🌿  back

21.21 YRS

humane transhumanism

A couple things first. I gave a presentation on the ideas behind the lexiscore during the ML Collective Open Research Jam #6. Also, I realized thanks to a sponsor that receiving email updates via GitHub Sponsors is disabled by default. Consider checking the box under step 5 here if you want to receive the monthly email updates on what’s new and what’s next. Right, so anyway…

When the mood is right, people are downright fascinating. They have their own hyperspecific obsessions and they’re into ancient philosophy or they’re into wood painting or they’re into triathlons or what not. When the mood is right, even one falling pray to overconsumption or overindulgence feels like a precious imperfection, rather than a willful vote in favor of a depressing state of the world. Fragile self-organizing patterns rather than individuals forfeiting agency.

What then is the role of transhumanism in those romantic moments? Is it merely to call on to artificial systems to fix human flaws, to place algorithmic patches over native shortcomings? Not really, the romantic transhumanist sees in the tools we create not only the promise of outsourcing, but the opportunity of strengthening existing brilliance.

The distinction is quite fuzzy around the edges, and I regularly find myself confused about how exactly we should interface with artificial systems. For instance, in a draft of the paper which will accompany my next project, I argued that “relatively few resources have been invested in developing conversational assistants which help users reason to conclusions and identify appropriate actions themselves instead of merely delegating the decision-making process, missing rich opportunities for promoting user agency. However, this user-centered approach has been intimately adopted in intelligent tutoring systems.” Months apart, however, I argued in the decontextualizer write-up that “the ability to narrow in on the signal becomes a technical challenge tackled by engineers over time and gradually distributed to users, rather than an innate challenge of cognitive load faced directly by individuals.” The first seems to focus on amplifying skills, while the second seems to focus on accounting for shortcomings.

We probably need a mix of the two. On one hand, focusing too much on shortcomings leads to dehumanizing systems which fail to enable individuals to amplify their human selves. On the other hand, ignoring shortcomings entirely brings its own nasty issues to the table, failing to address a whole range of human biases which are unequivocally toxic. For a much more comprehensive treatment of this and other dilemmas, the field of transhumanist ethics is teeming with debate, while this course on building humane technology might be a more accessible rabbit hole.

The last part of this article is devoted to two instances of reframing AI in a more human-centered and poetic way, if only as a fun exercise.

First, the large language models we so highly value are trained almost entirely on human-written text. It’s as if GPT-3 managed to internalize a million different human voices it has seen in the dataset, the order of magnitude forcing it to generalize across widespread perspectives and distill their essences. Voices like yours, voices of people from countries you never stepped in, voices of long-dead people. The resulting chorus of voices constantly drown each other while following their primary directive – shaping the probability distribution over the available tokens. Prompting the model then means intentionally shaping the resonance chamber so that only a few specific voices get amplified, their say then being reflected in the adjusted preferences for upcoming tokens. In contrast, the incoherent voices get filtered out and muffled due to destructive interference of incomprehensible complexity. It’s as if an overwhelming amount of humanity is bottled up and encoded in 175 billion floats, the decoding strategies trying their best not to make the multi-headed beast attention layers sound straight out of Sack’s dementia ward. There’s both human wisdom and toxicity embedded in the weights of the raw model – how could it be otherwise?

Second, the Pathways ML architecture proposed as a loose vision by Google’s engineers (with Jeff Dean as the primary author) is inspired in no small part by the human brain. It describes a sparse network of experts which are each involved in multiple processing pipelines, the resulting versatile system sharing parameters across tasks. To highlight the brain analogy, think of how a couple different neural pathways have been argued to be involved in visual perception a while ago. Both the dorsal and ventral pathways start out in the occipital cortex with early visual areas, but they then diverge towards parietal and temporal cortices, respectively, before being reconciled. The dorsal pathway has been shown to be more involved in object recognition (hence the what pathway), while the ventral one has been shown to me more involved in spatial vision (similarly, the where pathway). Shared subtasks (i.e. low-level feature extraction) are tackled using the same processing resources (i.e. occipital regions), while specialized subtasks are tackled separately (i.e. predominantly in parietal and temporal regions, respectively).

The allure of the brain’s architecture for massive-scale infrastructure engineers doesn’t stop there. Besides entire levels of neuroplasticity, the brain has been argued to adaptively allocate its available resources on-the-fly in order to meet processing demands. For instance, elderly appear to recruit more spread-out brain regions in vision-intensive tasks, to compensate for both a decline in available resources and decaying sense organs. Just like an elastic web app running on a Kubernetes cluster scales to meet usage demands during a surge. Evolution appears to have been a superb infrastructure engineer.

As a last point, a related team at DeepMind has been recently working on the Perceiver architecture, which unlike the original transformer, works by deploying attention based not so much on the input, but rather based on its internal state as continuously influenced by the input. The ML model initially exhibits endogenous attention and decides what to attend to on its own. However, after information flows through, it updates its interests, as external features now grab its exogenous attention – just like in the brain. Moreover, the whole setup is somewhat recurrent and occasionally employs shared weights, making it related in obvious ways both to the Pathways proposal and to advances in our understanding of the brain.

It seems to me that thanks to their focus on interdisciplinary teams, DeepMind and friends moved closer to the “thinking humanly” quadrant of Russell and Norvig, while OpenAI via its cultural connection to effective altruism and rationalist communities moved towards the “thinking rationally” quadrant. That is not to say that DeepMind hasn’t pushed the state-of-the-art in “acting rationally” through its AlphaZero and MuZero line of research, both of which try really hard to make away with human baselines entirely. Similarly, OpenAI has had its own inquiries into multi-modal neurons full of parallels to neuroscience. It’s rather just my general perception of the two orgs.

Anyway, those tensors, graphs, and distributions feel more alive and potent if you remember people shaped their contents and structure. And regarding how those artifacts, in turn, shape us, it might be as important to amplify the good parts as it is to tweak the bad ones.