Superhuman

From IMAX cinema to AI hub. 🎬 Take a look inside our new Berlin home! At Superhuman (formerly Grammarly), we’re building an AI productivity platform thanks to a growing team of more than 200 people in Germany. Our new hub at Das Center at Potsdamer Platz gives us more room to build together, focus, and connect. Curious how we turned a former cinema into a workspace for our teams, with a few nods to its movie history past? Read the story behind our new Berlin hub 👇

Tool-calling performance can drop as much as 85% as LLMs get more tools to choose from. Superhuman Go’s catalog has over 100 agents and 800 individual tools, so we studied why routing breaks at scale and what fixes it. 🗒️ In our latest engineering blog post, we dive into our team's research on the causes of Go’s routing performance degradation, compare different solutions to improve performance, and take a deep dive into the most promising one, which improved routing 10–17% on production queries: embedded shortlisting. ▶️ Read here: https://lnkd.in/dVWXWGNJ

It’s no secret that AI can help security teams scale operations. But actually achieving this reality is a totally different story. There are some classic LLM limitations (like context window limitations or hallucinations) and security-specific issues (like how to make sense of logs across data sources when there isn’t any shared key). But with the right architectural decisions, guardrails, and prompting, you can build a pretty effective AI threat-hunting agent. 💡 That’s exactly what we built with Barney, our AI threat-hunter agent, and it’s quickly become a key part of our security operations. In our latest blog post, we’ll share more about how we built Barney and where we are thinking of taking it next: https://lnkd.in/dMspWdAi

The red underline that appears as you type with Grammarly looks deceptively simple. Yet it’s the result of solving a tricky infrastructure problem. Suggestions need to appear instantaneously, across 40 million daily active users, without interruption. Our latest engineering blog post breaks down how we built the model serving infrastructure to make that possible at scale, plus lessons for other teams building LLM production systems. 👉 https://b.sprh.mn/4rd4STz

We just shipped Superhuman Go for Mac and Windows without rewriting either app from scratch. How did we do it? In our latest post, we share the architecture that took us from a scrappy Docs prototype to a cross-platform AI assistant, and why the decisions made early on made all the difference. 👉 https://b.sprh.mn/46K3ka8

What does building AI products actually look like before anyone knows what’s going to work? 👋 Join us on September 16 as we share the experiments, trade-offs, and real product decisions behind AI features at Superhuman. Register today: https://luma.com/rdzvmhud