Rapidata

Meet the 10 startups participating in the 2 week Startup Camp (All Industries) powered by Innosuisse and managed by Swissnex in San Francisco. From September 7 to 17, 2026, the startups are in Silicon Valley for 2 weeks of sessions, fireside chats, networking events, a pitch night, and more. Jannis Schönleber, CEO, and Nils Eyer, Head of Sales at 44ai AG Andrej Babic, CEO of Adiposs SA Mateo de Bardeci, CEO of DeepPsy AG Yegor Piskarev, CEO of Dexterous Endoscopes (hiring!) Lyle Halliday, CTO, and Nour Ghalia Abassi, CEO of DigeHealth Matthias Spühler, CEO, and Regula Spuehler, COO of heyPatient AG Andrea Moroni Stampa, Chairman, and Nathan Deutsch, COO of Lighthouse Tech 🕶 SA Milan Kuzmanovic, PhD, Chief AI Officer of nextesy AG Fayçal M'hamdi, Head of Go To Market, and Jason Corkill, CEO of Rapidata AG Federico Martinelli-Orlando, CEO of TALPA-Inspection AG Connect with them to learn more. Stay tuned for more information about the Startup Pitch Night that will take place on September 10, 2026 Managed by the Startup Program Team at Swissnex in San Francisco: Benjamin Leutwyler, Luna Dutli, Sven J.

If I tell you that human feedback could arrive fast enough, without sacrificing quality, to become "The reward signal" for model training? You'd probably wonder about the quality of that feedback. So let's dive in 😉 AI labs currently have two main options: approximate human judgements with a reward model, or collect human feedback asynchronously through traditional crowdsourcing platforms. Let's compare Rapidata through typical crowdsourcing. We sent the same 300 visual tasks to Rapidata and Prolific and published all 18,000+ individual responses. The result is the following: Prolific participants were more accurate individually: 98.0% vs. 91.6%. But once responses were aggregated, both crowds reached essentially perfect final-label accuracy. On this experiment, Rapidata delivered those labels: → 70× faster (15K answers in 2 minutes vs 3K in 66 minutes) → 8× cheaper → Both with >99.9% accuracy with enough answers per item At this speed, human feedback does not have to remain a slow and asynchronous annotation step, nor has to be approximated. It can be added as a direct reward signal during post-training or, for some training loops, replace the reward model entirely. We first explored this application while supporting the post-training of one of the leading image models out there. Full methodology, limitations, results, and raw dataset in comments.

2000 builders, working simultaneously from 13 different cities, on challenges from OpenAI, ElevenLabs, Databricks, Vercel and others. Last weekend (actually 10 days ago), we hosted the Zurich hub of Hack-Nation at Rapidata. The nice thing about this Hackathon is that it happened at the same time across 14 cities, from San Francisco to Delhi. In Zurich, we had to start at 6 p.m. so the challenges wouldn’t be leaked before the SF hub started at 9 a.m. Around 2,000 builders were therefore working simultaneously. We kicked off with a call between all the hubs, which gave it a bit of a sci-fi movie feeling: several cities around the world working on the same mission. In Zurich, we had around 40 builders who spent 24 hours working non-stop on sponsor challenges. The projects covered everything from AI applications in women’s health to an AI-powered audio negotiator etc... A big thank you to the Hack-Nation, Doruk Tan Öztürk and 0sec team, the mentors and judges, and the global sponsors and supporters. And, of course, thanks to everyone who joined us in Zurich and brought so much energy to the office. Ralf Boltshauser, Lucie Bierent, Caroline Knop, Vanda Alexeeva, Marco Pagano, Piotr Kleymenov (and many others). It was a lot of fun to host. We’re looking forward to seeing where some of these projects go next 🚀

Generative models keep improving. Output diversity doesn't. Reward models are too general and ain't keeping up. #CVPR2026 Between our Rapidata booth, our meetup on post-training and eval, we had 400+ conversations with researchers working on image, video, and world models. A few things kept coming up: • Almost everyone is using RL in post-training, but most approaches work with the same small set of reward models (PickScore, HPSv3, ...). These reward models are more general and mainly cover aesthetics and prompt alignment. They are useful but might not be the right objectives for many tasks. • Model quality keeps improving, but mode collapse is on everyone's lips. Outputs are getting excellent, yet increasingly converge toward the same styles (the cat might look great, but it's too often the same cat). • Human feedback is becoming a more central part of the training loop, with growing interest in online and continuously updated feedback rather than static datasets alone. At Rapidata, our goal is to support this shift by making human feedback available at an adequate speed for training cycles, with 6K+ human annotations per minute. 🚨 Another thing also stood out: the field needs more benchmarks that measure areas that lab customers care about and spend the $$$ for. We will be publishing more of these over the coming months. If there's a benchmark you'd like to see us publish (SVG generation, product position editing...), let us know in the comments. Thanks to everyone who stopped by our booth, joined the meetup, and shared their work. Looking forward to continuing the conversations. Bonus Pic: the robot figured I had a higher reward score