Martian is solving the largest problem in AI: making AI safe by letting us look inside the brains of models like GPT-4. This also lets us make models more effective and less expensive. All of modern AI exists because of the Transformer architecture. Nobody knew it at the time, but the invention of the Transformer solved a fundamental problem: previously, we couldn't put enough data through our models to get them to understand language. The result was GPT-3, ChatGPT, GPT-4, and a host of other models; models which actually work and understand natural language. Now that we have working AI, a different set of problems arises. How do we know what our models are doing, and how do we know that they will actually do what we want? From this perspective, the Transformer is a nightmare. It's a black box where you can see the inputs and outputs, but have no understanding of what's going on inside. We're building a safer alternative to the Transformer: one which is equally effective, but completely transparent and intelligible. This has a huge number of advantages: it can allow us to make models faster, cheaper, and more accurate. But most importantly of all, it allows us to create guarantees on AI behavior. We can start to say things like "this AI will never break the law" or, as AI becomes really powerful, "this AI will not end humanity." But we think that the implications are even bigger than that. When Google published the Transformer in 2017, nobody knew that it would lead to models like ChatGPT or potentially AGI. The Transformer ushered in a new era of AI. We think that the next era starts by understanding what's happening inside these models. The next era starts with Martian.
Something looks off?On-site & Remote