Institutional vs personal AI
Comparing Eunice and the aunties as two paradigms of agentic AI.
I keep circling back to William Gibson's Jackpot novels and more specifically Agency, because of their relevance to the current state of AI (and the world at large).
As background, Agency deals with multiple realities. In the canonical reality, we find ourselves in a dystopian future London, where technological progress has become detached from human prosperity. Most of the world's population has died out, yet humanity possesses advanced communications, nanotechnology, and AI.
The communication technology is so advanced that it allows the future to establish links with the past. Once such a link is established, the past timeline forks off, and becomes independent of the future it communicates with. Gibson calls this forked-off reality, a stub. One such stub is an alternate 2017 San Francisco, where much of Agency takes place.
And this is where we meet our two AIs.
In 2017 San Francisco, we meet Eunice, an apparently self-aware and adaptive AI agent. She very much resembles a GPT-Live model. She has real-time voice capabilities, is multimodal, and can spawn parallel sub-agents with tool calling capabilities. She has memory and displays signs of empathy and care for her user. I've previously argued that Eunice is almost here.
In future London, we encounter another form of AI, idispensibly utilised by a powerful law enforcement figure, Lowbeer, who refers to these systems as the aunties. The aunties are highly capable AI systems operating across Lowbeer's information and institutional networks. However it's the manner in which they are referred to, and their largely invisible presence, that really makes me think. What if Gibson was describing not just two kinds of AI, but two fundamentally different relationships between AI, humans, and institutions?
If we consider Lowbeer's aunties, they are embedded in a world where power is already highly concentrated. They extend Lowbeer's ability to observe, understand and act; and they are largely invisible. Eunice, by contrast, starts with Verity, an ordinary person testing an AI product in 2017 San Francisco. Yet their relationship evolves into something much more consequential.
The aunties amplify existing agency, while Eunice appears to generate new agency. She gives an ordinary individual access to capabilities that would otherwise require an entire organisation. The difference isn't necessarily how intelligent or autonomous these systems are, but how they distribute power.
For me personally, and for others building with frontier agentic AI, we now face a choice. This choice is less about intelligence or architecture, and more about ownership, authority, and alignment.
Should we opt for the auntie model: build a sophisticated intelligence layer around an existing organisation, give it access to institutional knowledge, tools, permissions and workflows? Its capabilities derive partly from the system it's embedded in.
Or should we go for the Eunice model: start with an individual and build an increasingly persistent, autonomous intelligence around them? It learns their context, coordinates resources, forms relationships and acts on their behalf.
But perhaps these two models aren't mutually exclusive. An agent could have the personality of Eunice and the institutional allegiance of an auntie. It could appear deeply personal while operating entirely within infrastructure owned and controlled by someone else.
There is also the question of dependence. An AI like Eunice could dramatically expand an individual's agency, but as its capabilities become essentialy, that individual may become increasingly dependent on it. The very technology that makes us more capable could also make us less independent, particularly if we don't control it.
We may already be building the intelligence of the aunties while aspiring to the agency of Eunice. The technology required to understand, remember, reason and operate tools is increasingly recognisable. The harder problem is deciding who grants an agent authority and whose interests it represents, and who ultimately controls it.
An AI agent might know you intimately, remember years of conversations, anticipate your needs, and act independently on your behalf. But if your employer owns the infrastructure, controls its permissions, and can switch it off, is it fundamentally the employee's Eunice, or the company's auntie?