How Centralized AI can be uselessly complex compared to Decentralized AI?
STORY OF A COINCIDENCE TURNED TO INNOVATION AND DANGER FOR HUMANS
In 2019, a small mistake reshaped the way people interact with the Internet. GPT-2 model creation, which is a improvised version of Generative pre-trained transformer model launched by Open AI in 2018 as the first iteration of the language model.
GPT-1 was formally designed to deliver consequent data out of the books to the internet, within a few clicks. It was trained with the exerted data from the physical books.
What are Generative Pre-trained Transformer models?
Generative Pre-trained Transformers (GPTs) are machine learning models that are used to formulate language processing tasks. It’s feeded with large number of data sets from sources such as Books, Wikipedia, Research pages, etc. to curate contexually relevant and semantically correct summaries.
In simple meaning, GPTs are enhanced versions of search engines which are independent and at the same time sharply updated language model on the surface of Internet.
Hence, they can be refined for a range of natural language processing tasks, including question-answering, language translation and text summarization.
The tale of an unfair decision reveals flaws in centralized models
GPT-2 model was trained with large sets of data, upto 8 million different web pages across internet. It turbo influxed huge understanding of common human needs on the internet. With right prompts, anyone could turn a single phrase to essay or translate a whole document.
However, GPT-2 had its share of constraints. It performed impressive with least reasoning tasks and massively lacked on questions that requires human kind-off thinking to answer, in an opinionated way.
Another issue they noticed was that it didn’t stick to basic human rules. It kept answering about things it shouldn’t, like how to do human trafficking, terrorist plans, and finding loopholes in the banking system. This worried Open AI because it’s important to follow rules and talk about safe things. Open AI needed to make sure it learns what’s okay to talk about and what’s not.
OpenAI employed a new method called “Reinforcement Learning from Human Feedback” (RLHF) to refine GPT-2. RLHF involves training an initial model, known as the “Apprentice”, to generate responses based on human feedback. This feedback is provided by a small group of evaluators who rate the responses according to predetermined guidelines. A separate model, the model, learns to emulate these human ratings. However, the models can be misled by the Apprentice into producing nonsensical but pleasing responses.
To counter this, a “Coherence Coach” is introduced, focused solely on generating coherent text. Together, these coaches guide the Apprentice to produce responses that are both coherent and aligned with human values. However, RLHF wasn’t intended to have a control over the model for not processing inappropriate content, but rather to improve the quality of GPT-2’s responses.
Problems with full control over LLM’s
- Security concern: Centralized systems, such as those used by many businesses and organizations, are often targeted by hackers seeking to access sensitive data, posing a threat to user privacy and the security of personal information
- Misleading users: It’s often found, governing bodies compel LLM companies to shape public perception by integrating flawed data into their models.
- One-size-fits-all philosophy: Centralized LLMs often prioritize a generalized approach, neglecting the unique needs and preferences of individual users.
- Revenue-first mentality: Big tech companies may prioritize profit over user experience and social impact, leading to compromised solutions.
