Search

The o1 model series is trained with large-scale reinforcement learning to reason using chain of thought. These advanced reasoning capabilities provide new avenues for improving the safety and robustness of our models. In particular, our models can reason about our safety policies in context when responding to potentially unsafe prompts, through deliberative alignment. This leads to state-of-the-art performance on certain benchmarks for risks such as generating illicit advice, choosing stereotyped responses, and succumbing to known jailbreaks. Training models to incorporate a chain of thought before answering has the potential to unlock substantial benefits, while also increasing potential risks that stem from heightened intelligence. Our results underscore the need for building robust alignment methods, extensively stress-testing their efficacy, and maintaining meticulous risk management protocols. This report outlines the safety work carried out for the OpenAI o1 and OpenAI o1-mini models, including safety evaluations, external red teaming, and Preparedness Framework evaluations.

Our main models of computation (the Turing Machine and the RAM) make fundamental assumptions about which primitive operations are realizable. The consensus is that these include logical operations like conjunction, disjunction and negation, as well as reading and writing to memory locations. This perspective conforms to a macro-level view of physics and indeed these operations are realizable using macro-level devices involving thousands of electrons. This point of view is however incompatible with quantum mechanics, or even elementary thermodynamics, as both imply that information is a conserved quantity of physical processes, and hence of primitive computational operations. Our aim is to re-develop foundational computational models that embraces the principle of conservation of information. We first define what conservation of information means in a computational setting. We emphasize that computations must be reversible transformations on data. One can think of data as modeled using topological spaces and programs as modeled by reversible deformations. We illustrate this idea using three notions of data. The first assumes unstructured finite data, i.e., discrete topological spaces. The corresponding notion of reversible computation is that of permutations. We then consider a structured notion of data based on the Curry-Howard correspondence; here reversible deformations, as a programming language for witnessing type isomorphisms, comes from proof terms for commutative semirings. We then "move up a level" to treat programs as data. The corresponding notion of reversible programs equivalences comes from the "higher dimensional" analog to commutative semirings: symmetric rig groupoids. The coherence laws for these are exactly the program equivalences we seek. We conclude with some generalizations inspired by homotopy type theory and survey directions for further research.

In recent work, Benjamin Schumacher and Michael~D. Westmoreland investigate a version of quantum mechanics which they call "modal quantum theory" but which we prefer to call "discrete quantum theory". This theory is obtained by instantiating the mathematical framework of Hilbert spaces with a finite field instead of the field of complex numbers. This instantiation collapses much the structure of actual quantum mechanics but retains several of its distinguishing characteristics including the notions of superposition, interference, and entanglement. Furthermore, discrete quantum theory excludes local hidden variable models, has a no-cloning theorem, and can express natural counterparts of quantum information protocols such as superdense coding and teleportation. Our first result is to distill a model of discrete quantum computing from this quantum theory. The model is expressed using a monadic metalanguage built on top of a universal reversible language for finite computations, and hence is directly implementable in a language like Haskell. In addition to superpositions and invertible linear maps, the model includes conventional programming constructs including pairs, sums, higher-order functions, and recursion. Our second result is to relate this programming model to relational programming, e.g., a pure version of Prolog over finite relations. Surprisingly discrete quantum computing is identical to conventional logic programming except for a small twist that is responsible for all the ``quantum-ness.'' The twist occurs when merging sets of answers computed by several alternatives: the answers are combined using an "exclusive" version of logical disjunction. In other words, the two branches of a choice junction exhibit an "interference" effect: an answer is produced from the junction if it occurs in one or the other branch but not both., Comment: 10 pages. Submitted to LICS 2011