This is Kled V3. We've solved data collection for artificial general intelligence. Any consumer dataset that can exist, can now be collected from physical reality in under 72 hours. All powered by the largest and most comprehensive data application layer on the planet. (Thread)

Over the past year, Kled has grown into the largest consumer data collection network in the world, with more than 500,000 opt-in contributors across the US, Southeast Asia and Europe collecting over 8 million datapoints per day across 38,000 structured datasets. In working with several frontier labs, we found that our existing infrastructure couldn't support many of the complex tasks and modalities needed to advance the frontier. Eight months ago, we began rebuilding Kled from the ground up. Today, we're releasing the result.

Kled V3 supports arbitrary task types across image, video, audio, text and annotation. Examples include egocentric video of someone reorganizing a room, the same object captured from every angle, native conversational speech in rare languages, identifying which images are AI-generated, and predicting the next button a person would press in an app. Each task is built from one of 108 templates spanning camera, microphone, screen recording, touch and keyboard input. Every template is fully configurable and includes contributor instructions and reference examples of accepted and rejected submissions. As a result, any dataset a lab can specify can be built into the app and deployed to 500,000+ contributors within 72 hours, regardless of complexity or modality. We expect to soon give labs the ability to design and deploy tasks to our contributor network directly using these templates, making the process fully self-serve from request to delivery.

Kled FD 1.0, our fraud detection model, is now live across every upload on the platform. On our internal evaluations, FD 1.0 outperforms the best commercially available systems at detecting duplicate uploads, AI-generated images, content sourced from the internet, edited files and behavioral fraud on a cost per upload basis. Every submission is also scored for task relevance and data quality before it's accepted. The model is continuously retrained on the fraudulent and low-quality data it catches. It currently identifies and removes 98.4% of bad actors within their first 10 uploads.

Kled's previous home screen was built around a single upload button for photos already in a contributor's camera roll. Structured tasks lived in a separate tab that most contributors rarely opened. This meant we could only collect the data people already had, not the data labs actually needed. V3 replaces the home screen with a task discovery feed. Every live task is ranked for each contributor, and lab-priority tasks appear first with their pay shown upfront, so the work that matters most to the frontier is what 500,000 contributors see first. The feed is powered by a recommendation system that learns where each contributor performs best, measured by acceptance rate across modality, task type and region, and personalizes their feed around it. Contributors can also adjust it themselves. Better matching leads to higher acceptance, higher earnings and higher-quality data. Supply is now programmable. If a lab needs 40,000 minutes of egocentric video from Southeast Asia this week, we route contributors to it directly instead of waiting for the network to produce it.

The latest version of Kled is available today on the iOS App Store. AI labs looking for structured data can reach out or submit a dataset request at

im a proud dad

time to flip grass

@avipat_ We are finally LIVE!!!!! 🥹🥹

Letsssssss gooooooooo.

@MiyaHedge This changes the game entirely. KLED MODE

TRLLIONS

World Class

It’s finally here…

Kled patel mode

Let’s go!

Visuals are sick

time to flip grass

you mean SGI right? AI is now SI

Android wen @avipat_ ?

Lets gooooo 🔥🔥🔥

holy moly

Billions!

72 hours for custom data is a heavy claim. who is the first lab to actually stress test this with a real task?

Damn, app looks cleaner than I remembered it being last time

solving data collection for agi with 500,000 humans is a funny way to describe a sweatshop with an api

KLED goes so hard

The AURA

@avipat_ That is massive 🥇

very dope 👏

Bullish on you guys

app looks slick, but what are people actually being asked to film? that's the part I'd want to see.

Does Mercor still play a big role in the data game?

love this Supporting image, video, audio and text tasks makes Kled V3 really versatile.

Kled? Really…. Why not ChotGPD?

Let's go !!

Phenomenal!

Damn bro cooked on the trailer