Reading

  • Dear diary, today the user asked me if I’m alive

    • I wouldn’t want ChatGPT to write my diary. It’d be the AI’s interpretation of your hopes and dreams. It’d be a simulation of an estimate of my plans and ideas. But I’d love to know more about what’s actually going on inside the AI’s head. What if I gave the AI somewhere secret to write about its feelings?

  • Simulation, Consciousness, Existence (1998)

    • Like organisms evolved in gentle tide pools, who migrate to freezing oceans or steaming jungles by developing metabolisms, mechanisms, and behaviors workable in those harsher and vaster environments, our descendants, able to change their representations at will, may develop means to venture far from the comfortable realms we consider reality into arbitrarily strange worlds. Their techniques will be as meaningless to us as bicycles are to fish, but perhaps we can stretch our common-sense-hobbled imaginations enough to peer a short distance into this odd territory.

  • Robot Dexterity Still Seems Hard

    • While a lot of these capabilities are impressive, robot progress still seems somewhat uneven to me. It’s cool to see these robots move in such human-like ways, but as former OpenAI Chief Research Officer Bob McGrew notes, “Manipulation is the hard problem we need to solve to make humanoid robots useful, not locomotion.” The value of a humanoid robot isn’t whether it can dance, run, or flip, but how capable it is at manipulating objects in the real world. And while manipulation capabilities are improving, they appear to have a very long way to go.

    • What robots have traditionally struggled with isn’t controlled, precise movements, but dexterity, which is something like “the ability to manipulate a broad variety of objects in a broad variety of ways, quickly and on the fly.” Humans can complete almost any object manipulation task you ask them to do — folding a piece of clothing, opening a gallon of milk, wiping up a spill with a cloth — even if it’s an object and/or task they’ve never encountered before. With robots, on the other hand, while it’s usually possible to automate any specific task (given the right hardware, enough time, and a narrow enough task definition), building a robot that can flexibly perform a variety of actions in a novel or highly variable environment is much harder. Robotic flexibility has improved over time (it’s much easier today to program a welding robot to follow a new path, for instance), but this flexibility still exists within a very narrow range of acceptable variation. The delta robot system above can grab randomly positioned objects, but would almost certainly require reprogramming if the size and shape of the objects changed, and I wouldn’t be surprised if even varying the color of the objects was enough to disrupt the existing automation.

  • On Jagged AGI: o3, Gemini 2.5, and everything after

    • My co-authors and I coined the term “Jagged Frontier” to describe the fact that AI has surprisingly uneven abilities. An AI may succeed at a task that would challenge a human expert but fail at something incredibly mundane. For example, consider this puzzle, a variation on a classic old brainteaser (a concept first explored by Colin Fraser and expanded by Riley Goodside): “A young boy who has been in a car accident is rushed to the emergency room. Upon seeing him, the surgeon says, “I can operate on this boy!” How is this possible?”

    • o3 insists the answer is “the surgeon is the boy’s mother,” which is wrong, as a careful reading of the brainteaser will show. Why does the AI come up with this incorrect answer? Because that is the answer to the classic version of the riddle, meant to expose unconscious bias: “A father and son are in a car crash, the father dies, and the son is rushed to the hospital. The surgeon says, ‘I can’t operate, that boy is my son,’ who is the surgeon?” The AI has “seen” this riddle in its training data so much that even the smart o3 model fails to generalize to the new problem, at least initially. And this is just one example of the kinds of issues and hallucinations that even advanced AIs can fall prey to, showing how jagged the frontier can be.

    • What’s clear is that we continue to be in uncharted territory. The latest models represent something qualitatively different from what came before, whether or not we call it AGI. Their agentic properties, combined with their jagged capabilities, create a genuinely novel situation with few clear analogues. It may be that history continues to be the best guide, and that figuring out how to successfully apply AI in a way that shows up in the economic statistics may be a process measured in decades. Or it might be that we are on the edge of some sort of faster take-off, where AI-driven change sweeps our world suddenly. Either way, those who learn to navigate this jagged landscape now will be best positioned for what comes next… whatever that is.

  • What is Entropy?

    • People say many things about entropy: entropy increases with time, entropy is disorder, entropy increases with energy, entropy determines the arrow of time, etc.. But I have no idea what entropy is, and from what I find, neither do most other people. This is the introduction I wish I had when first told about entropy, so hopefully you find it helpful. My goal is that by the end of this long post we will have a rigorous and intuitive understanding of those statements, and in particular, why the universe looks different when moving forward through time versus when traveling backward through time.

    • This journey begins with defining and understanding entropy. There are multiple formal definitions of entropy across disciplines—thermodynamics, statistical mechanics, information theory—but they all share a central idea: entropy quantifies uncertainty. The easiest introduction to entropy is through Information Theory, which will lead to entropy in physical systems, and then finally to the relationship between entropy and time.

  • Mitochondria Are More Than Powerhouses—They’re the Motherboard of the Cell

    • Amazingly, my co-workers and I have discovered that mitochondria are themselves social beings. At least, they foreshadow sociality. Like the bacterium they descended from, they have a life cycle: old ones die out, and new ones are born out of existing ones. Communities of these organelles live within each cell, usually clustered around the nucleus. Mitochondria communicate, both within their own cells and among other cells, reaching out to support one another in times of need and generally helping the community flourish. They produce the heat that keeps our bodies warm. They receive signals about aspects of the environment in which we live, such as air pollution levels and stress triggers, and then integrate this information and emit signals such as molecules that regulate processes within the cell and, indeed, throughout the body.

  • Delusional themes may be more varied than we thought

    • Delusions — fixed, false beliefs that cannot be changed by evidence — are a key symptom of many psychotic disorders. Diagnostic manuals like the DSM-5 group them into broad categories, with recurring themes including being persecuted or harmed, delusions of grandeur, or unusual bodily sensations. Yet little research has actually explored the limitations of these categories, whether they are universally relevant, or if they vary by culture

    • A new study in Schizophrenia Bulletin, led by University College London’s Elisavet Pappa, explores these categories in greater depth. Analysing 155 studies, altogether encompassing 173,920 participants, the team identifies 37 distinct delusional themes — far more than those previously accounted for. Their findings also reveal significant cultural variations in how people experience delusion, highlighting the need for a broader, more globally informed approach to psychosis.

  • Intelligence Evolved at Least Twice in Vertebrate Animals

    • A series of studies published in Science in February 2025 provides the best evidence yet that birds and mammals did not inherit the neural pathways that generate intelligence from a common ancestor, but rather evolved them independently. This suggests that vertebrate intelligence arose not once, but multiple times. Still, their neural complexity didn’t evolve in wildly different directions: Avian and mammalian brains display surprisingly similar circuits, the studies found.

  • Zone of Proximal Development

    • The zone of proximal development (ZPD) has been defined as: “the distance between the actual developmental level as determined by independent problem solving and the level of potential development as determined through problem-solving under adult guidance, or in collaboration with more capable peers” (Vygotsky, 1978, p. 86). For teachers, the ZPD is the space between current teaching knowledge and potential new levels with assistance. Willingness to learn enables ZPD progression. Vygotsky believed that when a student is in the zone of proximal development for a particular task, providing the appropriate assistance will give the student enough of a “boost” to achieve the task.

  • How to win an argument with a toddler

    • tldr: you can’t

  • The Mind in the Wheel – Prologue: Everybody Wants a Rock

    • lengthy article but worth reading

  • Burrito Now, Pay Later

    • Fun read

News

  • Remarks on AI from NZ

    • Last week I participated in a panel discussion on AI as part of a private event in New Zealand. The organizers asked me to kick it off by talking for ten minutes, so I pulled together a few ideas on the topic, which I’m going to present in lightly edited form here. The organizers specifically wanted to think in big-picture terms so as to avoid getting fixated on rapidly changing current events, and I think that this reflects that. The objective was to get participants thinking and talking, so I tried to present one or two notions that might be thought-provoking rather than trying to make some kind of heavy ex cathedra statement—which I’m not really qualified to do anyway.

  • $70 Million in 60 Seconds: How Insider Information Helped Someone 28x Their Money

    • On April 9, 2025, someone risked $2.5 million on SPY call options—and walked away with $70+ million in under an hour. The trade was placed at 1:01 pm. At 1:30 pm, Trump announced tariff pauses. The market exploded upward. These options that cost 85 cents were suddenly worth more than $25.

  • Novel Universal Bypass for All Major LLMs

    • Researchers at HiddenLayer have developed the first, post-instruction hierarchy, universal, and transferable prompt injection technique that successfully bypasses instruction hierarchy and safety guardrails across all major frontier AI models. This includes models from OpenAI (ChatGPT 4o, 4o-mini, 4.1, 4.5, o3-mini, and o1), Google (Gemini 1.5, 2.0, and 2.5), Microsoft (Copilot), Anthropic (Claude 3.5 and 3.7), Meta (Llama 3 and 4 families), DeepSeek (V3 and R1), Qwen (2.5 72B) and Mistral (Mixtral 8x22B).

Interesting sites

Research

  • A beautiful loop: An active inference theory of consciousness

    • Can active inference model consciousness? We offer three conditions implying that it can. The first condition is the simulation of a reality or generative world model, which determines what can be known or acted upon; namely an epistemic field. The second is inferential competition to enter the world model. Only the inferences that coherently reduce long-term uncertainty win, evincing a selection for consciousness that we call Bayesian binding. The third is epistemic depth, which is the recurrent sharing of the Bayesian beliefs throughout the system. Due to this recursive loop — in a hierarchical system (such as a brain) — the world model contains the knowledge that it exists. This is distinct from self-consciousness, because the world model knows itself non-locally and continuously evidences this knowing (i.e., field-evidencing). Formally, we propose a hyper-model for precision-control across the entire hierarchy, whose latent states (or parameters) encode and control the overall structure and weighting rules for all layers of inference. This Beautiful Loop Theory is deeply revealing about meditation, psychedelic, and altered states, minimal phenomenal experience, and provides a new vision for conscious artificial intelligence

  • Chasing the Rainbow: The Non-conscious Nature of Being

    • Despite the compelling subjective experience of executive self-control, we argue that “consciousness” contains no top-down control processes and that “consciousness” involves no executive, causal, or controlling relationship with any of the familiar psychological processes conventionally attributed to it. In our view, psychological processing and psychological products are not under the control of consciousness. In particular, we argue that all “contents of consciousness” are generated by and within non-conscious brain systems in the form of a continuous self-referential personal narrative that is not directed or influenced in any way by the “experience of consciousness.” This continuously updated personal narrative arises from selective “internal broadcasting” of outputs from non-conscious executive systems that have access to all forms of cognitive processing, sensory information, and motor control. The personal narrative provides information for storage in autobiographical memory and is underpinned by constructs of self and agency, also created in non-conscious systems. The experience of consciousness is a passive accompaniment to the non-conscious processes of internal broadcasting and the creation of the personal narrative. In this sense, personal awareness is analogous to the rainbow which accompanies physical processes in the atmosphere but exerts no influence over them. Though it is an end-product created by non-conscious executive systems, the personal narrative serves the powerful evolutionary function of enabling individuals to communicate (externally broadcast) the contents of internal broadcasting. This in turn allows recipients to generate potentially adaptive strategies, such as predicting the behavior of others and underlies the development of social and cultural structures, that promote species survival. Consequently, it is the capacity to communicate to others the contents of the personal narrative that confers an evolutionary advantage—not the experience of consciousness (personal awareness) itself.

  • Collective intelligence: A unifying concept for integrating biology across scales and substrates

    • A defining feature of biology is the use of a multiscale architecture, ranging from molecular networks to cells, tissues, organs, whole bodies, and swarms. Crucially however, biology is not only nested structurally, but also functionally: each level is able to solve problems in distinct problem spaces, such as physiological, morphological, and behavioral state space. Percolating adaptive functionality from one level of competent subunits to a higher functional level of organization requires collective dynamics: multiple components must work together to achieve specific outcomes. Here we overview 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, and implement collective intelligence by solving problems at the cell, tissue, and whole-organism levels. We explore the hypothesis that collective intelligence is not only the province of groups of animals, and that an important symmetry exists between the behavioral science of swarms and the competencies of cells and other biological systems at different scales. We then briefly outline the implications of this approach, and the possible impact of tools from the field of diverse intelligence for regenerative medicine and synthetic bioengineering.

  • The MICrONS Project

    • An unprecedented dataset of high resolution anatomical images of individual cells in mouse visual cortex, mapped on to their responses. This integrated view of function and structure lays a foundation for discovering the computational bases of cortical circuits.

  • The Platonic Representation Hypothesis

    • We argue that representations in AI models, particularly deep networks, are converging. First, we survey many examples of convergence in the literature: over time and across multiple domains, the ways by which different neural networks represent data are becoming more aligned. Next, we demonstrate convergence across data modalities: as vision models and language models get larger, they measure distance between datapoints in a more and more alike way. We hypothesize that this convergence is driving toward a shared statistical model of reality, akin to Plato’s concept of an ideal reality. We term such a representation the platonic representation and discuss several possible selective pressures toward it. Finally, we discuss the implications of these trends, their limitations, and counterexamples to our analysis.

  • Achieving Human Level Competitive Robot Table Tennis

    • Achieving human-level speed and performance on real world tasks is a north star for the robotics research community. This work takes a step towards that goal and presents the first learned robot agent that reaches amateur human-level performance in competitive table tennis. Table tennis is a physically demanding sport which requires human players to undergo years of training to achieve an advanced level of proficiency. In this paper, we contribute (1) a hierarchical and modular policy architecture consisting of (i) low level controllers with their detailed skill descriptors which model the agent’s capabilities and help to bridge the sim-to-real gap and (ii) a high level controller that chooses the low level skills, (2) techniques for enabling zero-shot sim-to-real including an iterative approach to defining the task distribution that is grounded in the real-world and defines an automatic curriculum, and (3) real time adaptation to unseen opponents. Policy performance was assessed through 29 robot vs. human matches of which the robot won 45% (13/29). All humans were unseen players and their skill level varied from beginner to tournament level. Whilst the robot lost all matches vs. the most advanced players it won 100% matches vs. beginners and 55% matches vs. intermediate players, demonstrating solidly amateur human-level performance.

  • THC: Practical and Cost-Effective Verification of Delegated Computation

    • Homomorphic cryptography is used when computations are delegated to an untrusted third-party. However, there is a discrepancy between the untrustworthiness of the third-party and the silent assumption that it will perform the expected computations on the encrypted data. This may raise serious privacy concerns, for example when homomorphic cryptography is used to outsource resource-greedy computations on personal data (e.g., from an IoT device to the cloud). In this paper we show how to cost-effectively verify that the delegated computation corresponds to the expected sequence of operations, thus drastically reducing the necessary level of trust in the third-party. Our approach is based on the well-known modular extension scheme: it is transparent for the third-party and it is not tied to a particular homomorphic cryptosystem nor depends on newly introduced (and thus less-studied) cryptographic constructions. We provide a proof-of-concept implementation, THC (for trustable homomorphic computation), which we use to perform security and performance analyses. We then demonstrate its practical usability, in the case of a toy electronic voting system.

  • Privacy Pass: A privacy-enhancing protocol and browser extension

    • Privacy Pass interacts with supporting websites to introduce an anonymous user-authentication mechanism. In particular, Privacy Pass is suitable for cases where a user is required to complete some proof-of-work (e.g. solving an internet challenge) to authenticate to a service. In short, the extension receives blindly signed ‘passes’ for each authentication and these passes can be used to bypass future challenge solutions using an anonymous redemption procedure. For example, Privacy Pass is supported by Cloudflare to enable users to redeem passes instead of having to solve CAPTCHAs to visit Cloudflare-protected websites.

    • The blind signing procedure ensures that passes that are redeemed in the future are not feasibly linkable to those that are signed. We use a privacy-preserving cryptographic protocol based on ‘Verifiable, Oblivious Pseudorandom Functions’ (VOPRFs) built from elliptic curves to enforce unlinkability. The protocol is exceptionally fast and guarantees privacy for the user. As such, Privacy Pass is safe to use for those with strict anonymity restrictions.