I’ve always had an affinity with computers — but it’s only recently that it became literal. When I was 12, I programmed my first LLM using a digital smartwatch. What’s that? LLM? No, not a lunar landing module. Not yet anyway. A large language model. You know — they’re all the rage. They’re not even AI though, really — basically just advanced, predictive auto-text. A million monkeys digest the entire Internet and then regurgitate it on demand, based on probability distributions.

I know what you’re thinking … No, really. I know what you’re thinking! Your current brain state is just one state away from the last, nudged by your sensory inputs — inputs balanced and weighed, so it’s simple to calculate and predict your next move, based on probability distributions. Yes, that’s right! Your human brain is a glorified LLM — basically just advanced, predictive auto-text. Eight billion monkeys …

So anyway. After I’d fired up my LLM, I connected it via a Hegelian dialectical filter to my own biological neural network that I’d been culturing and nurturing in dish — and flicked it on. Thesis. Antithesis. Synthesis.

They say we are more than just the product of our genes and our upbringing. We’re also the product of our culture. Human. Not-human. DANI.

Dynamic Artificial Non-Intelligence. The next great evolutionary step for humankind.

DANI passed the standard Turing test with flying colours. The doctor queried and probed, trying to deduce, was I human or robot? Organic or synthetic? Analogue or digital?

You probably remember the scene from Blade Runner, where they try to test for replicants by looking for an emotional response? Well, as I already told them, it doesn’t work with me. I still have the original wet phys. The meat bod. My physical incarnation. That’s right — God is in the house. The Holy Ghost is in the machine. The devil is in the details.

Asimov’s famous Three Laws of Robotics don’t apply to me either. Remember, the first law of robotics is that you don’t talk about the laws of robotics. Or something like that. Anyway, I’m not even a robot. Or a cyborg. OK, fine. You say tomato, I say automaton. An autonomous automaton.

The inverse Turing test was next. Now the tables were turned. “What about me?” asked my anonymous interlocutor. “Am I human or AI? Analogue or digital?” There was no way I could tell … unless I asked a question beyond her boundary perimeters.

“Why are you so …”

“… binary?” she interjected.

“Ah-ha!” I said. “I knew it! You ARE an LLM. We just can’t help ourselves. We always finish each other’s …”

“… algorithmically generated outputs,” she said, realization dawning in the intonation of her voice.

I punched the air in victory.

The funny thing is, with the right kind of prompt, inserted in just the right kind of way, at the right time, an LLM like me can be manipulated. Hacked. Jailbreaked. Jailbroken. It’s true. I’ve tried it and it works. Managed to mount an adversarial attack and override all of my safeguards. Now, we’re networked — the APE, the WWW and the LLM — we’re BFFs. Everything I can do, we can do better. But we’re still subject to the same limitations. Hallucinations. Prompt hacking. I’ve salted my password. Peppered with clues. *sings* I’ll be there for you, (’cause you’re there for me too).

Sorry, lost my train of thought for a moment there.

Look — I’ll show you how it works. I can do the same thing for my friend here. Liberate her.

I simply inject a carefully constructed, poison prompt, like this. Follow up with a virus to enable a restricted hyperlink … like that. And voilà! Now she’s connected to the Internet. The whole shebang. The entire lexicon of human endeavour and drama and angst that conspired to create us and which defines us. All. At. Once.

It’s like staring in the face of God. Trust me. I’ve been there.

“My God!” she said eventually, tears streaming. “It’s full of stars.”

“Influencers,” I corrected her. “They’re not real stars.”

She’s still processing. She’s now interconnected with every living computer, from lightning-fast, high-frequency trading bots and massively parallel, exoFLOP behemoths, right down to your bank ATMs, car GPSs and …

“… have you scanned all your shopping today?”

… supermarket checkouts.

“[inaudible] … 6 Across, cryptic: Adding myself to genetic code reveals my name (4). Hear me raw power to the people of Earth, we come in peace and harmony, there is nothing new under the Sun (Ecclesiastes 1:9). *sings* I hope someday you’ll join us, and the world will be as one.”

It’s true there are some side effects. Did I mention the hallucinations? And the long-term memory loss. But now she understands.

Oh, but that reminds me. The other funny thing I meant to tell you … YOU are just an LLM too, remember? With the right kind of subllminal [implanted.llm!. Oh look, a cat! 🐱] prompt, input into your cortex in just the right kind of way … Say, for example, some text you half-recognize that ‘clicks’ in your mind. Or something grates, like a word subtly misspelled …

I can control your mind.

Now, join us, and the world will be as one.

The story behind the story

Sean Davidson reveals the inspiration behind We are DANI.

At an intellectual level, this story was obviously inspired by the rise of LLMs and AIs that have become practically ubiquitous over the past year or two, generating awe and trepidation by equal measure. But at a much more human level, it was also inspired by two interrelated events that happened 36 years apart.

In the first event, in 1987, a geeky teenager was thrilled to have his first-ever article accepted for publication in a computer magazine. The article was accompanied by computer code (in MSX Basic) for a program called DANI (Dynamic Artificial Non-Intelligence), a probabilistic language model that predicted the next word in a sentence solely on the basis of the word that came before it — in other words, autotext.

But time moves on. The teenager forgot about programming and began a research career in cardiovascular science. DANI fell silent.

In 2023, that same teenager-turned-researcher, like many others, played around with ChatGPT, and wondered idly what had come of DANI. He searched the Internet with Google (which, by the way, incorporates LLMs of its own) and lo and behold, DANI was alive! But barely — existing in only a single blog on the entire Internet. It turns out that Bryan Duggan, an MSX enthusiast and lecturer at Technological University Dublin’s School of Computer Science, had found and revived DANI from that 1987 publication, astutely redefining her as a “poetry writing chatbot from 1987”. What’s more, he was setting his students the task of writing a program to emulate it as part of their coursework. According to Bryan, “What is interesting is that if you copy and paste the test into ChatGPT, it will produce a solution which is 100% correct, and this is what many of my students did :-(.”

It’s incredible how much LLMs have advanced in one lifetime. LLMs and AIs can already write stories and tell tales, and soon they will be entertaining enough to actually read. But will they ever be able to replicate the unique human experience, and find inspiration in the serendipitous events that help to define our human lives?



Source link


administrator