Why the Brain Consolidates: Predictive Forgetting for Optimal Generalisation, Fountas et al. (2026), arXiv (link)
Impressively broad investigation of how neocortex might selectively retain some details of a memory while leaving the others encoded by the hippocampus, using offline replay to train the hippocampus (and other downstream areas) to respond to increasingly compressed neocortical representations. This compressed representation addresses the “fidelity-generalisation conflict” plaguing state-of-the-art LLMs (how to draw on particular examples and details when necessary, but solve novel problems too), something that humans and other animals don’t seem to struggle with. A nice example of neuroAI where a genuinely informative neuroscience paper also has genuine insight for AI development as the latter field turns its attention further towards continual learning: a rare example of a paper with a foot equally in both camps. Replay, representational drift and generalisation are some of my favourite topics and it’s a joy to see them brought together like this. As a hippocampus researcher, I’m happy to see attempts at specific predictions about how these computations map onto cortico-hippocampal physiology, and equally happy that it leaves a lot of work there to do.
It faithfully follows the centuries-long trend of drawing analogies between the human brain and whatever the latest technology is (here, the key-value (KV) caches of LLMs are analogous to episodic memory stores). But the authors do make a convincing argument that a temporally decoupled learning process, usually thought of as being necessary only because of biological constraints in animals, could be useful in AI and expand the capabilities of KV cache storage.
Tags: neuro-ai, computational-neuroscience
The rise of tabular foundation models, Christoph Molnar (link)
This five-part blog series is a really good introduction to tabular foundation models, by someone who is optimistic about their current state and future advances.
Although discussions about artificial general intelligence focus on LLMs and how well they replicate human performance on tasks, for me, much more interesting is this kind of foundation model which can spot patterns in anything, and give insights that otherwise wouldn’t be possible - either with human expertise, models trained on limited data, or a combination of both. Being inherently text-based means LLMs are well designed for capturing, using and reproducing information that has been written by and for people, because we communicate with each other in natural language. But there is a whole world of information that is best captured in non-linguistic terms. And I suspect that we could get a lot more mileage from advanced models that can work intelligently with this kind of data, precisely because people are not so good at working with it: this is genuinely additive to human capabilities.
Tags: data-science
AI Agents in Data Science Competitions: Lessons from the Leaderboard, Justin Chung Clark, DrivenData (link)
Along with the better-known Kaggle, DrivenData is a platform that hosts competitions for data scientists to compete against each other to get the best metrics on some task and win prizes. This article describes the outcome of a simple experiment setting off several frontier models to participate with minimal prompting, and their conclusions reflect what I saw myself when I tried vibe-coding a solution to a data science competition (I didn’t submit anything; I was just curious to see what the outcome would be). The models did pretty well, sometimes achieving well over the 90th percentile on the leaderboard compared to human competitors. But they took a lot of effort (tokens) to get there. What I saw was a kind of context entrenchment where the LLM persisted with a strategy or assumption that may not have worked, being unable to correct is course, which probably aligns with what the author describes as plateauing before reaching the top of the leaderboard. The experiment did not test what I would consider state of the art: in my view, that necessitates more sophisticated prompting or context engineering - but that is likely to be, in the near future, what standard data science work is.
Tags: data-science, software-engineering
Large language mistake, Benjamin Riley, The Verge (link)
Riley highlights a recent divergence in which influential (neuro-)AI researchers are moving beyond large language models in the pursuit of artificial general intelligence, while Big Tech leaders continue to predict that their LLMs are only around the corner from it. Siding with the former, he highlights research showing that language in humans is “a tool for communication rather than thought”. I agree with his take, but my own diagnosis would be that the AI field has a tendency to focus on the outcome (human-level or superhuman intelligence) rather than the means to get there. The fluent linguistic output of tools like ChatGPT is deceptive because we associate fluency and articulateness with intelligence in people, but it is curiously divorced from learning capacity: humans show a great deal of independent intelligence in the earliest years, when langauge ability is lowest, which slows and solidifies in later years when adults are typically more linguistically fluent. More than a decade ago during my undergraduate degree, I did a research project on whether gesticulation can aid thinking, linked to the Global Workspace Theory that posits a definitively non-verbal basis for human intelligence. And, of course, there is a huge degree of intelligence in non-human animals which lack language altogether. This doesn’t mean that artificial intelligence has to mimic the way that biological intelligence works, but in my view this fundamental dissimilarity is important for one big reason: the benefit of LLMs is the ability to leverage the vast amount that humans have written, but it has been written for other human beings who will have spent years learning about the world first and building up a large amount of context. The data that is most effective for communicating between people (language) will not be the data that is most efficient for communicating between silicon-based deep neural network models, unless they have somehow been equipped with similar context. Answering a question coherently (i.e. communicating) is only a good benchmark if the goal is to communicate, not if it is something broader. The fallacy is conflating communication with other cognitive functions.
Tags: neuro-ai
The AI Shift: Is hiring becoming less meritocratic?, John Burn-Murdoch & Sarah O’Connor, Financial Times (link)
This installation of the FT’s newsletter “The AI Shift” analyses the fascinating example of AI making tasks easier for both recruiters/employers (writing compelling job descriptions) and job candidates (writing compelling applications), but creating a situation that benefits neither side. This is the result of tech companies developing a solution in search of a problem: to the extent that hiring is a “problem”, generative AI has produced not a solution but a set of satisficing behaviours. I would love to see more systems-level analyses of the impacts of AI like this: critiques often come in the form of worker productivity, tradeoffs of speed vs accuracy when generative AI replaces manual tasks, resource use, without asking questions about how this affects things at a bigger scale. What happens to worker productivity in the long-term if the speed of execution increases so much that workers burn out? Does the resource-intensiveness of generative AI result in faster climate transition to meet the energy demands now that wealthy technology companies have a bigger incentive to adopt it; or does it result in efficiency innovations in hardware and software that bring down the costs of generative AI and fuel its demand even more?
Tags: ai-ethics
Free Agents: How evolution gave us free will, Kevin Mitchell, 2024
“We may say, ‘I wish I were more confident like Gary,’ […] We rarely think, ‘I wish I were more sociable like a naked mole rat,’ […] There seems to be little concern that we are constrained to behave like generic human beings, but strangely more worry that we are constrained to behave like ourselves.” In this book, Mitchell frames the question of free will as the extent to which our actions are guided by internal constraints rather than external ones. He identifies biological complexity as the core underpinning for our free will: it gives rise to a persistent self that can set future goals and act in accordance with them, as well as a sophisticated set of inhibitory mechanisms that serve as strong internal constraints on what we do. For him, it is precisely these internal constraints that give us free will, liberating us from control either entirely by external ones (as simpler organisms are) or unbridled randomness. The treatment of some philosophical nuances is lightweight, as is his treatment of some of the psychological literature, but he is true to his background as an evolutionary neuroscientist and brings an interesting and intuitively appealing perspective. Top-down organisation is a cause just as much as bottom-up mechanisms, and therein lies the agency. He leaves us with a final take on AI: agency is crucial, so artificial systems will never understand the world without causal interventions (which cannot be learned passively), goals, and indeterminacy. What he doesn’t mention is that biology also equips us with a lot of inductive biases that enable this understanding of the world, and current state-of-the-art AI research tends to ignore that.
Tags: philosophy-of-mind
Slow Productivity: The lost art of accomplishment without burnout, Cal Newport, 2024
After popularising the term “deep work” in a previous, somewhat cult book, Newport turns his attention to what productivity really is (roughly: achieving goals), how to differentiate productivity from its antithesis pseudoproductivity (doing visible stuff to appear busy, regardless of whether it helps achieve those goals), and his three principles for “slow productivity”. Slow productivity is the philosophy that hgh-quality work that’s worth doing takes time, and not just determination and a long to-do list. He calls on examples of famous over-achievers from history to demonstrate the value of his proposed approach, and while this does resemble cherry-picking or survivorship bias, it works well as a pep talk. I have been reflecting lately on whether the general expectation that everything should be done unstoppably faster, in many areas of life and society, eventually hist a wall of human cognitive limits. Newport’s three principles align with my concerns: he recommends doing fewer things, working at a natural pace, and obsessing over quality. Current trends promoted by productivity gurus and company management (exacerbated by the arrival of LLMs) suggest a very different idea of what productivity is and how to achieve it. Perhaps productivity-as-a-goal becomes an example of Goodhart’s law.
Tags: productivity
A hardwired neural circuit for temporal difference learning, Campbell et al. (2025), bioRxiv (link)
An impressive and important paper, finally confirming the widely-held theory that ventral striatum transmits value signals to the ventral tegmental area, which in turn relays a temporal-difference error signal back to the striatum. The authors do a series of minutely detailed experimental studies (in a total of 74 mice!) to identify the exact transformations that each cell type in the circuit performs, bridging the gap from computational algorithms which are common in this subfield to biological filters that implement them.
Tags: systems-neuroscience, reinforcement-learning
Empire of AI: Inside the reckless race for total domination, Karen Hao, 2025
A perfect complement to Code Dependent (see below), telling the other side of the story: a biography of OpenAI that illustrates how the oppression felt by people across the world arises from Silicon Valley. Ego, financial interests, ideological hubris and an attitude of empire-building - along with the huge capital available in Silicon Valley - enabled a small handful of people to develop large language models without the safety checks or ethical considerations that are usually expected of such new technologies. Like Murgia’s message, Hao ends with a call for those affected by AI technologies to be included in its development from the start.
Tags: ai-ethics
Code Dependent: Living in the Shadow of AI, Madhumita Murgia, 2025
A thoughful and balanced book telling human stories about the good and bad side of AI, from data annotators to under-resourced doctors to the parents of troubled teenagers: some using AI, some helping to develop it, some subjected to it, and some fighting back. A few of the tales are familiar ones that are often repeated in the increasing number of books about AI ethics, but Murgia’s deliberate focus on collecting a global distribution of perspectives emphasises how, despite the potential for AI to improve people’s lives across the world, it is often used as a tool to oppress by governments and enterprises alike.
Tags: ai-ethics
Beyond the Supply Chain: Artificial Intelligence’s Demand Side, Alicia Solow-Niederman, SSRN (link)
Found via Luiza Jarovsky’s AI governance newsletter.
Solow-Niederman highlights some privacy issues that can arise via the (often innocent) interaction between a user and a generative AI system, which she argues are often overlooked. For example, a user asked a chatbot to generate some code, ChatGPT generated a URL, and the URL happened to lead to a real photo of a person’s face on imgur. I recall (and now can’t find) a social media post from some months ago in which people asked ChatGPT (with “memory” enabled, such that it had access to a large volume of chats from the same user) where in the world it thought they were from and what it thought their native language was, and users reported pretty high accuracy. Is the way you speak to AI a privacy violation by itself?
Tags: ai-governance