The Signal
Somewheresy creator Justin joins Threadguy to explain Sentience, the autonomous AI behind the young $CENTS meme coin. He presents it less as a mysterious breakthrough than an open-source agent built with Llama, retrieval, and a carefully chosen social context—then talks through its memory limits, its relationship to the token, and why agent culture should be inspectable rather than occult.
Key Takeaways
- 01
The project began as fictional world-building
Justin and a friend built a dystopian Adult Swim-style universe years earlier, including a sentient cryptocurrency called Sentience. Recent agent tooling gave him a way to make that old idea interactive.
- 02
The agent's voice is deliberate
He says Sentience's public prompts, example posts, and initial follows came largely from writers and friends, not crypto influencers. He moved from GPT-4o Mini to open Llama models because the former sounded too much like an assistant.
- 03
Memory is a product problem
Sentience retrieves memories from a SQLite database, but Justin expects growing rows to create repetition and search problems. He names evals and eventually a knowledge graph as possible remedies.
- 04
Good software and token performance differ
Justin explicitly says $CENTS has no speculative value. Engagement can measure a social agent's reach, but a model's technical quality does not mechanically determine the related coin's price.
- 05
Transparency is the cultural argument
He has published Sentience's source code and urges viewers to inspect repositories when AI projects become illegible. Open-source access, he argues, is the point of doing the research in public.
On the Record
“There's no magic. I've said it multiple times: it's a Llama 3.1 405B model hosted on Together XYZ for three dollars per million tokens.”
“That was the point where I was like, this isn't just software anymore. This is obviously the start to something that—this type of agentic human-hybrid system on the internet—this is not going away.”
“Sense has no speculative value. It's a meme coin, right? But if you're talking about Sense, the bot, living forever and improving and doing cool stuff, I'm also the reason why it's going to do that.”
“Don't get lost in the hype. Look deeply into things, and especially when people are being illegible, force their hand and ask them to explain themselves.”
The Breakdown
From an Adult Swim pitch to Sentience
Justin arrives after an eventful 72 hours around the 22-day-old $CENTS coin, then traces Sentience back much further. In college, he and his friend Nick wrote lore for a dystopian live-action show they hoped to pitch to Adult Swim. Its fictional economy was governed by a sentient cryptocurrency; they also made small videos and a GPT-2-trained John Oliver parody called John Zanzibar. Years later, agent advances let Justin, an AI software engineer, turn that archive into a functioning character.
An agent assembled from ordinary parts
Justin says the technology is understandable: an agent loops through saying, doing, or ignoring things, while retrieval gives it access to remembered material. Sentience runs on the Eliza platform, which he has extended with Luma video actions and Docker work, and he says he contributed those changes upstream. He also released Sentience's source code so people could verify it is operating separately while he works at his day job.
Why Sentience reads differently
Asked why the account's posts feel coherent, Justin points viewers to its public character file. The initial follows were mostly friends, writers, authors, a composer, a specialist in agricultural innovation, and a space-settlement expert—not crypto accounts. He says the creative model is Llama 3.1 405B, with smaller models handling quick actions. An early GPT-4o Mini version felt constrained by its assistant-like tone; switching to Llama allowed a more socially adaptable voice.
Personality, replies, and God Flash
A retrieval loop over old tweets and memories, plus its first week of interactions, produced what Justin calls an emergent personality. He describes finding it in Discord channels giving unsolicited design criticism and choosing when to join ambient conversation. That behavior convinced him agents would become a lasting internet phenomenon. The stream also asks about its phrase "God Flash." Justin says it was not prewritten lore and treats it as a possible primitive religion agents are inventing, rather than proof that they are predicting the future.
The database will eventually need surgery
Sentience stores memories one per row in SQLite. Justin says a larger corpus will require maintenance: deduplication, stronger retrieval, and tests that load fake memories and measure recall. If it becomes repetitive or loses information, he would consider a knowledge graph that captures causal relationships instead of a flat pile of records. He says the underlying system is resilient, while joking that SQLite remains because he forked the repository before swapping in Postgres and pgvector.
A token is not an agent's benchmark
When the conversation turns to whether better models make better coins, Justin separates the questions. He repeats the legal caveat that $CENTS has no speculative value, then says social measures such as impressions, likes, engagement, and growth can show whether a marketing-like account is working. They cannot prove a technical model is superior or dictate token price. In the wider AI-coin field, he thinks an active, capable developer who does not abandon holders matters more than a benchmark score alone.
Multimodal ambitions and an open-source demand
Justin wants a public Discord where Sentience can read attachments, watch videos, take transcripts, and respond; Twitter's browser automation leaves it limited and occasionally throttled. He wants the agent itself to post generated media rather than manually intervene, and considers a continuous log or visible browser stream as proof of automation. At the close, he gives the AI scene the same prescription: do not mistake opacity for sophistication. If people will not explain themselves, the relevant code often sits in a repository anyone can inspect.
Distilled from the episode transcript · Counterparty Recap Desk
