TimeLike™ is a first-of-its-kind AI platform designed to help human beings find near-optimal solutions to complex and consequential real-world problems even in highly complex, rapidly changing, and unfamiliar real-world decision-making situations.

TimeLike uses computer modeling and simulation and model-based methods to greatly augment human beings’ innate ability to construct mental models of the world around them and use them to predict the probable consequences of different possible courses of action, so that they can swiftly zero in on the best course of action open to them, and start carrying it out it. TimeLike can often also be used to automate the process of carrying out the chosen course of action.

a. How is TimeLike different from LLMs and other forms of Generative AI?

Over the last few years the term “AI” has become almost exclusively associated, both in the popular press and the popular imagination, with LLMs, such as Chat GPT, and multimodal GenAI models, such as GPT-4o, Gemini, and Claude 3. All of these models share a few key characteristics: they have all been trained, at enormous expense, on vast amounts on enormous static datasets of human-generated data – text, images, sound and video – and they are all very good at generating new data sequences that are strikingly similar, in the sense that they obey the same statistics, as the datasets they were trained on. Because the outputs of these models so closely resemble human-produced text, images, sounds, and video, many people who don’t understand how the technology works tend to anthropomorphize, and assume that these models are in some sense “intelligent”, in roughly the same sense that human beings are intelligent.

In fact, as a number of prominent critics of the hype surrounding these kinds of GenAI models have pointed out, these models have no understanding of the actual real world implications behind the words, images, sounds and videos they have been trained on, or those they generate. They have no understanding of time, space, causality, the laws of physics, or even the difference between truth and falsehood. As a result, their outputs, when examined closely, very often make no sense at all. This phenomenon has come to be known as “hallucination”, but that term gives the misleading impression that this kind of behavior is the exception, rather than the rule. It would be more accurate to say that these kinds of GenAI models never do anything but “hallucinate”; it’s just that sometimes it is less obvious. TimeLike has almost nothing in common with these kinds of GenAI models. For one thing, TimeLike has a built-in “understanding” of time, space, causality, the laws of physics, and the difference between truth and falsehood. For another, TimeLike is not designed to mimic the behaviors of human beings, and certainly not to replace them; instead it is designed to greatly augment the innate capabilities of human beings, using computer modeling and simulation them to construct more accurate and reliable predictive models of the situations in which they find themselves, especially when those situations are much too complex, complicated, rapidly changing, and/or unfamiliar for the unaided human mind to be able to accurately model.

In Vernor Vinge’s famous essay on The Coming Technological Singularity, he described two very different possible paths to achieving “superintelligence”, one based on purely artificial intelligence, or AI, and the other based on using computer technologies to greatly augment human intelligence, which he called intelligence augmentation, or IA. Using Vinge’s terminology, the approach TimeLike implements, i.e. simulation-enabled cooperative human intelligence, or SECHI, would be considered a form of IA, not AI. Alternately, to use more current a term, TimeLike is a form of human-centered AI.

b. How is TimeLike different from other kinds of AI?

There are many other kinds of AI besides LLMs and multimodal GenAI, too many to discuss here, so instead we will focus just on those AI approaches that, like TimeLike, do incorporate some sort of built-in understanding of time, space, causality, the laws of physics, and describe how TimeLike differs from those. This includes, for example, Yann Lecun’s Joint Embedded Predictive Architecture (JEPA), Nvidia’s Physical AI, Joshua Tenenbaum’s Computational Cognitive Science, Pascal/ISI’s Simulation Intelligence, and Eric Xing’s Physical Agentic, and Networked (PAN) world models. We consider all of these kinds of approaches to be promising, interesting, and potentially highly synergistic with TimeLike’s simulation-enabled cooperative human intelligence (SECHI) approach, but, to the best of our knowledge, none of

these other approaches supports all of the key features and functionality provided by TimeLike. First, to the best of our knowledge, unlike TimeLike, none of these other approaches has been expressly designed to support maximal generality in modeling, so that it could be used to model any possible decision-making situation consistent with the laws of physics at any level of fidelity, or mix of fidelities, that might be necessary and appropriate to provide reliable guidance in making whatever decisions our users might need to make – and, as it turns out, surprisingly many real-world decision-making situations do in fact require this kind of extreme generality. Second, to the best of our knowledge, unlike TimeLike, none of these other approaches has been expressly designed to support cooperative decision-making at scale, enabling large numbers of users (up to billions) to coordinate their decision-making and actions intelligently, effectively, and efficiently, so that they readily find and take advantage of omni-win-win opportunities while deftly avoiding multipolar lose-lose traps.

Yes. Unlike most other forms of AI, TimeLike does not provide, nor rely upon, it’s own algorithms for machine learning, and does not provided any native mechanisms designed to extract useful information from large datasets made of unstructured data. Instead, TimeLike relies upon its users – some of them – to provide high quality, verified and validated composable causal computer models of all of of the real or hypothetical real-world systems and effects that are pertinent to any of the kinds of real-world decision-making situations they want or need to be able to use TimeLike to support. For example, if we want to be able to use TimeLike to support all the decision-making pertinent to running a smart factory, or a smart supply chain, or a smart city, we are generally going to need the help of human experts familiar with those kinds of systems, and expert in constructing and using computer models of all the pertinent subsystems effects. As we’ve indicated, TimeLike is not designed to replace human expertise and intelligence, but rather to augment it, making it possible for human beings to make much more intelligent and effective use of the many highly complex and complicated systems we have to deal with in today’s ever more complex, complicated, and interconnected world.

At first blush, this requirement – the need for help from humans expert in the application domains of interest and familiar with computer modeling and simulation and model-based methods – might seem like it would be a significant limiting factor on the usefulness of the technology. And, in a sense that’s true – if we are ever to be able to use TimeLike for all the many different kinds of decision-making it could potentially be useful for, we are going to need a great deal of help from scientists and engineers expert in many different domains of science and engineering. However, this is not nearly as big a problem as it might seem. In the first place, we aren’t going to need all of those models, and all of that specialized expertise all at once; we can and should pick and choose our initial use cases carefully, looking for “low-hanging fruit”, while at the same time developing and following a roadmap designed to get us to where we ultimately want to be. Also, as a practical matter, most scientists and engineers today need to have at least a reasonable degree of familiarity with the tools and methodologies of computer modeling and simulation and model-based methods, because these have become essential tools for people working in these areas. And, for the same reason, many of the specialized computer models we are going to need already exist in some form that is at least reasonably close to what we are going to need for use in TimeLike. As a result, it seems likely the largest part of the computer modeling that work will be needed will be in the area of model integration, rather than having to develop lots of new models from scratch. And TimeLike has been expressly designed to provide a near ideal platform for general purpose model integration.

TimeLike is the latest and greatest of a series of roughly half a dozen increasingly powerful, general, and flexible software platforms for modeling and simulation and model-based decision-making that we have designed, developed, and tested and refined through use in demanding real-world applications over the last several decades, mostly in connection with our past work for the US Department of Defense. Most but not all of TimeLike’s key features and functionality have been previously implemented, tested and refined through use in the context of one or more of these earlier software platforms. However, some of the desired features and functionality have never before been implemented, and we anticipate that some of these will likely require some iterative refinement, guided by the feedback we receive from early adopters, before they will be ready for general use. Also, all of those earlier platforms were designed for use within the relatively low-stakes context of model-based engineering – where a bug or a software crash won’t kill anyone – and by only a limited number of users, and within a controlled and secure environment, e.g. a private compute cluster, or one of the supercomputing centers operated by the US DoD. Ultimately, we want TimeLike to be suitable for use in highly consequential applications, by any number of users, and within an open computing environment, and that raises a whole host of new issues related to security, reliability, robustness, DevOps, etc. However, we don’t need to wait until we have satisfactory solutions for all of those requirements before we can start to use TimeLike in the design and development of those kinds of systems; that generally shouldn’t be necessary not until we are nearly ready to deploy those systems in an operational environment.