a16z: Why we invested in Yupp

Author: Chris Dixon, founder of a16z crypto; Elizabeth Harkavy, partner of a16z crypto; Translation: Jinse Finance xiaozou

Modern AI systems rely not only on computing power and algorithms, but also on human feedback. Companies use post-training optimization techniques such as human feedback reinforcement learning (RLHF) and direct preference optimization (DPO) to improve models. These techniques can reduce bias and enable models to respond to prompts with higher quality and more coherent responses - which is critical to accelerating AI development. Model evaluation is also critical, but model optimization can only be achieved by defining what is "better" first.

The challenge arises from this: companies are reluctant to share - they treat data and training processes as confidential. This results in AI model evaluation relying on limited information from closed systems or static benchmarks that are divorced from actual applications, which seriously restricts the room for model improvement. Users are also in an information blind spot, not knowing how their feedback affects the model, or even confirming whether the feedback is adopted. Some model leaderboards and crowdsourcing platforms attempt to improve transparency, but they often do not allow users to trace their contributions or provide tangible rewards for participation. Platforms that claim to be fair and transparent often rely on good faith rather than enforceable standards.

We believe that cryptography can bring transparency and ownership to this gray area of ​​AI. Blockchain can help contributors easily obtain rewards, while providing AI developers with reliable guarantees of the quality and source of feedback data. Users are incentivized, developers get trusted data, and everyone can achieve two-way auditing in this open market. To this end, we led the $33 million seed round of Yupp, a consumer product that allows everyone to explore and compare the latest AI models for free.

Yupp uses a crowdsourcing model for model evaluation: users enter a prompt word, view multiple AI-generated responses in parallel, and then select the best answer. Their choices generate a digitally signed "data package" of preference data that is extremely valuable for post-training optimization and evaluation of AI. Users can not only use the latest models for free, but also receive rewards based on the feedback they provide.

Yupp’s design turns human judgment into a renewable economic resource. As new interaction data is generated, old data “expires”, forming a natural flywheel: more adoption leads to fresher evaluations; fresher evaluations lead to better models; better models lead to more usage. All participants – both regular users and AI model developers – can join in and see transparent rules that apply to everyone, ensuring the market remains credibly neutral. No one can hide scores, and no one can manipulate rewards or outcomes.

The founding team has deep experience in both AI and crypto. They worked together on consumer-facing machine learning products in the early days of Twitter. Pankaj Gupta was the global head of consumer engineering at Google Pay and Coinbase, and Gilad Mishne was the head of machine learning at GoogleX. The startup team brings together senior engineers from Google, Coinbase, and top labs.

AI requires a robust and reliable evaluation system based on large-scale human input, and crypto is the trust machine that achieves this goal. Yupp is committed to becoming the basic evaluation layer of future AI by allowing users around the world to contribute feedback to improve models. We are honored to support Yupp and look forward to assisting them in building a closed feedback loop on the chain to ensure that the results of AI innovation are shared by all builders.