In the world of computing and artificial intelligence, few topics generate more interest than the rise of so-called “basic models”. These models can be considered meta-AI, but not Meta-AI, if you know what I mean: systems that integrate large neural networks with even larger datasets. They are capable of processing many things but, more importantly, they are easily adaptable to all areas of information, shortening and simplifying what was previously a laborious process of training AI systems. If the basic models live up to their promise, it could bring AI to much wider commercial use.
To give an idea of the breadth of these algorithms, GPT-3, a basic model for natural language processing released two years ago, contains more than 170 billion parameters, the variables that guide functions within a model. Obviously, this is a lot of parameters and it gives an idea of the complexity of these models. Along with this complexity comes considerable uncertainty, even among designers, about how they work.
At a recent Stanford University conference, scientists and engineers described how the arrival of the base models was made possible by substantial advances in hardware engineering that reduced data processing costs by reducing the time and energy the system uses to manage while performing its data analysis. The result is that AI research has succeeded in creating generic models in the sense that they are essentially pre-trained using a huge single data set and can perform a variety of different tasks with relatively little programmer input rather than being tailored to a single task and dataset. An AI scientist compared it to learning to skate. If you can walk, you have most of the skills needed to skate; minor changes and a little practice are enough.
As one can imagine, a quantum leap such as this stirs controversy, starting with whether the very term “foundation model” signals an effort by a single institution – Stanford, which pioneered the year a research center on foundation models – to exercise intellectual hegemony. and epistemic closure on the field of AI. Envy and professional and institutional rivalries come into play (“Leave it to Stanford to pretend they have the foundation!”). However, this appears to be a mix of genuine “don’t rush” concerns and worries about how the terminology might affect the distribution of investment capital and research support.
Others are confused by the basic models because any flaws or biases in these models (and in everything equally important, it will be impossible to entirely eliminate biases and errors) are likely to reproduce themselves in the systems and descendant apps. If, for example, some sort of racial bias is present in a base AI model (remember Controversy over the “unprofessional hair” algorithm?), this bias would be embedded in other systems and tasks potentially manifesting in discriminatory outcomes. Like a DNA mutation, these flaws, replicated in many AI systems, could metastasize and become devilishly difficult to fix and eliminate. The argument goes, we could end up with some of the worst instincts in the human world replicated in the virtual world, with consequences for real, living human beings.
Some AI scientists are also expressing concerns about the environmental impact. These basic models, despite their increased material efficiency, require huge amounts of electricity which must come from. . . somewhere. As they are replicated and commercialized, where will that power and the infrastructure to deliver it come from? (AI is not the only IT area where this is a problem; specifically, cryptocurrency mining is consume massive amounts of energy.) Although renewable energy sources are preferable, the rapid growth in demand likely means more coal, oil and natural gas in the near term, with a concomitant increase in carbon emissions as well as more power lines at high voltage, electrical substations and likes to deliver it. Then there is the shortage of rare earth elements needed to produce the necessary hardware systems. New sources must be found and exploited, with the associated impacts on the environment.
IIn this whirlwind of social, political, environmental, cultural and monetary anxiety, Stanford philosopher Rob Reich (pronounced “Rishe”, not “Rike” and not to be confused with the other Bay Area professor and former secretary of labor in the Clinton administration, Robert Reich), who helped co-found the university’s Institute for Human-Centered Artificial Intelligence, entered the conversation with an analysis of some of the broader ethical considerations the AI field should think about. At the recent Stanford meeting, Reich told the assembled industry scientists and scholars that his primary concern was not their individual moral compasses, but rather the virtual absence of conceptual and institutional frameworks to define and guide research. on AI.
He compared his concerns to those expressed by Albert Einstein about the age of the machine: these advances are the promise of a new era of prosperity and opportunity. Without ethical boundaries, it mentioned (borrowed from Einstein), it’s like placing “a razor in the hands of a three-year-old child”. Technological development always goes beyond ethical reflection, exposing society to dangers, if not from bad actors, perhaps from immature actors.
The current state of development of AI ethics, Reich said, resembled that of an adolescent brain: full of a sense of its own power in the world but lacking the developed frontal cortex necessary to restrain his less thoughtful impulses. Government regulation, he argued, is premature (we haven’t done the necessary ethical thinking about what such regulations should be). Moreover, the lag between the emergence of problems and the enactment of laws and regulations means that powerful AI tools will be widely used long before a regulatory framework is in place. As the government slowly moves towards a regulatory regime, AI developers must learn to control themselves.
Reich said that CRISPR, gene-editing technology, provides a contrasting template for developing and structuring ethics to go along with basic patterns. Reich Credits Jennifer Doudnawith who Emmanuelle Charpentier, received the 2020 Nobel Prize in Chemistry for developing CRISPR, for pioneering the effort to establish safeguards around gene editing. Reich told a story that Dudna tells in his memoirs and elsewhere of waking up from a nightmare in which Adolf Hitler had access to CRISPR technology. She immediately began organizing a network of scientists to build voluntary ethical controls to govern the field.
The key principle followed by the network was “no experimentation on human embryos”. Journals agreed not to publish papers by scientists who violated the ban, and scientific bodies promised to exclude such scientists from professional conferences. This self-government is no guarantee against bad actors, like the Chinese scientist He Jiankuiwho used CRISPR to produce genetically modified babies and lost his job and was sentenced to prison, but that’s a reasonable start.
Such clear professional standards, if they can be put in place quickly enough, could stop some types of problems in AI before they start. With enough self-policing, the world could buy the time needed to develop more detailed ethical exploration, laws, and regulatory standards to guide research and use of AI, and the institutions to enforce them.
This is a compelling argument because it recognizes that while the law is often a weak means of limiting bad behavior, peer standards can be an effective deterrent. More importantly, while ethical concerns about racial bias and the environment are important, Reich’s engagement with AI scientists at this higher and more general level of conversation underscores the importance of applying the Tenet of Jurassic Park to AI: Before doing something, it might be a good idea to ask yourself if you really should do it.