A 500% surge in one day, understanding Yzi Labs' investment in Tensorplex Labs?

Author: BUBBLE, BlockBeats

This morning, Yzi Labs announced an investment in Tensorplex Labs, which immediately caused its Bittensor subnet project Dojo to surge by 500%. Last year, decentralized AI startup Tensorplex Labs completed a $3 million seed round of financing. A year later, Tensorplex Labs has made contributions to the Bittensor ecosystem in various areas, including Dex, data dashboards, and AI. This article will briefly analyze what the two-year-old Singapore company Tensorplex Labs has done to become a core participant in the Bittensor ecosystem.

The DeFi Core of Bittensor

Staking and Lending

Before the release of dTao, the Bittensor ecosystem was quite lacking. Tensorplex contributed core DeFi products within the TAO ecosystem, one of which is Tensorplex Stake, a cross-chain staking product that launched stTAO, a LSD ERC20 "Liquid Staking Derivative" token. This allows EVM investors to participate in the TAO ecosystem and enjoy staking rewards without downloading a Substrate-compatible wallet, while also manually delegating. Additionally, it enables staking assets to be bridged to networks like Bittensor to participate in validator operations, for which Tensorplex established a minimal trust bridge between Substrate and EVM. Furthermore, it incubated lending protocols like TaoBank, allowing users to obtain interest-free loans using staked TAO as collateral. The peak TVL for staking reached over $30 million, making Tensorplex one of the largest validator nodes on Tao.

Exchange + Data Dashboard Integration Backprop

Just last month, as dTAO was officially launched, @BackpropFinance also officially went live. At this point in time, as one of the three Dexs that can trade dTAO, "hotkeyswap, taostats" not only allows users to trade various Alpha subnets but also provides a comprehensive data dashboard that integrates trading dashboards, address analysis, on-chain information queries, market analysis, and various other products into one. The familiar UI combined with a complete data dashboard significantly lowers the participation threshold for new users joining the TAO ecosystem, making Backprop the most popular Dex product in Bittensor currently.

Development of the AI Ecosystem

LLM

Tensorplex has made multiple attempts in large models, creating a model that surpasses all open-source counterparts in the decentralized LLM pre-training competition of @MacrocosmosAI "Subnet 19". With over 60 Bittensor subnets and validators running on AI chains like @flock_io and @AlloraNetwork, the latest top model Sumo was released on Bittensor Subnet 9 in 2023. After increasing the parameter limit to 7 billion in April 2024, the performance of Sumo-T9-7B even surpassed well-known models like Falcon 7B and Llama 2 7B.

Tensorplex Stream

Tensorplex has also made attempts in the area of Agents. Tensorplex and @Corcel_X "Subnet 18" built the first Web3 RAG chatbot "Tensorplex Stream", which gathers information from cryptocurrency podcast records to train the AI Agent, issuing a summary and compilation of the latest cryptocurrency information. Now, in addition to the Chatbot, they have developed Podcasts, allowing the AI to "speak" the information it has collected. The advantages of RAG in collecting the latest information in the rapidly evolving cryptocurrency market have also been widely applied.

Dojo Subnet

The Dojo Subnet has become the most prominent subnet in dTAO within the past 24 hours, surging by 500% after Yzi Labs announced its investment in Tensorplex Labs. Dojo is a data collection platform, but it is not just a data collection platform; it integrates humans as observing third parties into a comprehensive ecosystem. Human intelligence can directly influence AI development, training, and agent operations both on-chain and in the real world.

The main workflow allows anyone to generate unique synthetic tasks using LLMs, which could involve game physics feedback, image generation quality, physical algorithms, etc. Through Dojo, "artificial" feedback data is collected and added to response obfuscation techniques.

Tasks are then distributed by miners, and the work of human validators is to confirm the basic facts generated by the LLM, determining the quality of responses provided by various contributors and allocating rewards based on multidimensional metrics: human consistency "weighted Cohen's Kappa", quantifying stability "intraclass correlation coefficient ICC", and factual fidelity "Spearman correlation". This data will be used to improve open-source models and optimize them.

Dojo will use technical means to prevent Sybil attacks and ensure that contributors are real humans. "After the system matures, roles to prevent attackers will be added, forming a contribution chain." So far, Dojo has collected over 3 million human-generated data points.

The investment in Tensorplex Labs by Yzi Labs is just a microcosm of the gradual market acceptance of the Bittensor ecosystem. The "halal" Bittensor generally does not receive much attention in the market. In fact, as an established AI ecosystem, the various project parties within it possess considerable capability and patience, creating a serious atmosphere of building. Although there are still issues such as shell projects in the subnets or the dTAO system not being fully developed, we still look forward to Bittensor producing more excellent products. BlockBeats will continue to pay attention.