Tag: chatbots

Categories
Society Technology

Not waving but drowning beneath grey goo data

A few decades ago, someone or other caught the imagination of nerds by warning about ‘grey goo’ – a sludge of self-replicating nanobots that once they got going would eventually submerge the planet beneath depthless layers of their endlessly churned-out carcasses.

That or they ate the world to the core with their ceaseless reproduction.

Either way, not so much Terminator as a silicon Sperminator.

Now, you might protest that I’ve been a bit vague with my introduction. Who exactly said this? When? Where?

But let’s face it – this is the direction we’re headed.

I vaguely remember the details. Either I’ll have to Google it or you will. And right now – mostly to make a point – I can’t be arsed.

You knock yourself out though if you like. I’ll wait…

…you’re back already? I suppose some of us slip down the Net’s rabbit holes more easily than others, eh?

Anyway, back to grey goo.

No goo zone

Or maybe not as grey goo hasn’t happened yet. So far the only unnatural junk piling up on the planet is what we humans put there.

Yes, okay, mud in some forms is a sort of carbon-based life’s grey goo. I’m thinking rich peaty loams and forest canopy floors. Oil sands.

And yes, we could have a long chat about the evolution of grey goo that eats grey goo and where that would all go in a billion years.

But this isn’t the article for such diversions.

My point is we’ve escaped Goo-meggedon so far. At least in the physical world, which still matters the most but won’t before the century is out.

Let the goo times flow

Back over in Not Real Life however – the light-fast world of data and compute and 4Chan LOLs and our future – things aren’t going so well.

Just two months ago I wrote that soon we’ll have to demonstrate our identities to prove we’re not AI-bots. To authenticate our humanity on everything from Twitter to (eventually) talk radio phone-ins.

But by ‘soon’ I mean I was thinking a few years. Not months.

However ChatGPT has put the vanguard of machine learning into the public’s hands with a global gusto. And the resultant cacophony of coverage and wonder has even outdone Dall-E and Stable Diffusion and the other image-focused AI systems’ debuts earlier this year.

Indeed no sooner than you can say “a gun is a dangerous weapon, don’t point it at anything important and treat it with respect” people are already flooding Internet forums created for human pronouncement and consumption to post ChatGPT’s (stupendous) verbal vomit.

To give just one example, Liam Tung of ZD Net reports:

Stack Overflow, a site where developers can ask and answer coding questions, has temporarily banned the use of text generated from ChatGPT, a chatbot released by Open AI last week.

ChatGPT is based on OpenAI’s GPT-3 language model.

People have quickly discovered that, while it answers prompts in a “human-adjacent” way, there can be flaws in the answers it gives.

Basically, keen Stack Overflowers (I guess? I’m not a local) have been spamming the site with ChatGPT-created content. Which is a problem when even the bot’s creator, OpenAI, stresses its precocious child can deliver “plausible-sounding but incorrect or nonsensical answers.”

Liam Tung – who may or may not be an organic life form himself – continues sagaciously:

This appears to be a key cause of its impact on Stack Overflow and its users who are seeking correct answers to coding problems.

Additionally, because ChatGPT generates answers so quickly, some users are supplying lots of answers generated by it without parsing them for correctness. 

Oh Liam! In just a few casually typed out / generatively predicted words there, you’ve raised so many questions about the future of knowledge, civilisation, blogging for early retirement, and getting a robot to do one’s schoolwork.

But it’s a common kernel at the heart of all that which we’ll be choking on today.

The goo goo trolls

You see, what has struck me most forcefully in the past fortnight is the sheer volume of data these things are going to create.

So far it’s just everyone and his dog. But soon it will be everyone and their robot. Then their robot’s robot. And so on.

Listen carefully! Do you hear a tap-tapping?

Yes, it’s the sound of Lithuanian troll farms and Texan entrepreneurs alike bashing out code and sending forth bots to drown us all in this calorie-light info-crud in order to earn a few dollars from Google’s Adsense. Or perhaps to tilt us to vote this way or that.

Look, I don’t know their nefarious plans. I’m one of the good guys.

But even as I’ve been pecking away for hours like every other curious pigeon on ChatGPT’s levers in the hope of another crumb of dopamine, it struck me that humanity’s death by AI might be even dumber than I previously suspected.

True, I’ve led the field among very obscure pundits in warning that just getting rid of 90% of low-level knowledge jobs could be more than enough to rupture society. At least for a few decades and a civil war or two.

In other words we don’t need a post-singularity-level evil and scheming AI to explain why, with regret, the human race has to go, as it dangles us over a bubbling vat of metaphorical sulphuric acid.

No, just a cheap and dumb-ass bot that can be copied-and-pasted over the world’s office employment might do the trick.

But what’s clear from even these early skirmishes with ChatGPT is an even more insidious risk – the danger of our emerging ‘other’ reality being rapidly flooded with data goo. ‘Plausible-sounding but incorrect or nonsensical’ gunk – or even mostly right but still exceeding mediocre verbiage – that crowds out what little signal is left in a world that already seems to be turned up to 11 on the noise front.

Holly goo lightly

Of course there will be solutions. Every platform already has its own Deckardian countermeasures seeking to resist or destroy this stuff.

But the models will improve. And it will be ever harder to tell their output from a real person’s most precious shared thoughts.

At best it’s yet another arm’s race that you and I will increasingly be viewing out of a passenger window rather than the cockpit, while software does all the heavy-lifting.

Leaving us wondering if we’re on a nice getaway flight to someplace warm, or actually cooped up inside a missile heading for impact.

Or else I guess it breaks the Internet and we go back to fire and parchment and worrying about the wolves.

(But no we won’t go back. As I wrote previously we’ll sign-in as provably real people and fence this stuff out. But we’ll have to do that, for as long as we can. Nobody wants the wolves again.)

Hah! Stick those 1,153 words in your natural language model’s pipe and smoke them. I won’t go down without a fight.

Categories
Technology

We don’t serve their kind here

I am old enough to remember when pleading emails from Nigerian princes locked out of their multi-million fortunes weren’t just a meme.

They actually showed up in your inbox.

Curiously, the initial giveaway with those emails wasn’t even the craziness of the claim.

(The gist of which was that you, Mr Nobody Orother of Upper Nowheresville, was the only hope this poor man had of releasing his fortune – if only you could wire him £10,000 to facilitate the transfer).

No, what marked the emails out was they were invariably shot through with spelling and grammatical errors.

It seemed odd, given the supposed importance of the communication.

Indeed it seemed odd from the perspective of perpetuating a crime.

Did fluent English-speaking Nigerians refuse to work with fraudsters?

Was there something about fraud that caused a word-perfect email to decay into a tell-tale red-pen-fest for any teacher of 12-year olds?

Had the emails been copied and pasted too many times?

Or was it some elaborate way the fraudsters had to track who’d been sent what email, before the coming of marketing response analytics? The same way publishers will put a spelling mistake into a dictionary to spot and prove a counterfeit?

We never learned, as far as I know.

But at least the incompetence made deleting the emails easy.

Death by botulism

Thirty years on, and armies of impersonators assail us on the Internet.

Or rather, they assail our digital identities as we parade them on Twitter or in the comments of a YouTube video.

Casually called bots, they’re nothing like the Robbie the Robots of 1950’s imagination.

Rather these are one-trick one-track ponies whose sole function appears to be to sing the praises of Putin or else tout crypto.

For now pattern recognition again matches and dispatches them.

But for how long?

Anyone who has played with language prediction models like GPT-3 knows they are becoming scarily good at stringing sentences together.

Indeed while many in the field of AI have been complacently (to my eyes, anyway) shrugging their shoulders at the speedy rise of these potential Turing Test busters, at least one Google employee got the sack for saying his favourite chatbot seemed sentient.

An interesting discussion for another day. (Although not one to have with your GPT-3 bromance buddy if you want to keep your job…)

For the purposes of this post I’m more concerned about the apparently imminent chatbot-singularity.

Ticked off bots

That’s not an official term, incidentally – chatbot-singularity.

It’s a phrase I’ve just coined to describe when as many neural net chatbots as can be pumped out are wandering around the Internet indistinguishable from humans.

Indistinguishable at least to anyone but professional AI-witch hunters.

(Think Blade Runner. But with less film noir and nudity.)

After the chatbot-singularity, it will be really hard to know who is human online. Let alone who is the dog of New Yorker cartoon fame.

And it gets tricksier.

Researchers have been trying to train digital agents for years to start their knowledge (/language) land-grab with a keyboard and mouse.

In other words, before you throw the whole Internet at your natural language model so it can learn how to answer anything (which, yes, is what’s going on and if you’ve not been keeping up then your blown mind is excused) you first let it learn about pointers on screens and QWERTY keyboards.

It’s laborious, but once achieved a natural language model could then be prompted to do active things on the Internet that are again indistinguishable (from the Internet’s perspective) to a human.

You think I exaggerate?

As Lex Fridman recently pointed out in a podcast that covered all this territory, how many times have you ticked a box on a data entry form to ‘prove’ you’re not a robot?

Yes, that’s literally the test.

You supply no proof. Being able to tick the box is proof enough.

I’d say that particular security barrier hasn’t got long left to live.

Papers please

Muse on this for a few minutes and I suspect you’ll end up reaching the same conclusion as me.

Which is that you’re eventually going to have to show your passport or your driver’s license to write a post on Reddit.

Not literally. But in some digital form.

Because Reddit (presumably, though we’ll see how society progresses…) doesn’t want robots writing posts as if they were human.

Which chatbots can pretty much now do. (In fact, they can interview each other, complete with deep fake voice impersonations).

And once they can get themselves email addresses and jump through human-ish hoops with their keyboard and mouse skillz, the walls keeping out their conversations will crumble.

So yeah, after that you’ll need to show you are you.

Probably you’ll authenticate a device at first. Your phone or your laptop. But if you use some other device you’ll have to re-authenticate.

Eventually it’ll be biometric, maybe linked to wearables.

The point is some node on The Internet won’t have to compare the data coming from you to the data from Joe Terminator to decide which of you is flesh-and-blood.

You’ll have proven that at some previous stage in the chain, via your birth certificate or whatever, and you’ll point to the proof to continue.

You and whose army?

What if you want an AI agent to do your bidding?

Isn’t any code that interacts with other code on the Internet a bot from this perspective?

Well yes.

Which means traces of you-proof will probably be encoded into anything you initiate and launch on the Internet.

Perhaps absolutely everything you do digitally.

At a simultaneously more concrete and more trivial level, you’ll also eventually own personality-complete human-like chatbots who know what you like, and do stuff for you. They’ll be knitted to you and your reputation the way the valets of old were tied to their masters.

We’re just too slow and digitally cumbersome for this not to be part of our increasingly digital future.

So people or systems (with the right permissions) will have to be able to see they’re your bots, via the integrated you-proof they carry.

A brave new world where your digital DNA leaves traces everywhere.

Oh, and did a thinking bubble just appear above your head complete with the words Use Case For The Blockchain?

Yes I’m inclined to agree.