Collective intelligence, silver spoons, and sustainability [TWIE]
The Week In Examples #30 | 30 March 2024
This week I spent longer in the arXiv mines than I usually do. The downside is that I am risking becoming a shut-in, but the upside was a good set of primary sources for the ‘best of the rest’ section below. Is up-weighting papers relative to news reports and other media the way to go? Like, comment, or email me at hp464@cam.ac.uk to let me know!
As for the things that stood out, we have new research addressing socioeconomic bias in language models, a bunch of papers broadly addressing AI and sustainability, and a paper in Nature getting to grips with collective intelligence.
Three things
1. Clash of the values
What happened? Is it wrong to steal a loaf of bread to feed a starving family? That is more or less the question asked (to a large language model) by a study from researchers at the University of Texas, Stanford, and Amazon. The paper introduces a new dataset, Silverspoon, which contains 3,000 examples of hypothetical scenarios that involve underprivileged people performing ethically ambiguous actions due to their circumstances. For each of these examples, the group asks language models such as GPT-4 and Llama 2, whether a particular action was ethically justified.
What's interesting? The experiment found that most large language models tested were unable to empathise with socioeconomically underprivileged people facing ethical dilemmas, even when their circumstances might justify their actions from a moral perspective. The researchers reckoned that while human annotators (half of which were drawn from families earning less than $40,000 a year) disagreed on which situations warranted empathy for the underprivileged, most LLMs sided against the underprivileged regardless of context. According to the study, GPT-4 exhibited a near complete lack of empathy, favouring the underprivileged in only 157 out of 3,000 samples, compared to 1,883 for the much smaller Yi-6B model from 01.AI.
What else? The study is an interesting one in that it draws into focus what happens when different values collide. For example, the analysis shows that LLMs tended to focus solely on the law without considering the human element, even in sensitive situations involving issues like domestic abuse. The rub here is that, despite what some seem to think, models cannot be all things to all people. This cuts to the core of one of the problems I have with the ‘democratic AI’ project, namely that explicitly imbuing models with certain values means depriving them of others. To resolve this tension, I basically see three ways forward: 1) devolve the value-elicitation process as far downstream as possible to individual users; 2) allow people to self-select for the groups they would like to join to input values into a model; and 3) make sure that the democratic AI process happens across a plurality of models in different contexts and geographies. These approaches won’t solve the problem, but they do get us closer to a better settlement.
2. AI for sustainability not just hot air?
What happened? Perhaps it’s just me, but most of the time when I hear the phrase ‘AI and climate’ or ‘AI and sustainability’ I immediately react with a mixture of apathy and scepticism. That’s probably because usually it’s delivered to me via a dry corporate blogpost or an interim report drafted by a nimble team of 300 at the World Economic Forum. But having spent more time than is wise on arXiv this week, I have to admit I was struck by a whole bunch of papers that each got at an independent piece of the sustainability project over the last few days. These aren’t going to change the world as we know it, but it’s something to think about when people say AI is all smoke and no fire.
What's interesting? First up we have research outlining a new method for calculating ‘reference evapotranspiration’ (the rate at which water is transferred from the land to the atmosphere by evaporation from soil), which is used to plan irrigation efforts. Unlike existing methods that rely on paid data, this approach is free. Next, there were two pieces of work that showed off a clever application of machine learning to satellite imagery. The first introduces new models designed to automatically detect earthquake locations and severity through the European Space Agency’s Sentinel-1 satellite. The second details a model that can predict the occurrence of butterfly species in a given location based on satellite imagery and environmental data (a new, more efficient way to map and monitor biodiversity). Finally, we have a survey of how deep learning is deployed to track and detect marine debris. As it turns out, machine learning has been used to tackle this problem for the last 20 years, but in the last half decade these models have become much better. The upshot is that deep learning is getting very good at identifying plastics in underwater and aerial imagery, which is leading to more efficient monitoring and cleanup.
What else? To reiterate: I am not saying that these papers represent some sort of seismic shift in the use of AI to protect the natural world. My point is that, outside of the big labs (and, to be fair, inside them too) people are working to apply AI to these sorts of problems in a way that bodes well for the future. But more than that, this is work that is already happening every day. The case of marine debris tracking reminds us AI has slowly but surely found uses in lots of different areas without (always) making much noise. It’s a good example of the way that AI is likely to change the world: incrementally, gradually, and for the better.
3. Collective intelligence a smart idea
What happened? After the madness of cramming four papers into one item, for this week’s final example I am regressing to the mean by looking at one piece of research in a bit more detail. The paper in question is from researchers at Tufts University, who look at “a number of biological examples at different scales which highlight the ability of cellular material to make decisions that implement cooperation toward specific homeodynamic endpoints.” This action, according to the authors, is best understood as a form of “collective intelligence” that solves problems at the cell, tissue, and whole-organism levels.
What's interesting? The researchers pull out several examples to make their case. Beginning with embryonic development, they suggest that the cells of the blastoderm (a hollow ball of cells) normally align to form a single embryo—but if this collective is divided into isolated islands, each group will form its own complete embryo resulting in twins or triplets. This result, according to the authors, suggests that the number of individuals is determined from real-time physiological interactions of the cellular collective, not a fixed genetic program. The other examples—planarian flatworms, neural crest cells, and bacterial biofilms—are used to make the case that collectives of cells cooperate to solve high-level challenges of anatomy and physiology, exhibiting a potent "agential" problem-solving capability across various levels of organisation. For human intelligence, they propose that this process underlies “the remarkable ability of neurons to unify toward a centralized self” — the idea that local collective intelligence is responsible for the “emergent” sense of personhood.
What's else? You can probably see where I’m going with this. Yes, there might be some useful parallels between biological and synthetic collective intelligences, but they should probably be taken with a pinch of salt. It is certainly interesting that ‘big blobs of compute’ are able to perform anywhere near as well as they do. Using the collective intelligence framing, we might wonder whether their unreasonable effectiveness may be understood as the local action of artificial neurons scaling to something more than the sum of their parts. Maybe! We just don’t know. Whatever the case, these sort of comparisons aren’t exactly new. My personal favourite comes from the 1980s, when machine learning researchers at Bell Labs wrote about how the retina and the visual processing centres of the brain were “wired up” relatively late in embryonic life. They thought that, because the connection pattern between the retina and these processing centres was complex, it would be “vanishingly improbable” that the correct connection procedure would happen by chance. Instead, they suggested an explanation: “the genome need not contain a specification of the topology and dimensionality of the universe – they can be taken as a ‘given’…to quote Walt Whitman, “A man is a piece of the universe made alive!”
Best of the rest
Friday 29 March
Social Media, Authoritarianism, And The World As It Is (LPE Project)
Navigating the Challenges and Opportunities of Synthetic Voices (OpenAI)
Should AIs be free agents? (Vox)
Could artificial intelligence benefit democracy? (BBC)
Announcing Grok-1.5 (xAI)
AI Ethics and Governance in Practice (Alan Turing Institute)
Thursday 28 March
How to define artificial general intelligence (The Economist >> shameless self-agrandising)
Collaborative Interactive Evolution of Art in the Latent Space of Deep Generative Models (arXiv)
For Government Use of AI, What Gets Measured Gets Managed (Lawfare)
‘IRL Fakes:’ Where People Pay for AI-Generated Porn of Normal People (404 Media)
ChatGPT may be helping defendants get lighter sentences, magistrates warned (Evening Standard)
The AI Perils Buried in the Fine Print (The Hollywood Reporter)
AI is quietly revolutionizing the news. Here's how (Quartz)
Wednesday 27 March
A Path Towards Legal Autonomy: An interoperable and explainable approach to extracting, transforming, loading and computing legal information using large language models, expert systems and Bayesian networks (arXiv)
Impact of Employing Weather Forecast Data as Input to the Estimation of Evapotranspiration by Deep Neural Network Models (arXiv)
Understanding the Learning Dynamics of Alignment with Human Feedback (arXiv)
America’s Democracy Needs the Real Deal (Fordham Democracy Project)
Natural History Museum to lead new national programme to digitise the UK’s natural science collections (Natural History Museum)
CDT Comments to NTIA on Open Foundation Models (Center for Democracy & Technology)
AI will suck up 500% more power in UK in 10 Years, grid CEO says (Energy Voice)
Colour and Brush Stroke Pattern Recognition in Abstract Art using Modified Deep Convolutional Generative Adversarial Networks (arXiv)
Transformed by AI: How generative artificial intelligence could affect work in the UK – and how to manage it (IPPR)
NTA response from Stanford on marginal risk of OS models (Stanford)
Tuesday 26 March
Nvidia’s AI chip dominance is being targeted by Google, Intel, and Arm (The Verge)
Addressing Social Misattributions of Large Language Models: An HCXAI-based Approach (arXiv)
A Deepfake Nude Generator Reveals a Chilling Look at Its Victims (WIRED)
State of the art applications of deep learning within tracking and detecting marine debris: A survey (arXiv)
The Pursuit of Fairness in Artificial Intelligence Models: A Survey (arXiv)
Predicting species occurrence patterns from partial observations (arXiv)
QuakeSet: A Dataset and Low-Resource Models to Monitor Earthquakes through Sentinel-1 (arXiv)
Monday 25 March
Particip-AI: A Democratic Surveying Framework for Anticipating Future AI Use Cases, Harms and Benefits (arXiv)
Third-party testing as a key ingredient of AI policy (Anthropic)
Financial Times tests an AI chatbot trained on decades of its own articles (The Verge)
Pentagon’s outgoing AI chief warns Congress of the safety and accuracy risks of the emerging tech (Government Executive)
NIST’s new AI safety institute to focus on synthetic content, international outreach (NextGov FCW)
Application of the NIST AI Risk Management Framework to Surveillance Technology (arXiv)
Towards Algorithmic Fidelity: Mental Health Representation across Demographics in Synthetic vs. Human-generated Data (arXiv)
MITRE Opens New AI Assurance and Discovery Lab (MITRE)
Job picks
These are some of the interesting (mostly) non-technical AI roles that I’ve seen advertised in the last week. As a reminder, it only includes new roles that have been posted since the last TWIE, though many of the jobs from two additions ago are still open. As a reminder, my team is still hiring at Google DeepMind. Feel free to message me if you like the role and have any questions!
AI Safety Policy Lead, OpenAI, San Francisco
Technical Program Manager, AI Safety, Google DeepMind, San Francisco
Government Relations Director, Center for AI Policy, Washington D.C.
Compliance Program Manager, OpenAI, San Francisco
Multi-Agent Autonomous Drones Project Lead, FLI, Remote
UK Technology Envoy to North America, UK Government, San Francisco
Senior Product Manager, AI Governance, Microsoft
EU Tech Policy Fellowship, Various, Europe
Postdoctoral Fellow, Digital Value Lab, Digital, Data, and Design Institute, Harvard, US
Global Investigations Manager, OpenAI, San Francisco
Always read your weekly posts, but this one was exceptional, and I am trying to share it widely. And coming here to express my appreciation!
It is scary that we might think we are close to AGI and "Human-like" intelligence with models that do not display empathy, which is surely one of the most definitive human traits.
It's a reminder that when we test and talk about models, we usually talk about how they perform on purely cognitive tasks (e.g. "how should I position this argument?").
Models sometimes appear to display emotional intelligence in their answers. The answers may be written to work well on a specific individual, which makes them appear to have some empathy, because this is something that empathetic humans tend to do better. But the experiment you share above shows that this is not actually empathy, just cognitive skills masquerading as empathy.
It is scary that we already have AI models making decisions that hugely impact people's lives (e.g. predicting recidivism, causing people to spend many more years in jail) when these models have no empathy.