Unconventional AI was founded on a suspicion: that the dominant recipe for building capable AI systems, more parameters and more attention over ever longer contexts, is one path among several, and that the alternatives stay underexplored not because they are less promising but because they are harder to fund.

The Unconventional Grant exists to fund those alternatives. We asked for work that questions an assumption the field has stopped examining, and we said plainly that we would rather back a well-posed risk than a safe increment.

We did not know what to expect from a pilot. Nearly 100 submissions came in, from researchers who clearly did not need convincing that the unexamined assumptions are where the interesting problems are. That response tells us something we suspected but could not prove: there is a great deal of serious unconventional work out there waiting on someone to fund it.

Today we are announcing our first cohort. Five projects, at Stanford, the University of Texas at Austin, Yale, MIT, and UC Berkeley, each receiving $100,000. Below, each team describes its own project. These are their public abstracts, printed in their words.

The 2026 cohort

Retention-Aware Differentiated-Memory Scalable AI Processors

Thierry Tambe, Stanford University

Today’s AI processors retain every bit of on-chip memory for far longer than most AI values actually remain useful. Our project explores a different approach: treating data retention as an architected resource, much like precision, bandwidth, or locality. We are building a retention-aware AI computing system that combines conventional SRAM with dense short-term memory (StRAM), such as gain-cell embedded DRAM, and long-term memory (LtRAM). A new compiler will determine how long different pieces of data need to remain valid and then place and schedule them in the most appropriate memory. Short-lived activations, intermediate results, speculative tokens, and temporary agent state could reside in dense, energy-efficient short-term memory, while frequently reused or persistent data would remain in SRAM or LtRAM. The central idea is that processors should not spend area and energy preserving information long after the computation no longer needs it.

This approach directly attacks the growing memory wall in AI systems. SRAM has become increasingly expensive in chip area, while emerging logic-compatible memories can provide greater density and lower leakage but only retain data for a limited time. Rather than hiding that limited retention through continual refresh, we want the hardware and compiler to exploit it safely. Our goal is to leverage this cross-layer to substantially increase effective on-chip memory capacity and bandwidth, and reduce memory-system energy without sacrificing correctness or model quality. We will evaluate the concept across generative AI, long-context models, vision-language workloads, and agentic systems, and ultimately validate it in silicon with a retention-aware compiler. Along the way, we expect to learn which AI data structures have predictable enough lifetimes to safely use short-retention memories, how much compiler and hardware support is required to manage them, and the quantifiable benefits after accounting for migration, variability, and other real-system overheads. More broadly, we hope to demonstrate that bounded retention can become a useful design resource rather than a limitation, opening a new path toward denser memory and more energy-efficient AI processors.

Fractal Basins Trap Latent Reasoning

William Gilpin, University of Texas at Austin

Reasoning allows artificial intelligence models to revise and correct their mistakes, enabling recent frontier advances in software engineering, mathematical theorem solving, and autonomous task planning. Modern recurrent reasoning models dynamically allocate additional computation to difficult problems, yet the physical mechanism governing this slowdown remains unknown. We hypothesize that reasoning slowdowns are a consequence of transient chaos, a general phenomenon in high-dimensional dynamical systems like turbulent fluid flows and biological signaling networks. To test this theory, we will identify and quantify the emergence of fractals, a physical signature of transient chaos, across a diverse range of modern reasoning architectures and tasks.

We expect that as reasoning models encounter increasingly difficult tasks, the complexity and depth of fractals in their latent dynamics will increase. Using dynamical systems theory, we will identify how the false starts and dead ends that characterize difficult problems become trapping regions in latent space that delay convergence. We will develop mechanistic interpretability tools based on saddle-point theory in order to identify weakly unstable, nearly correct solutions associated with reasoning traps. Building on tools from optimization theory and statistical physics, we will identify optimal protocols for mitigating the influence of reasoning traps and thus escaping from fractal basins.

We aim to establish a framework for using nonlinear dynamics to distinguish intrinsic computational complexity from avoidable architectural or training limitations in reasoning models. It will yield diagnostics for model-task alignment, interpretable accounts of reasoning failures, and test-time interventions to reduce inference cost without retraining or reducing accuracy. More broadly, our work aims to connect modern large reasoning models to universal principles governing computation, relaxation, and control in complex nonlinear systems.

Noise as a Resource (Or at Least Not a Hindrance): Physical Neural Scaling Laws for Co-Designed AI Hardware

Logan Wright, Yale University

Current AI systems are essentially digital computer simulations of complex analog machines. While these complex machines remain partially mysterious, the field and industry of AI have developed a vast repertoire of empirical engineering wisdom, but also, and increasingly, a theoretical understanding of these simulated machines. This understanding, the physics of AI, includes concepts like neural scaling laws, which have become arguably the most important principled engineering strategy behind modern AI. This physics also includes interpretable approximate models of AI systems, which represent one of the main means by which scientists are decoding the black box of AI, and are one of the most promising means by which AI engineers can in turn develop strategies to optimize and control its performance, including ensuring safe operation.

A key goal of my laboratory at Yale is to consider what the corresponding theory and engineering wisdom should look like for physical, analog, realizations of these complex analog machines, physical neural networks. Some of this work is translational: naturally much of the established wisdom and theory that applies to digitally simulated analog neural networks partially applies to real physical analog neural networks. However, it is not all translational because real physical neural networks are not exactly the same as digital simulations: they include noise, they drift due to the environment around them, they do not always mimic the exact mathematical form of idealized neural networks, and they may have physical constraints, such as a limited amount of space they can occupy or energy they can consume. But, on the other hand, they can be described using existing physics because they are actual, rather than digitally simulated, analog physical machines. With a gift and support from Unconventional AI, my laboratory intends to apply our developing physics of physical neural networks to electronic physical neural networks, with the aim of deriving engineering wisdom for this new paradigm of physical AI hardware, identifying design principles for efficient scaling, and using these to propose codesigned physical architectures and algorithms that respect the physical constraints and capabilities of realizable electronic hardware and its surrounding infrastructure.

In addition, we plan to contribute infrastructure to enrich the community of physical AI science, such as a “flipped” seminar series intended to provide an accessible introduction to scientists, engineers, and students interested in this topic.

Deeply Looped Recurrent Transformers

Yoon Kim, MIT

Transformers are the currently dominant architecture for large language models and AI more broadly. Key to their effectiveness is their scalability, i.e., continued increases in performance as the model size, in particular model depth, increases. However, in ordinary Transformer couple computational depth with model size: deeper Transformers requiring more parameters, which places demands on memory amount and bandwidth during training and inference. Looped Transformers (also sometimes called recurrent-depth or recursive Transformers) decouple depth with model size by reusing portions of the network. Recent works have found that looped Transformers can increase model performance without using more parameters, and can moreover enable improved reasoning capabilities on certain domains. However, existing works have generally explored looped Transformers in the context of tens of loops, observing that gains via looping saturate quickly. This project seeks to explore looped Transformers in regimes where the number of loops is in the hundreds or thousands. To make training and inference tractable, we will also explore loopable primitives where the majority of the parameters can live “on-chip”, thereby substantially reducing memory bandwidth.

Brain-Inspired Recurrent Architectures for Efficient State Tracking

Wanyu Lei, Daniel Kunin, and Friedrich T. Sommer, UC Berkeley

Modern AI systems are powerful pattern recognizers, but inefficient state trackers. State tracking is the ability to maintain an internal state by integrating past actions and observations relevant to the future. This ability is essential for intelligence, including reasoning, planning, memory, control, and navigation. Biological intelligence evolved recurrent architectures to solve state tracking tasks efficiently. In contrast, modern attention-based foundation models retain the full history of the past irrespective of its future relevance. This strategy for state tracking incurs substantial computational and memory costs that grow quadratically with sequence length. Our premise is that efficient state tracking is not merely a capability of biological intelligence, but a centerpiece for a different kind of artificial intelligence.

Our project aims to establish a general theory of efficient state tracking and use it to develop recurrent architectures whose internal dynamics are derived directly from the mathematical structure of state updates. In particular, we will leverage representation theory and harmonic analysis to develop the theoretical framework, then use the resulting principles to design architectures with nontraditional artificial neurons and fixed recurrent matrices that efficiently represent and update information over time. We will study state tracking across a hierarchy of increasingly complex computational problems, from sequential group composition, to finite-state automata with irreversible state transitions, to context-free languages with richer forms of memory. Our preliminary results show that these mathematically designed recurrent architectures can learn from substantially fewer examples and generalize to sequences longer than those seen during training. We will investigate whether these principles continue to hold as state-tracking problems become more expressive and complex. We expect such architectures to offer advantages over attention-based transformers in energy efficiency, sample efficiency, interpretability, and generalization beyond the training horizon. More broadly, we hope to learn whether state tracking can serve not simply as one capability that AI systems must acquire, but as a foundational design principle for the next generation of artificial intelligence.

What happens next

From here the work is theirs. We are not steering the research or holding anything back for ourselves; the one thing we asked for is that it be published in the open. Awardees will do that through their usual channels, and we will point to it as it lands.