“The weirdness right in front of us”. In trying to ensure AIs align with human interests, can we be open to learning new things about them - and ourselves?
Photo by Valentin Petkov on Unsplash. Hackaton, Manchester, United Kingdom
We find our feeds overwhelmed with speculation about Artificial Intelligence these days (a long standing theme of interest here). No doubt these are triggered by the startling achievements of what is becoming widely known as Generative AI - the image makers that create moody masterpieces out of textual prompts (known as ‘text-to-X’), the chatbots (notoriously ChatGPT) [link to service] that talk back to you as convincingly as a therapist (or academic, or favourite poet…).
But the pieces triggered are deeply philosophical and questioning - as if the exponential explosions of intelligence, real or simulated, is shaking us into imagining what it might be like to live in a world with such presences. If we don’t grant them consciousness and agency, then how do we place them into “service” to worthwhile, not self-destructive goals? (The Fully Automated Luxury Whatever.) If we decide they do have agency and consciousness at some point, are we in a similar situation as the expanding reach of animal rights? Who has the right to “switch off” both forms of sentience?
Here’s a small tour d’horizon of what’s come to us this week:
Futurist and academic Andy Miah - a friend of the AG - conducted his own conversation with ChatGPT a few days ago, asking it direct questions about its impact on humanity. Some of the exchange below:
You suggest that artificial intelligence is only ever a product of what it is programmed to do, but can’t artificial intelligence become its own authority and go beyond what we programme it to do?
While it is true that artificial intelligence is a product of what it is programmed to do, it is also possible for AI systems to learn and adapt over time. This means that AI systems may be able to go beyond their initial programming and develop their own behavior and decision-making processes. However, this does not necessarily mean that AI systems will become their own authority and operate independently from human control.
In most cases, AI systems will still be subject to the constraints and limitations of their design and implementation, and will ultimately be guided by the goals and objectives that have been set for them by their human creators. While it is important to carefully consider the potential implications of advanced AI systems, it is unlikely that AI will become a truly independent and autonomous entity.
How will we know when artificial intelligence is determining its own aims, beyond what humans ask it to do?
It is unlikely that artificial intelligence will ever reach a point where it is able to determine its own aims beyond what humans ask it to do. While AI systems may be able to learn and adapt over time, they will always be subject to the constraints and limitations of their design and implementation.
Furthermore, the development of advanced AI systems is closely tied to the goals and objectives of their human creators, so it is unlikely that AI systems will be able to develop their own aims that are completely independent from human input.
You get a sense of what some significant AI figures are saying about this stage of the machinery - which is that it’s involved in some kind of modelling of the mind of the human user, the answers essentially tailored to responding satisfyingly to the most recent question put. Not too bad, if the AIs are designed to make us happy…
The interesting question, as Benjamin Bratton and Blaise Agüera y Arcas put it recently in Noema, is this:
Instead of a groundhog-day rehashing of debates about whether machines have souls or can think like people imagine themselves to think, the ongoing double-helix relationship between AI and the philosophy of AI needs to do less projection of its own maxims and instead construct more nuanced vocabularies of analysis, critique, and speculation based on the weirdness right in front of us.
We’re inclined to stay with the weirdness a little longer - not doing what Bratton and Arcas warn us against, which is “premature ontologisation”. IE, we don’t know quite what’s the nature of this new stage of AI’s inner and outer reality is, yet - and we should let it unfold some more.
Photo by Gertrūda Valasevičiūtė on Unsplash
Of course, the anxieties about “alignment” - whether we can embed these rapidly advancing machines with a respect for human life, and the best of our values - are entirely legitimate. This Quanta piece covers many of them - and reveals entertainingly that there are “Orthodox” and “Reform” positions. (The Orthodox worry about the machines sparking to life, then conspiring against us; Reformers think their powers could be at the service of human malevolents well before that). There is a bit of software that might help here:
Many in the alignment community think the most promising path forward is a machine learning technique known as inverse reinforcement learning (IRL). With IRL, the machine is not given an objective to maximize; such “inserted” goals, alignment proponents believe, can inadvertently lead to paper clip maximizer scenarios. [Where a powerful AI is given a goal to make paper clips, and subsumes the whole world to that end…]
Instead, the machine’s task is to observe the behavior of humans and infer their preferences, goals and values. In recent years, researchers have used IRL to train machines to play video games by observing humans and to teach robots how to do backflips by giving them incremental feedback from humans (people viewed short clips of a robot’s various attempts and chose the one that looked best).
It’s unclear whether similar methods can teach machines the more subtle and abstract ideas of human values. The writer Brian Christian, author of a popular science book about AI alignment, is optimistic: “It’s not such a stretch to imagine replacing the nebulous concept of ‘backflip’ with an even more nebulous and ineffable concept, like ‘helpfulness.’ Or ‘kindness.’ Or ‘good’ behavior.”
Another fascinating conversation we observed was from the integral thinkers at Integral Life, coming together and posing what they called a “(Ma)Turing Test” to a GPT chat-bot (punning on the Turing Test, the classic proof of AI intelligence). As they explain below, setting up their Saturday event:
Developmental psychologists have demonstrated that a human being grows through a successive progression of increasingly complex stages of meaning-making.
These stages of ego/self-identity development can be thought of as coherent structures or patterns of how the “self” of each stage organizes and navigates itself in relation to the world (e.g., think of how the 3 year old self organizes itself and its world versus the 12 year old self versus the 30 year old self, and you’ll begin to appreciate the progression of qualitatively more advanced maturity that occurs).
Maturity of a given adult’s ego/self-identity is assessed through language, specifically how they reply to a set of specific sentence prompts, prompts like “A true friend…“, “My mother and I…“, “Rules are…” and a few dozen others.
Using this same method, we just assessed GPT-3, the world’s most advanced sentence-generating Artificial Intelligence [er, not quite], having it complete sentence prompts for the Maturity Assessment Profile developed by Dr. Susanne Cook-Greuter.
This test and its resulting data has been collected on hundreds of thousands of sentence replies over the course of four decades, and shows that more mature “selves” evidence a growing sophistication, self-awareness, sensitivity and perspectival fluidity as they progress to later, more complex stages of self-development.
So what is the maturity level demonstrated by the world’s most advanced AI? It is impressive, promising and thought-provoking…
Tantalising. If their results become public, we’ll post them here.
And finally, a video from our long-standing friends London Futurists - on AGI (artificial general intelligence) and the future of ethics. Blurb beneath the embed:
“When artificial intelligence exceeds human thinking in all categories of reasoning and understanding, what conclusions will it reach about the future of ethics? Will such systems - AGIs - take greater care of us humans, than the regard we show to rodents? To what extent can design choices made by human developers influence the decisions that AGIs will take? Or is any such discussion premature or misguided, given apparently more pressing problems facing human civilisation as 2023 approaches?
“This London Futurists webinar took place on 17th December 2002 and featured the ideas of Daniel Faggella, the founder and CEO of Emerj Artificial Intelligence Research. Daniel has researched and written extensively on topics such as:
A forthcoming "moral singularity"
Scenarios for the emergence of AGI
Why ideas of "friendly AI" are fraught with difficulty
Possible trajectories for posthumans in the wake of advanced AI”
***
We’ll be keeping a regular weekly eye on AI talk, particularly as it may assist and develop human communities, from now on.