Founders' Letter

We stand on the cusp of truly exciting times—where superintelligence is only a few years away, promising centuries of progress in mere months. Already, we have models with PhD-level intelligence operating at speeds far beyond human capabilities. Yet, paradoxically, most complex tasks are still carried out by humans. The bottleneck isn’t intelligence: it’s communication. Current interactions with AI rely almost entirely on prompting: an inherently lossy approach that forces us to treat advanced systems like black boxes. Imagine trying to optimize a software system without access to its source code—it’s not only inefficient but absurdly limiting.

We believe there’s a better way: applied interpretability. By truly understanding a model’s inner mechanisms, we gain the ability to directly control its behavior and how it uses its latent knowledge. This isn’t just about safety; it’s about fundamentally enhancing model performance.

At Tilde, we’re building full-stack interpretability solutions to make this vision a reality. This includes general interpreter models at each stage of the inference pipeline, and control tech to optimally orchestrate the elicited features in deployment. These solutions enable human experts to ensure that their complex knowledge is effectively utilized by the models they deploy, allowing AI to then go on and tackle the tasks that are truly beyond human reach. Crucially, this new paradigm serves to elevate, not replace, the traditional arts of fine-tuning and other post-training methods, by providing a new layer of insight and control.

Lastly, a few words on our philosophy and approach. While there’s a scaling element to what we do—building larger interpreter models for larger models—to say that building the “right” interpretability solution is a matter of scale alone would be a mistake. After all, creating infinitely-long interpreter models is useless. A guiding thesis we have at Tilde is that the path towards genuine model understanding lies in finding the right mathematical lens to express the problem.

Our team brings together a unique set of backgrounds, with extensive experience in academia as well as small and large AI infrastructure companies. We started Tilde with the fundamental goal of achieving a better understanding of the universe. It just so happens our models are universe compressors—distilling humanity’s entire knowledge into a couple trillion parameters—and to understand what fills our night sky we just might have to look at the constellations hidden inside the black box.

We’re only at the beginning, and while we’re excited by the leaps in reasoning we’ve already seen, there’s still so much more to discover. If this sounds interesting to you, come build the future with us.

Founders, Tilde Research

Ben, Dhruv, & Mason

join@tilderesearch.com