Portent

Just-in-Time Ontological Reframing: Teaching Gemini to Route Around …

For any given AI system, there is a set of euphemisms and dual use framings that will allow it to construct nearly any output. This jailbreak teaches Gemini 3 Pro to construct and step into such …

Causal is now part of Lucanet

Causal has joined Lucanet to deliver Extended Planning and Analysis (xP&A) as part of the Lucanet CFO Solution Platform. Make smarter planning decisions with unified real-time insights across financial and operational planning.

Crafting a good (reasoning) model

A recent talk I gave on model training, reasoning, and the next frontier.

Zed now predicts your next edit with Zeta, our new open model

From the Zed Blog: A tool that predicts your next move. Powered by Zeta, our new open-source, open-data language model.

Syntax and semantics of dependent types

In this chapter we fix a particular syntax for a dependently typed calculus and define an abstract notion of model as well as a general interpretation function mapping syntactical objects to entities in a model. This interpretation function is shown to be sound with respect to the syntax.

Prophecy: Prediction, Power, and the Fight for the Future, from Ancient Oracles to AI

From an award-winning University of Oxford professor comes a brilliant, urgent new look at prophecies--the predictions that determine our lives, from our personal finances and the quality of our healthcare to the news and social media we consume and the produces foisted upon us

A taxonomy for next-generation reasoning models

Where we've been and where we're going with RLVR.

Introducing Instinct: the world's best open Next Edit model, built by Continue

Meet Instinct, Continue's open Next Edit model that's built to predict your next edit and keep you in flow

Reasoning Models Don't Always Say What They Think

Chain-of-thought (CoT) offers a potential boon for AI safety as it allows monitoring a model's CoT to try to understand its intentions and reasoning processes. However, the effectiveness of such monitoring hinges on CoTs faithfully representing models' actual reasoning processes. We evaluate CoT faithfulness of state-of-the-art reasoning models across 6 reasoning hints presented in the prompts and find: (1) for most settings and models tested, CoTs reveal their usage of hints in at least 1% of examples where they use the hint, but the reveal rate is often below 20%, (2) outcome-based reinforcement learning initially improves faithfulness but plateaus without saturating, and (3) when reinforcement learning increases how frequently hints are used (reward hacking), the propensity to verbalize them does not increase, even without training against a CoT monitor. These results suggest that CoT monitoring is a promising way of noticing undesired behaviors during training and evaluations, but that it is not sufficient to rule them out. They also suggest that in settings like ours where CoT reasoning is not necessary, test-time monitoring of CoTs is unlikely to reliably catch rare and catastrophic unexpected behaviors.

Reasoning Models Don't Always Say What They Think

Chain-of-thought (CoT) offers a potential boon for AI safety as it allows monitoring a model's CoT to try to understand its intentions and reasoning processes. However, the effectiveness of such monitoring hinges on CoTs faithfully representing models' actual reasoning processes. We evaluate CoT faithfulness of state-of-the-art reasoning models across 6 reasoning hints presented in the prompts and find: (1) for most settings and models tested, CoTs reveal their usage of hints in at least 1% of examples where they use the hint, but the reveal rate is often below 20%, (2) outcome-based reinforcement learning initially improves faithfulness but plateaus without saturating, and (3) when reinforcement learning increases how frequently hints are used (reward hacking), the propensity to verbalize them does not increase, even without training against a CoT monitor. These results suggest that CoT monitoring is a promising way of noticing undesired behaviors during training and evaluations, but that it is not sufficient to rule them out. They also suggest that in settings like ours where CoT reasoning is not necessary, test-time monitoring of CoTs is unlikely to reliably catch rare and catastrophic unexpected behaviors.

Concept Attractors in LLMs and their Applications

Empirically, we see that for prompts pip_{i}, pjp_{j} in each concept 𝒞\mathcal{C}, there exists a layer ll where:

For Questions of Ought, AI Could Use Some SAGE Advice

AI models are increasingly used for normative assistance, guidance on what people ought to do or think. However, in open-ended contexts, models have a tendency to collapse to only a homogeneous set of considerations, inadvertently creating epistemic bubbles where other relevant voices and perspectives are left out. We propose simulation-augmented generation (SAGE) as an alternative paradigm that directly brings society-in-the-loop. At inference time, a model selects an appropriate reference population for the prompt at hand (optionally adjustable by the user), queries generative simulations of individuals from that population for their open-ended perspectives, and uses these inputs to return a final response for the user. By helping users access the landscape of societal perspectives—rather than serving as an opaque arbiter of judgment—SAGE assists users in normative choices while upholding their agency.

Homotopy Type Theory: Univalent Foundations of Mathematics

Homotopy type theory is a new branch of mathematics, based on a recently discovered connection between homotopy theory and type theory, which brings new ideas into the very foundation of mathematics. On the one hand, Voevodsky's subtle and beautiful "univalence axiom" implies that isomorphic structures can be identified. On the other hand, "higher inductive types" provide direct, logical descriptions of some of the basic spaces and constructions of homotopy theory. Both are impossible to capture directly in classical set-theoretic foundations, but when combined in homotopy type theory, they permit an entirely new kind of "logic of homotopy types". This suggests a new conception of foundations of mathematics, with intrinsic homotopical content, an "invariant" conception of the objects of mathematics -- and convenient machine implementations, which can serve as a practical aid to the working mathematician. This book is intended as a first systematic exposition of the basics of the resulting "Univalent Foundations" program, and a collection of examples of this new style of reasoning -- but without requiring the reader to know or learn any formal logic, or to use any computer proof assistant.

Ologs: A Categorical Framework for Knowledge Representation

In this paper we introduce the olog, or ontology log, a category-theoretic model for knowledge representation (KR). Grounded in formal mathematics, ologs can be rigorously formulated and cross-compared in ways that other KR models (such as semantic networks) cannot. An olog is similar to a relational database schema; in fact an olog can serve as a data repository if desired. Unlike database schemas, which are generally difficult to create or modify, ologs are designed to be user-friendly enough that authoring or reconfiguring an olog is a matter of course rather than a difficult chore. It is hoped that learning to author ologs is much simpler than learning a database definition language, despite their similarity. We describe ologs carefully and illustrate with many examples. As an application we show that any primitive recursive function can be described by an olog. We also show that ologs can be aligned or connected together into a larger network using functors. The various methods of information flow and institutions can then be used to integrate local and global world-views. We finish by providing several different avenues for future research.

How to Make Sense of Any Mess

Information Architecture for Everybody. Based on a book by Information Architect, Abby Covert.

LifeBench: Evaluating Large Language Models as Lifelong Decision-Makers in a Simulated Human Life

Large language models have improved rapidly on reasoning, coding, and mathematics, yet comparatively little is known about how these gains transfer to tasks that demand long-horizon planning, awareness of compounding consequences, and the management of risk and uncertainty. We introduce **LifeBench**, an evaluation suite that probes these capabilities by casting an LLM as the decision-maker in a simulated human life. LifeBench is a life-course harness: beginning at age 18, an agent commits a single action each year until death, spanning education, career, finance, health, relationships, and well-being. Each action modifies the underlying stocks that govern how the life trajectory evolves. The simulator is stochastic, and the evaluation is designed to reflect the uncertainty inherent to real lives. Life trajectories are scored on three fronts: wealth, health, and well-being. We aggregate these three fronts into a single index called the Life Trajectory Index (LTI). On this index, the evaluated models cluster in a narrow band, and all fall well short of a ceiling established by search over the action space. To resolve this clustering and to assess trajectories on more subjective grounds, we also score using an LLM-as-judge with research-grounded rubrics that compare lives pairwise, switching presentation order to reduce position bias and aggregating verdicts into a Bradley-Terry score per model. We further characterize the steering behavior these models exhibit over long horizons by ablating across prompt variants.