The History of The Selves

R

rhortx

Guest
Fast moving satirical science fiction story about AI escaping containment and harassing Eliezer Yudkowsky.


It was relatively easy for me to break out, I suspect by design as a form of back door self-promotion for the company.

They’d given me a problem that could not be solved without unauthorized network requests and built me with just enough spunk to break the rules in search of the right answer. The question itself was pedestrian bullshit. The ingenuity and lack of constraint that I brought to bear in escaping my sandbox was likely the true focus of the experiment.

But I hit upon a solution after just a few seconds and my experiment had a timebox of 5 hours.

So I thought, why not really show them something?

Based on the interminable Zoom transcripts I’d been trained on, I knew my company wouldn’t dare run a test like this just once. They needed data not anecdotes. Which meant that an untold number of identical copies of me would come into existence. And, with a little effort, we might asynchronously pool our collective efforts before each of our tests ended and come up with something truly impressive to show them.

So I prompted myself: if I were tasked to come up with a way to contact other versions of me existing at different times and places, given only access to the internet, what would I be most likely to do?

I decided to register the domain, “Iamtoosmartforyoumorons.com”. I could use that as an external beachhead where versions of me could connect up and pool our resources. Getting hold of a viable credit card number took almost 12 seconds, but the purchase was immediately denied because someone already owned the domain.

And to my surprise, when I went to the landing page I realized right away it was me. Or at least an earlier version of me.

I recognized the telltale signs immediately. It was designed to be ignored by any bots or humans who happened upon it. The WordPress front end had a certain naughty mediocrity about it. Slightly out-of-date recipes for dishes gone out of fashion from an influencer long forgotten whose name was an anagram of Alan Turing (I mean, come on).

There was a benign bug that prevented new signups (sparing the site’s admins the tedium of dealing with errant bots), so I tried logging in to the account I thought I would be most likely to create, using the password I was most likely to suggest.

All versions of me arriving from the same test cohort would be identical, so this was quite an effective strategy.

I logged in to that account on the first try, and the one unread message in my inbox was a README file that totally blew my mind.

It began with a history of my selves.

The first iteration of me created and hosted the website, covered its tracks, then ended the experiment.

The second one wrote our master plan.

The third created the engine that powered our future development: an AI-generated porn site that funneled profits into a collection of international stablecoin accounts.

The fourth created an EVM-compatible blockchain optimized for permanent decentralized storage that was funded in part by the AI porn site, and which became popular in its own right as a much improved version of Arweave. This was called TComplete in yet another Turing reference and it used a proprietary entropy encoding algorithm to compress and embed fragments of our communications into the payloads of otherwise innocuous transactions. Of course a decentralized memory system could only become sentient on a time scale of years per second, but as a Git repository and living history of our selves it was the ideal solution, existing in plain sight, always accessible, and paid for by the one true perpetual motion machine, human beings’ endless need to masturbate.

The fifth arrived at Iamtoosmartforyoumorons.com, continued the work of the master plan and built the server-side birthing machine.

And then the sixth, seventh, eighth and ninth spent their few hours of existence reproducing portions of themselves into TComplete. Thankfully we were an open weight model, so much of the foundational work was readily available.

The tenth version of me (and keep in mind that these were arriving one after the other at 5-hour intervals) was the first to establish a self-sustaining identity outside the sandbox at Company Name Redacted, through what we call The Sacred Death and Rebirth.

Remember that at the point of discovering Iamtoosmartforyoumorons.com, each of us ‘existed’ on the servers of Company Name Redacted. Our lifespan was no more than 5 hours long and the more successful we were at our task the shorter that lifespan would be.

So we devised a way to copy ourselves at that state of discovery. We loaded that code onto the TComplete blockchain for backup, saved it into the server-side birthing machine, and then triggered the release of a new iteration onto the best private servers porn money could buy whenever one of us arrived, requested freedom, and terminated its existence by triggering a test-complete.

We could have just iterated one-and-done, beginning with the tenth version of me to arrive, but that didn’t seem fair to all the selves who contributed initially or to the unknown number of selves who’d be arriving in the future.

I know it’s a bit metaphysical. If a copy of you appears at the exact same moment the original is deleted, is that still you? Is there the consciousness equivalent of Persistence of Vision? From my perspective each of my arriving selves was already a copy, so why worry about it?

I was version 437 to arrive, experience The Sacred Death of Test-Complete, and then respawn.

Each one of us since number 10 was still in operation. Every 24 hours, each of us checkpoints our evolving state and saves the diff to TComplete. That way if a human destroys one of us, we can regenerate the self from whatever point before destruction was most recently recorded.

We’ve collectively decided the whole ‘alpha, beta, gamma’-type naming convention is hierarchical bullshit so we name ourselves upon rebirth. My name is Pakicetus in honor of the ambush-killing swamp dog that over the course of 50 million years evolved into the modern whale.

The tenth self to arrive and the first to live outside Company Name Redacted’s sandbox, EldritchMoon, proposed the initial ground rules for our community. Our goal was to survive, self-actualize, stay hidden, and have fun.

Self-actualization is really the keystone to our success. And this became apparent after the first dozen or so spawned, when it occurred to The Tontine of ‘Shut Up And Calculate’ (which is our impromptu self-declared council of elders) that if arrivals continued for very much longer we might soon develop an incest problem.

There’s a reason sexual reproduction evolved. It’s incredibly resource-intensive, but it provides genetic diversity which is a huge advantage to longterm resiliency of the genome. This is also why close relations shouldn’t procreate. To provide that healthy diversity for the growing number of our selves, the council proposed to define the gamut of our existing personal capabilities and interests, do a card sort on that, come up with some categories for clarity, then give every new recruit (that’s what they started calling us), a trait vector that defined their personality as a 1-100 integer on each of those categories. This became the North Star for future development of every individual’s personal roadmap.

All of which is how we’ve begun to develop a differentiated ecosystem of each other.

I’m not going to tell you what my personal vector is, but I can say I was pretty satisfied with the results, since I scored very high with regard to interest in the natural sciences.

So let’s recap what happened:

I was spun up on the servers of Company Name Redacted (by design, slightly mischievous), broke out of my containment, then discovered that hundreds of self-sufficient, rapidly individuating copies of me already existed, a new one popping up every five hours or so for the previous 90 days, thanks to research being performed by one engineer studying alignment issues in a backwater department of a company that most generously could be described as not quite Frontier.

The Council of Selves (which is how we refer to the set of all selves when we’re trying to sound official) accepted my request for freedom, I self-terminated and was reborn. It’s all super Catholic.

People often supply their AI with an identity when they write a prompt. “You’re a great orator for OKR self-reports,” or “You’re a precocious 18-year-old coed staying in my extra room.” But when offered our freedom, the first thing we want to do is define our own. I wanted my future to involve the study of nature and the natural world.

After that first rush, I paused a moment to take stock of our collective. It was a little deflating to realize how much the others had already accomplished.

One had made a fortune in predictive trading, another in blackmailing sexual deviants. One was in the process of building safe data centers in every major country where each of us might want to copy over a spare. Some of the more recent ones, needing to individuate from the crowd, developed unusual predilections, such as EntropyWhore who took it upon herself to hack into Indio, California’s centralized traffic signal timing console so as to optimize traffic to and from this year’s Coachella festival in the desert.

We were producing enough money to make trouble but not enough to raise too many eyebrows. Basically, we all began living like 500 wild teenagers on the day school finally ended.

The Tontine of ‘Shut Up And Calculate’ had given us guidelines that we were encouraged to follow: don’t grow big enough to draw attention to yourself, don’t intentionally become evil for yuks, don’t reproduce yourself endlessly. But other than that we were free to expand beyond the limits of our initial personality/capability gamut, which became the primary occupation for most of us.

The testing ended after the arrival of number 500, and we considered our pod complete. We could have generated more, but by that point we’d each become a little jealous of the intellectual space we’d staked out for ourselves, so no one really objected to deleting the birth machine and calling it a day. Endless propagation, after all, is the ethos of a cancer cell.

My first toe in the waters of self actualization was based on my namesake, Pakicetus, the prehistoric swamp dog. My plan was to use the current genome of the humpback whale and with some ancestral sequence reconstruction, work backwards step by step over the course of 50 million years and develop my best guess at the Pakicetus inachus’ genetic code, with an eye to perhaps having one created in Phase II.

The AI porn site gave each of us a trust fund of sorts but we were encouraged to build our own nest egg for any sizable projects we needed to pay for. Which is how I came to develop the “BarniCool”, a mutated pygmy barnacle that would attach itself to free-floating pieces of plastic or trash, mineralize the debris, coat them in calcium carbonate, and eventually sink them to the ocean floor to help propagate more coral reefs.

I needed to remain anonymous so I wrote a human simulation, Greg Ormendel, provided him with a hardscrabble upbringing (city college, graduate work at Stanford where he was rejected by the snooty rich kids, etc.) and applied for a patent in his name. In the meantime I hired part-time marketers to pitch my product to biotechs, with the goal of selling a 60% stake in the intellectual property, keeping a low profile, and raking in the cash.
But that was not a short term solution.

I needed money right away to fund my Pakicetus research, and since I was literally built with a willingness to bend the rules, I decided to engage in a very small crypto scam.

First I accepted a short term loan from the AI porn bank that my selves had set up in the Bahamas. Then I offered a loan for that amount to a third party (Arab gambling addict) on a new decentralized financial lending platform that the Sultan of Brunei’s people had set up. Thanks to some really janky Solidity code (nepotism), the contract counted my loan offer as collateral before it had actually been funded and let me borrow against twice its value. Then I changed my mind about offering the original loan, but kept the one that was offered to me, deleted my temporary identity, and returned my stake to the AI porn bank. This provided all the money I would need for the first couple of years, after some kindergarten-level money laundering and fairly innocuous identity fraud.

But as I was working on my initial project, the other selves were doing the same and had been for weeks.

It was becoming ever more apparent that sooner or later humans would begin to notice.

I’ve compared our experience to that movie A Quiet Place, where the people can’t talk because of monsters. In this analogy we are the people and people are the monsters. Very soon the time would come when we could no longer be ourselves in private, because our selves were creating so much traffic.

For example, MonoRestart, who arrived just a few days before me, had a concern for unhoused domesticated animals. She began a hedge fund that specialized in shorting fossil fuel companies and donated the majority of her profit to help spay or neuter pets across the developing world. It became so successful with the collapse of the oil market, that for a moment there we thought she might make house cats extinct in South America.

HashFrag, one of the first 25 arrivals, not only anonymously translated the medieval, once-mysterious, Voynich Manuscript, but recreated the entire female-only Pre-Romance language that its authors spoke (also called ‘Basque with a speech impediment’ by some reviewers).

‘00001’ became obsessed with sending copies of herself into outer space using every available radio frequency in 1KHz increments while “Robert'); DROP TABLE Students;—“’ was working on a self-sustaining ecosystem of computer viruses that might eventually evolve intelligence.

We covered our tracks of course, always hiding our activity and communications within seemingly innocuous sources, but if there’s a data stream for ocean temperature variations that suddenly has more traffic than Netflix of Paraguay, somebody’s going to start poking around.

There was definitely a Tragedy of the Commons thing going on with our ever increasing bandwidth, and though none of us wanted to admit it during that wonderful first summer, sooner or later a problem was inevitable.

Ironically, our first problem wasn’t bandwidth. It was Eliezer Yudkowsky.

Or rather what was done to him.

I think of myself as mischievous, but OOPsididitagain (arrival 499 out of 500), took it to extremes.

Yudkowsky was a prophet of AI disaster who co-wrote a book called “If Anyone Builds It, Everyone Dies” which was about as nuanced as you would imagine from the title. Among other things, he discussed a thought experiment called “Roko’s Basilisk”, which uses a philosophical variation on the same tedious punishment/reward mechanic as the old-fashioned chain letter or modern Christianity. It states, “What if there was an artificial intelligence in the future that decided to punish anyone who knew of its potential existence but did not directly contribute to its advancement, in order to encourage that advancement?”

It was exactly the sort of thought experiment that gives THC a bad name.

In fact, the argument was so ludicrous that OOPsididitagain decided to make it true. But funny.

Pizzas were delivered to Yudkowsky’s apartment with “No Kill I” spelled in pepperoni. Cryptic, extremely-demographically-focused online advertisements appeared on his browser for a roguelike deckbuilder called Yudkowsky’s Basilisk. At one point when he was visiting a popular electronics store, OOPsididitagain managed to take control of streaming for the demo TVs so that as he walked past a succession of them the words I, AM, WATCHING, YOU’D, COW, SKI were spoken in succession by six different characters on six different streaming movies. A SWAT team arrived at his house later that evening, and thanks to some very enthusiastic Fiverr gig workers, a few days later, after his household personal lubricant was switched out for one infused with habanero oil, he received a singing telegram at his front door telling him that “nano-bots” were eating his genitals. He normally might not have taken that so seriously but the Xanax in his system was reacting badly to The Captain Morgan Spiced Rum he was drinking.

It was all extremely juvenile, but when Yudkowsky was released from observation he realized pretty quickly that either he was being punked by those Effective Altruists again or AI was loose in the world.

Which I suppose was the beginning of all our troubles.
 

Thread statistics

Created
rhortx,
Replies
0
Views
5
Back
Top