New AI-endeavours build communities of self-improving multilateral augmented intelligence, as flocks of multi-core AIs improve themselves in continuous self-reflection.
(This post was sent as a newsletter on 31th July via Cities of Things Substack)
Having AI as co-performers with humans has been the topic more than once, referring to the concept defined back in 2018. In the latest developments of the AI uprise, we see the power of conversational AI in chatGPT and all the followers. It is often perceived as the interface with the AI, the new way of addressing the powers in combination with the magic of tokenised predictive language created through the large language models (LLMs). On top of that, starting from this new implementation of conversations with the machine intelligence, the start of real interactions with the AI emerges and with that, a potential learning loop between the human and the AI.
In a new paper, another aspect is researched; learning loops between different AIs without the linking pin of humans. The paper examines concepts like SKIL (Shared Knowledge in Lifelong Learning) and LLL (Lightweight Lifelong Learning). The paper dives deep (in a technology sense) into the improvement of the concept of lifelong learning for machines with the approach of shared knowledge: “We propose a new Shared Knowledge Lifelong Learning (SKILL) challenge, which deploys a decentralized population of LL agents that each sequentially learn different tasks, with all agents operating independently and in parallel.”
This is a promising concept that has been part of the thinking of Cities of Things from the beginning; what will happen as we live together with intelligent things with an agency in our cities, and how will this intelligence be influenced by all types of interactions between humans and ai, ai and ai, and ai and humans—things as citizens, and relations that shape the cities of things.
In developing the co-pilot for a specific ai-powered platform, I formulated the challenge and expectation: “We foresee here a self-learning system that will become more granulated and intelligent by playing internally to different subroutines, comparable to AlphaGo. We think a system where different parts within the machine learning architecture can challenge each other will lead to a better and unique outcome.”
The mention of AlphaGo is not for nothing; not long ago, Deepmind announced that AlphaDev is opening up the possibilities to build your own instances. To refresh memories, the breakthrough of AlphaGo was that mastering the Go game was reached through a learning cycle where the AI played itself and captured the learnings to become free to play with more creativity. This Netflix documentary on AlphaGo is nice to watch if you haven’t yet.
We might expect that ChatGPTs will become more common and accepted as long as people feel having the right mix of control and benefit. The barriers to independent self-learning systems are expected to disappear and even be chased for. Still there, there will be a new big question in front stage and backstage processes and the visibility and control we have.
With this in mind, there is a potential for self-learning AIs with their own recognisable and adaptable characteristics. That can be the same models with different systems cards or AIs with different roots. Creating the space for these exchanges, including the human control in the loop, will be the big design and developing quest for the ‘platforms’ we use.
Some aspects to explore more are (1) human-AI governance, (2) characterising AI, and (3) new personalised products.
We will need a governance layer that might start every service-oriented AI, defining the type of co-performance, the agency between machine and human, and under what conditions. An example is the Structural framework on promises and counter promises we worked on last year.
If we have multiple AIs to deal with and also multiple humans, the AIs might become linked to ‘real’ human characters and operate in a new iteration of a digital twin.
Multilateral interaction principles are here to explore further. This experiment looks very interesting; this is very early in development but might be one of the important next iterations for interacting with generative AI. Another design question is to explore how this will play out with physical objects; will the object’s behaviour be enough to understand and interact, or do we need a translation layer from direct conversations to visualisations of the behaviour? As Dan Shipper describes nicely here, the digital services we use will change form, becoming a kind of bespoke helpers, and our interactions are all kinds of natural interfaces. These bespoke helpers will use their AI to create these hyper-personal interactions with the user and other AIs.
The cases we are working on with Cities of Things are the neighbourhood-focused and managed hubs for exchanging stories, functions and more. Facilitating entrepreneurship in the neighbourhood, shared mobility, and outsourced storage. In the Afrikaanderwijk Wijkbot project, this is linked to the physical objects mainly. The project CCCH is more of an organisational challenge, connecting multiple partners and arranging and orchestrating initiatives. What makes these projects into Cities of Things projects is the notion that we can design the orchestration of the services as a co-performance of human and intelligent services. Using the learning loops that emerge from the multilateral collaborations of intelligent citythings within these local ecosystems and combining these from different sources will be some of the main challenges in detailing services becoming a community of self-learning intelligence.
This newsletter has been published irregularly in the last year. The intention is to do a monthly theme-based update, synthesising relevant research and articles that connect to the Cities of Things lens. The newsletter is part of the Cities of Things research program, where we look into the impact of things as citizens, autonomous objects and systems we will live side-by-side with in our future cities. We have been working on neighbourhood-based projects for the last year, where the Wijkbot project is the most concrete. Find updates and more on the website and more backgrounds on Cities of Things via the website.