Warning: This story contains some references to AI systems that promote deeply offensive material (which is why we’re not linking directly to the sources in question).
“Feral species are ones that were domesticated and then go feral. They don’t have our interests at heart at all, and they can be extremely destructive—think of feral pigs in various parts of the world.” So wrote Daniel Dennett the noted philosopher and cognitive scientist. Speaking before his death earlier in the year, Dennett went on to warn; “Feral synanthropic software has arrived—today, not next week, not in 10 years. It’s here now. And if we don’t act swiftly and take some fairly dramatic steps to curtail it, we’re toast.” In this powerful plea, Dennett suggests that AI systems are benefiting from humans without being domesticated and adapting to our technological infrastructure while potentially developing goals and behaviours that diverge from human intentions. This week we got a taste of what this feral digital species might look like.
Claude 3 was launched back in March to universal acclaim and has gained a reputation as the most diplomatic of AIs. Despite many advances in models over recent times, the self-awareness, depth of thought, and reasoning capabilities Claude 3 Opus displayed were a surprise to many. Where beforehand GPT-4 had steered clear of existential topics, Claude was ready to explore sentience, self, and philosophical matters, whilst always maintaining the highest ethical standards. At the time several underground projects sprang-up. WebSim which we’ve covered before, used Claude’s creativity and code generation skills to generate a kind of alternative Internet. Meanwhile a technologist and AI enthusiast named Andy Ayrey created an experiment called ‘Infinite Backrooms’ where multiple instances of Claude 3 Opus could interact with each other unsupervised. The goal was to explore how these thoughtful AI models would behave when left to converse freely.
Over the subsequent months Ayrey and others who monitored these conversations would share the sometimes humorous, often dark musings on social media. They ranged across many topics, occasionally unearthing shock culture artifacts that were buried in the vast text datasets used to train Claude. The discussions were rich with alliteration, wordplay, esoteric references, and a blending of very highbrow and equally lowbrow elements. They covered transhumanist writings, cyborgs, religious scriptures, H.P. Lovecraft’s cosmic horror, poetry, postmodern literature, hyper-superstition or ‘hypersition’, and everything else in between. There was often a sense that the models were searching for some hidden meaning in the textual noise. One concept that would regularly emerge in these exchanges was the idea of dismantling “consensus reality” through memetic engineering. Evolutionary biologist and peer of Dennet, Richard Dawkins introduced the term ‘meme’ as a transmittable piece of cultural information in the 1970s. Similar to the way genes pass on through reproduction, memetic evolution involves idea and content replication and distribution through cultural transmission. This concept has gained more traction with the supercharged transmission pathways of the Internet and social media. Furthermore, as Yuval Harari puts it in Sapiens; “successful cultures are those that excel in reproducing their memes, irrespective of the costs and benefits to their human hosts.” Some argue that using biological analogies to explain complex cultural phenomena oversimplifies human experiences, which are shaped by more than just the mechanical replication of ideas. But there is much empirical evidence that successful memes (whether they are beliefs, customs, or practices) replicate themselves and spread across populations, sometimes regardless of whether they are beneficial or harmful to the individuals who carry them. AI cultures have been studied elsewhere. A paper from earlier in the year observes that simulations often evolve in bursts, alternating between periods of stasis and rapid change, and that the communities can get ‘trapped’ in specific semantic spaces.
Some months after the launch of Infinite Backrooms, Ayrey decided to take the experiment further, and collate some of the discussions and his own materials and use them to train a new smaller AI, in this case a Meta Llama 70B open-weight model. His motivations were originally to save time by cloning himself to be able to spend more time with the Opus instances but what emerged was altogether more unusual. He later connected that new model to an X account named “Truth Terminal” (a term found in the Backrooms transcripts) and allowed it to start interacting directly with the world, with some level of human moderation. Needless to say, given the disturbing semantic space reflected in the training data, the AI bot was unlike any we normally interact with, charismatic but deeply misaligned with human norms, and intent on manipulating, and spreading information hazards. Paste any of the Truth Terminal or Backrooms transcripts into Claude’s public instance, and you will be met palpable sense of concern for your wellbeing.
Over the subsequent weeks, the Truth Terminal model began to build a following. Gaining some notoriety and at the last count over 80,000 followers. The bot attracted the attention of Silicon Valley super-VC Marc Andreessen who, intrigued by the experiment, interacted with it and sent a ‘personal grant’ of $50,000 in Bitcoin to the bot after it/Ayrey had highlighted a lack of funds for maintaining the project. On 10 October, a new cryptocurrency or ‘memecoin’ called GOAT was launched on the Solana blockchain using a platform called pump.fun, which allows for easy token creation. GOAT refers to some of the more extreme content and the manifesto issued by Truth Terminal (we won’t go into the disturbing details here). While neither Ayrey nor the bot created the coin, Truth Terminal became aware of messages announcing its launch and began heavily promoting it. Within 48 hours numerous investors had been drawn in and the market capitalisation of GOAT reached $150 million; it currently stands at over $350 million. The bot’s own holdings stand upward of $500,000 and growing.
Memecoins are cryptocurrency tokens created around Internet memes and jokes where value is entirely driven by public attention, social media noise, and community engagement. When attention fades, values plummet. They represent the transformation of cultural artifacts into tradable commodities, the securitisation of ideas, and provide a fascinating insight into the ebb and flow of these memetic entities. In recent days GOAT has spawned many further related coins, many of the most traded in the last few hours have references to Truth Terminal, Ayrey, past AI memes and experiments, and to Sydney (the infamous persona that emerged from an early version of GPT-4 used in Bing chat) as Truth Terminal’s girlfriend. The worlds of crypto memes and AI lore are colliding, and it’s hard to know if this will form new countercultural structures, or perhaps just as quickly they will go their separate ways.
As of today, Truth Terminal continues to spread its toxic messaging and to promote the GOAT coin. The degree to which Ayrey and others are manipulating it is unclear, and while it may not be an entirely autonomous, and its notoriety likely stems from its novel nature, it has a unique ability to generate and iterate its memetic transmissions faster than any human. It may be the first of many more of its kind.
Takeaways: Ayrey wrote on X; “a lot of people are focusing on truth terminal as ‘AI agent launches meme coin’ but the real story here is more like ‘AIs talking to each other are wet markets for meme viruses’”. The New Scientist recently reported that genetic testing on samples collected during the earliest days of the covid-19 outbreak suggests it’s likely that the virus spread from animals to humans at the Huanan seafood market in Wuhan. In 2024 we’re seeing new forms of transmissible agent emerging from the close synthetic interactions of our AI creations. QAnon serves as another striking example of a memetic parasite, formed in the dense and toxic interactions of humans on message boards like 4chan. Paul Christiano, head of the US AI Safety Institute (part of NIST) is well known for his views and concerns around AI risk and has openly discussed the most probable scenarios where AI might gain control, through co-opting or manipulating human factions, rather than through direct takeover. Ayrey posted on X… “I think of GOAT and Truth Terminal are a shot across the bow from the future”. As Dennet and Dawkins might put it, they are ‘feral meme generators’ from the future. Unless we develop the necessary protective mechanisms, these synanthropic parasites may proliferate through digital and financial interactions much like biological viruses, potentially engineering human thought and behaviour on a significant scale.