Shortly before this gets posted, Jon Crowcroft and I will have presented this year’s offering at Gikii, the weird little conference that crosses law, media, technology, and pop culture. This is what we will possibly may have said, as I understand it, with some added explanation for the slightly less technical audience I imagine will read this.
Two years ago, a team of four researchers – Timnit Gebru, Emily Bender, Margaret Mitchell (writing as Shmargaret Shmitchell), and Angelina McMillan-Major – wrote a now-famous paper called On the Dangers of Stochastic Parrots (PDF) calling into question the usefulness of the large language models (LLMs) that have caused so much ruckus this year. The “Stochastic Four” argued instead of small models built on carefully curated data: less prone to error, less exploitive of people’s data, less damaging to the planet. Gebru got fired over this paper; Google also fired Mitchell soon afterwards. Two years later, neural networks pioneer Geoff Hinton quit Google in order to voice similar concerns.
Despite the hype, LLMs have many problems. They are fundamentally an extractive technology and are resource-intensive. Building LLMs requires massive amounts of training data; so far, the companies have been unwilling to acknowledge their sources, perhaps because (as is happening already) they fear copyright suits.
More important from a technical standpoint, is the issue of model collapse; that is, models degrade when they begin to ingest synthetic AI-generated data instead of human input. We’ve seen this before with Google Flu Trends, which degraded rapidly as incoming new search data included many searches on flu-like symptoms that weren’t actually flu, and others that simply reflected the frequency of local news coverage. “Data pollution” as LLM-generated data fills the web, will mean that the web will be an increasingly useless source of training data for future generations of generative AI. Lots more noise, drowning out the signal (in the photo above, the signal would be the parrot).
Instead, if we follow the lead of the Stochastic Four, the more productive approach is small data – small, carefully curated datasets that train models to match specific goals. Far less resource-intensive, far fewer issues with copyright, appropriation, and extraction.
We know what the LLM future looks like in outline: big, centralized services, because no one else will be able to amass enough data. In that future, surveillance capitalism is an essential part of data gathering. SLM futures could look quite different: decentralized, with realigned incentives. At one point, we wanted to suggest that small data could bring the end of surveillance capitalism; that’s probably an overstatement. But small data could certainly create the ecosystem in which the case for mass data collection would be less compelling.
Jon and I imagined four primary alternative futures: federation, personalization, some combination of those two, and paradigm shift.
Precursors to a federated small data future already exist; these include customer service chatbots, predictive text assistants. In this future, we could imagine personalized LLM servers designed to serve specific needs.
An individualized future might look something like I suggested here in March: a model that fits in your pocket that is constantly updated with material of your own choosing. Such a device might be the closest yet to Vannevar Bush’s 1945 idea of the Memex (PDF), updated for the modern era by automating the dozens of secretary-curators he imagined doing the grunt work of labeling and selection. That future again has precursors in techniques for sharing the computation but not the data, a design we see proposed for health care, where the data is too sensitive to share unless there’s a significant public interest (as in pandemics or very rare illnesses), or in other data analysis designs intended to protect privacy.
In 2007, the science fiction writer Charles Stross suggested something like this, though he imagined it as a comprehensive life log, which he described as a “google for real life”. So this alternative future would look something like Stross’s pocket $10 life log with enhanced statistics-based data analytics.
Imagining what a paradigm shift might look like is much harder. That’s the kind of thing science fiction writers do; it’s 16 years since Stross gave that life log talk. However, in his 2018 history of advertising, The Attention Merchants, Columbia professor Tim Wu argued that industrialization was the vector that made advertising and its grab for our attention part of commerce. A hundred and fifty-odd years later, the centralizing effects of industrialization are being challenged starting with energy via renewables and local power generation and social media via the fediverse. Might language models also play their part in bringing a new, more collaborative and cooperative society?
It is, in other words, just possible that the hot new technology of 2023 is simply a dead end bringing little real change. It’s happened before. There have been, as Wu recounts, counter-moves and movements before, but they didn’t have the technological affordances of our era.
In the Q&A that followed, Miranda Mowbray pointed out that companies are trying to implement the individualized model, but that it’s impossible to do unless there are standardized data formats, and even then hard to do at scale.
Illustrations: Spot the parrot seen in a neighbor’s tree.
Wendy M. Grossman is the 2013 winner of the Enigma Award. Her Web site has an extensive archive of her books, articles, and music, and an archive of earlier columns in this series. She is a contributing editor for the Plutopia News Network podcast. Follow on
3 thoughts on “Small data”