OORT details 2026 RLHF roadmap questions

OORT’s 2026 roadmap put two pieces of its business on the same page: a larger edge-computing network and a bigger push into reinforcement learning from human feedback, or RLHF. In a January 23 roadmap post, the company said it wants more than 100,000 active Deimos II nodes and described RLHF and predictive data collection as potential revenue drivers for DataHub. OORT’s commercial RLHF page makes a similar pitch from the services side. The company says it provides “high-precision human feedback” through a decentralized workforce, with contributors in more than 100 countries, multi-layer identity verification and “100% bot-free” operations. Those claims help explain the questions that surfaced around the roadmap this week. The discussion was less about whether RLHF is useful than about how a decentralized RLHF pipeline would work in practice when the system has to filter bad inputs, move feedback quickly enough for training loops, and decide how contributors get paid. (oortfoundation.org) ### What exactly did OORT put in its 2026 roadmap? The January 23 roadmap said OORT plans to shift Deimos II from storage toward decentralized AI data processing. (oortech.com) The company said it already had more than 58,000 active nodes and set a 2026 goal of 100,000-plus nodes capable of running small language models and other data-processing algorithms locally. The same roadmap tied that network to DataHub, OORT’s data marketplace and task system. OORT said it wants a “comprehensive gamified incentive system,” an open task launchpad and a developer marketplace for data collection and processing workflows. It also said “RLHF and predictive data collection have the potential to bring substantial revenue to DataHub.” (oortfoundation.org) ### Why did RLHF become the pressure point? OORT’s RLHF product page says tasks can be “launched and completed in hours rather than weeks,” and says verified contributors and behavior monitoring improve trust in the feedback pipeline. Those are the operating claims users immediately tested against the roadmap. Community questions centered on familiar RLHF failure modes: whether a large open contributor base can be protected from noisy or malicious data injection, whether real-time or near-real-time human feedback is possible at useful latency, and whether compensation should be paid in a stable asset or in OORT’s token. (oortfoundation.org) Those questions track the mechanics of the model OORT is proposing rather than a dispute over the company’s stated direction. (oortech.com) ### Why were reward functions part of the conversation? A May 21 post highlighted RULER, a natural-language reward framework that uses plain-English criteria to score agent trajectories. The write-up said the approach addresses brittle reward design by replacing hand-coded metrics with language-based evaluation. That discussion intersected with a broader critique associated with Andrej Karpathy: scalar or hand-built reward functions can be unreliable for complex agent behavior. The RULER summary said natural-language reward systems can make reward design easier to update as tasks change, though it also described implementation challenges and the need for careful evaluation. (blockchain.news) ### Where does that leave OORT’s roadmap? OORT’s own materials already frame quality control as a selling point. The RLHF page cites identity checks, behavior monitoring, bot detection and a trained subset of contributors for red-teaming, while the roadmap frames the wider system as a “verifiable” and privacy-preserving data engine. The next concrete reference points are the company’s published pages rather than a new filing or product launch. (blockchain.news) OORT’s foundation roadmap dated January 23, 2026 and its live RLHF services page are the clearest statements of record on node targets, incentive design and how the company says it will run decentralized human-feedback work. (oortfoundation.org) (oortech.com)