Decentralized Reinforcement Networks for Large Language Models
The purpose of this article is to highlight the many privacy challenges that plague enterprise-AI solutions. Devr technology not only solves these privacy challenges, but our privacy routing solution creates organic, self-organizing decentralized reinforcement networks that can supercharge LLMs.
Large Language Models (LLMs) like ChatGPT accelerate productivity, but they come with significant risks and liabilities. Enterprises are blocking employee access due to IP concerns [Ref. 1]. Regulators are scrambling to catch up with the pace of AI advancement [Ref. 2].
It will be challenging for the many companies entering the AI race to deal with the turbulence of emerging tech markets.
Devr is a new Internet protocol for the governance of decentralized privacy networks (DPN), powering a new era for data sharing economies. Devr solutions address the challenges posed by traditional data sharing infrastructure, which has remained unchanged since the early 2000’s.
DEVR PRIVACY STACKS
Devr Privacy Stacks (See Fig 1) provide privacy templates to accelerate innovation in key industries and market verticals. Our stacks provide the basic privacy framework, along with tools, designs, and source code to jump-start your innovation in emerging technology.
Figure 1
Currently available privacy stacks include:
- Patient Opt-in: Allow patients with smart health devices to regulate their own privacy
- Expert Systems
- Industry 4.0: Rapid PoCs for predictive maintenance at scale
- Prescriptive Medicine: Instant solutions for prescriptive medicine
- Decentralized Fraud Prevention: Innovative privacy-enhanced fraud detection
- Data Safety for Enterprise-AI: Governance tools for enterprises to safely interact with LLMs
- Content Creation: Build tools to protect creator and derivative IP with granular traceability
DATA SAFETY FOR ENTERPRISE AI
The Devr Privacy Stack for Data safety for Enterprise-AI provides data governance tools for enterprises to safely interact with LLMs. AI/ML/LLM solution providers can rapidly develop full data economies with enterprise customers by bundling Devr's Privacy Stack for LLMs into their architectures.
The Privacy Stack for Enterprise-AI includes:
Browser Plugin
Devr’s Trust Plugin transforms any data producer into a Web3-encrypted-endpoint, allowing the data privacy journey to begin. Our plugin for browsers turns Web 2.0 applications into Web3-encrypted endpoints, ensuring that any consumer data collected is encrypted before hitting the wire.
Devr Safe for Consumer and Employee Data
Devr Safes provide Web3-encrypted storage for consumer, patient, or employee data. Data stored within a Devr Safe is accessible through tightly enforced credentials to ensure data protection for sensitive data.
Smart Blocker Application
The Devr Smart Blocker is an innovative utility, which allows enterprises to define and regulate data policies for how employees interact with 3rd-party AI solutions such as LLMs.
The journey towards enterprise data safety begins by understanding the difference between data and metadata. In the figure below (Figure 2), it is important to note the difference between the QUESTION asked by an employee and the surrounding METADATA, which is tightly associated with both the employee and the QUESTION.
Figure 2
The enterprise must have access to both the QUESTION as well as all the METADATA in order to understand how and by whom the 3rd-party AI solution is being used. While it is important for the enterprise to have access to the METADATA, it is equally important to hide/obfuscate the METADATA before sending the QUESTION to the 3rd-party AI.
Devr's Privacy Designer provides an intuitive user interface to assign privacy filters at a per-cell, or per-field level. Privacy filters can range from zero-privacy to full-privacy, and the full spectrum of filters in-between. See Figure 3.
Figure 3 describes the DPN between an enterprise and a 3rd-party AI solution provider
Figure 4
Figure 4 describes the flow along the DPN between an enterprise employee and a 3rd-party AI solution provider.
- An employee posts a question to the LLM
- The Smart Blocker intercepts and approves the question
- Dynamic privacy policies are applied to the question
- The LLM responds, and the Smart Blocker intercepts the response
Figure 5
Figure 5 describes a successful question + successful response.
- An employee asks a question
- The Smart Blocker approves the question
- The LLM responds to the question
- The Smart Blocker approves the response
These four transactions are recorded on the DPN.
Figure 6
Figure 6 describes an unsuccessful question.
- An employee asks a question
- The Smart Blocker rejects the question
These two transactions are recorded on the DPN.
Figure 7
Figure 7 describes a successful question + unsuccessful response.
- An employee asks a question
- The Smart Blocker approves the question
- The LLM responds to the question
- The Smart Blocker rejects the response
These four transactions are recorded on the DPN.
Figure 8
Figure 8 describes a crucial aspect of the DPN. The Smart Blocker represents the governance policies of the individual enterprise using the LLM service. Mapping successful question-response combinations and unsuccessful question-response combinations from all the individual enterprises within a DPN provides powerful and high-quality reinforcement training data for LLMs.
Please subscribe to our newsletter and visit our website to understand more about our tools and solutions for DPN. You can explore our Devr Privacy Stacks, templates that provide pre-built privacy acceleration for different market verticals. We provide a free tier that allows you to explore basic concepts of privacy and understand how data can be granularly regulated to address a dynamic and growing list of risks and liabilities.