Download App
iOS & Android

Negentropy Capital: Why We Invested MetaGPT?

Cointime Official

ChatGPT, while popular, is not without flaws

For nearly a year, the crypto market has been in the shadow of a bear market, as if nothing could get the market excited. However, ChatGPT, an AI text generator, was recently released. Its user base surpassed one million in five days, and tens of millions of people rushed to play it, igniting the global market's enthusiasm for AI technology.

Despite its success, general-purpose AI focused on natural language processing, such as ChatGPT, continues to face a severe lack of training data sets. The ChatGPT prototype GPT-3 makes use of five data sets:

1. Common Crawl: A data set with over 45 terabytes of web text that accounts for 60% of GPT-3 training data.

2. WebText2: A text data set derived from Reddit links that accounts for 22% of GPT-3 training data.

3. Books1: A collection of books from Project Gutenberg and other sources that accounts for 12% of GPT-3 training data.

4. Books2: An Internet-based book data set that accounts for 3% of GPT-3 training data.

5. Wikipedia: A data set that includes all English Wikipedia articles and accounts for 3% of GPT-3 training data.

Even if the five data sets are already rich, they contain too much noise, errors, duplicates, irrelevance, or inappropriate content (particularly due to the low quality of the Common Crawl, which provides 60% of the content). These are also severely downgraded because they are unsuitable for GPT learning and promotion.

However, no good data set of the same size is currently available for large language models.

As a result, the actual development of ChatGPT has been severely hampered. To solve these problems, MetaGPT's vision is to use Web3 motivation as the infrastructure for the 0-1 part of AI training.

MetaGPT: A bridge between AI creators and AI users

MetaGPT deftly combines AI and Web3, and its unique Train to Earn model encourages users to participate in the training and optimization of AI models in the Web3 industry. Kryptal, MetaGPT's upcoming AI robot, aims to provide users new to the Web3 world with more user-friendly access to the latest crypto industry knowledge base.

According to Cointime, Web3 investor Negentropy Capital officially announced a $2 million investment in MetaGPT.IA on February 10, 2023.

MetaGPT is also about to release its latest AI product, Kryptal, which will provide Web3 industry robots trained in vertical industry knowledge bases.

About the Data sets: MetaGPT -- Crowdsourced AI error correction mechanism

We discussed the scarcity of good data sets for AI training earlier.

MetaGPT, on the other hand, allows anyone to participate in the feedback adjustment part of the unsupervised large language model and Earn incentives (Train to Earn) to build more usable general-purpose AI. The precise procedure is as follows:

1. Use the OpenAI team's prompt dataset, which includes a large amount of prompt text.

These texts describe the task, such as "write a poem" or "tell me a joke."

The human tagger then provides the desired output for each prompt text, along with a score ranging from 0 to 5, indicating the output's quality.

As a result, a supervised data set is generated that can be used to train GPT-3 to produce output that is more in line with user expectations.

2. Manual feedback is required for further fine-tuning of GPT-3.

Create a feedback dataset with a large amount of user input, such as "Give me a joke" or "Give me a pie chart," and so on. The human feedbacks then provide the best output for each user input, along with a score ranging from 0 to 1, indicating the output's reasonableness. As a result, a reinforcement learning dataset is created, which can be used to train GPT-3 to produce more rational outputs based on user input.

Because it provides diverse, high-quality, and real-time data to help models better understand and satisfy user intentions, crowdsourced user involvement can aid in manual feedback tuning of large language models. MetaGPT, for example, allows users to rate or comment on the model's responses or suggestions, and then adjusts the model's parameters or policies based on user feedback to improve the model's accuracy, friendliness, and utility.

In this way, we can better align the model with the user on various tasks, increasing user satisfaction and trust.

Simultaneously, MetaGPT employs the GPT-3 open interface to enable anyone to train the Web3+AI project's open source fine tuning model on it. Anyone who trains a fine-tuning model suitable for a specific field has the option of opening for a fee or for free, with the fee going entirely to the model trainer. The free model will also reward MetaGPT-based equity tokens based on usage.

MetaGPT seeks to bridge the gap between AI creators and AI users in order to promote AI innovation and adoption. And this goal, by establishing a Train to Earn, may exist to achieve the possibility.

About Negentropy Capital

Negentropy Capital, founded in early 2020, is a globally diversified venture capital firm focused on the digital assets, cryptocurrency, and blockchain technology industries. It primarily consists of four functional departments: strategic merger and acquisition, strategic investment, asset management, and international cooperation. Metavers Ecological Fund, NFT Special Fund, and DeFi Special Fund are its three special funds. Negentropy Capital focuses on upstream and downstream blockchain investments, focusing on the blockchain industry's cutting-edge positioning and aiming to empower visionary crypto believers and teams through capital means.


All Comments

Recommended for you