peak entertainment, and texting was an exercise in finger gymnastics and existential dread. We're talking about the 1990s, specifically the era of T9 predictive text. Call it The Ghost in the Machine is Just T9 on Steroids.

You’d punch in "4-6-6-3" and, like magic, "good" would appear. If you wanted "home," you’d hit the "next" key. It was clunky, it was frustrating, and it was, in its own primitive way, AI.

Fast forward to today, and we're all collectively losing our minds over ChatGPT. It writes essays, codes software, and occasionally tells us how to build a nuclear reactor out of household items. It's hailed as the dawn of a new intelligence, a sentient digital being.

Here's the dirty little secret, though: ChatGPT is just T9 on a truly epic dose of steroids. The fundamental trick? Identical. The core flaw? Scaled up to terrifying proportions.

The OG Oracle: T9's Reign of Error

Let's rewind to 1996. A Seattle company called Tegic cooked up T9, short for "Text on 9 keys." It was a revelation for anyone tired of mashing "7-7-7-7" for an 's'. Nokia, Motorola, and Samsung quickly licensed this wizardry, making it ubiquitous. Tegic, by the way, was later gobbled up by AOL in 1999, then by Nuance in 2007. https://en.wikipedia.org/wiki/Tegic

How did it work? You’d press each key once, and T9 would guess the word. It prioritized the most frequent match. So, "4-6-6-3" would give you "good" before "home," "hood," or "hoof." A quick tap of the "next" key would cycle through the alternatives. It even learned your preferred words, subtly adapting to your linguistic quirks. https://www.omnicalculator.com/how-does-predictive-text-work

T9's inventor, Martin King, started with an eye-tracking device for disabled users before founding Tegic in 1995. He passed away in Seattle at 60, leaving behind a legacy of slightly less painful thumb-typing. https://gizmodo.com/how-t9-predictive-text-input-changed-mobile-phones-5646477

But T9 had a dark side, a precursor to our modern AI woes. Its dictionary was tiny. Like, "my-cat-has-a-bigger-vocabulary" tiny. This led to what we affectionately called "junk words" – T9's confident, yet utterly nonsensical, guesses. https://alchetron.com/T9-(predictive-text))

You'd be trying to type something perfectly reasonable, and T9 would offer up a word that made you question reality. Its dictionary was so small, it probably thought "existential dread" was just a typo for "excellent bread." This was the ancestor of the AI hallucination, a bug that would one day plague systems with billions of parameters.

From Dumbphone to Supercomputer: The Scale-Up

Now, let's talk about ChatGPT. It's a "large language model," a fancy term for a statistical next-word prediction engine. It doesn't understand in the human sense; it just calculates the most probable next word based on the words that came before it.

Sound familiar? It should. That's exactly what T9 was doing.

The difference, of course, is scale. T9 worked with a dictionary of words, a paltry collection compared to what came next. ChatGPT? It works with a dataset the size of the entire internet. https://tanurhansen.medium.com/clm-series-chapter-3-do-you-remember-t9-heres-your-first-encounter-with-artificial-intelligence-6b075cb69213

Imagine T9, but instead of a few thousand words, it had read every book, every article, every tweet, every comment section, and every recipe for artisanal sourdough ever published online. That's ChatGPT. It's the same fundamental trick, just executed with a truly staggering amount of data and computational power.

The Emperor's New Autocomplete

And with that scale comes the same fundamental bug, amplified a millionfold. T9's "junk words" were annoying. ChatGPT's "hallucinations" can be terrifying. It confidently invents facts, cites non-existent sources, and occasionally makes up entire legal precedents.

Beneath the hood, ChatGPT is still just a glorified parrot with a PhD in making stuff up. It's a master of sounding plausible, even when it's utterly wrong. It doesn't know it's wrong because it doesn't know anything. It just predicts.

When T9 offered "duck" instead of "fuck," it was a minor inconvenience. When ChatGPT fabricates a court case or gives dangerous medical advice, the stakes are a little higher than an embarrassing text message. The "confident wrong guess" bug, once a minor annoyance, is now a feature of our most hyped AI.

Why We Keep Falling For It

Why do we keep falling for it? Because the scale is so immense, the output so often coherent, that it feels intelligent. We project human understanding onto a statistical engine. We want to believe in the magic, in the emergence of true intelligence.

But it's just a very, very good autocomplete. It's a sophisticated pattern matcher, not a thinker. The "magic" is in the sheer volume of data it's processed, allowing it to string together words in ways that often mimic human thought. It's a parlor trick, albeit one performed on a stadium-sized stage.

The Future of Confident Nonsense

So, what does this mean for the future of AI? It means we need to temper our expectations. These systems are incredibly powerful tools for generating text, summarizing information, and even assisting with creative tasks. But they are not infallible oracles. They are not sentient.

They are, at their core, descendants of T9. They're brilliant at predicting the next word, but they're just as capable of confidently guessing wrong, only now with the gravitas of a supercomputer and the reach of the internet. The ghost in the machine isn't a ghost at all; it's just the echo of every word ever written, occasionally tripping over its own feet.

Sources & further reading

https://en.wikipedia.org/wiki/Tegic https://www.omnicalculator.com/how-does-predictive-text-work https://alchetron.com/T9-(predictive-text)) https://gizmodo.com/how-t9-predictive-text-input-changed-mobile-phones-5646477 * https://tanurhansen.medium.com/clm-series-chapter-3-do-you-remember-t9-heres-your-first-encounter-with-artificial-intelligence-6b075cb69213