LLMs vs Human Learning
Posted on 10 May 2023; 01:00 AM IST. Last Updated 07 Aug 2023; 09:15 AM IST.Summary: This article explains the key concepts and differences between LLM’s and human learning.
Large Language Models (LLM’s) are a recent invention, which have become very popular, and many frameworks like ChatGPT, which are based on LLM’s are now available, in the software industry. LLM’s could not only process natural language constructs, but could also generate natural language constructs like humans.
LLM’s are undoubtedly a great achievement, and the rest of the article dives into the various aspects of this new technology. The earliest LLM’s are based on Recurrent Neural Networks (or RNNs), and the current breed are based on Transformer model technology. The details of RNN’s and Transformer models are beyond the scope of this article.
What is explained in this article is -
- What is the LLM Learning Strategy?
- What problems LLM’s could solve?
- Can LLM’s make AI fly?
- Can LLM’s replace human workers?
- What could have been the goal of AI?
- What is the LLM Learning Strategy?
LLM’s learn by digesting very large corpus of text. This is a fairly expensive process, but that is how it works, and could be visualized as a “brute force” approach to learning.
Humans do not learn like LLM’s, since humans typically learn the alphabet of a language, then a few simple words, and finally sentences.
Most humans could appreciate the intrinsic beauty of a language, as they learn a language. For humans learning is an emotional process. These emotions when intertwined with other emotions like a scenic waterfall, could produce poetry. The generative scheme (of the language) in case of humans is “emotional”.
LLM’s cannot process emotions like humans, they makeup for this deficiency by imitating (or copying), what was injested already, in the form of a corpus.
- What problems LLM’s could solve?
LLM’s are best suited for -
a) content generation, story telling, etc.
b) text translation from one language to another.
c) simple chat applications.
LLM’s are undoubtedly a part of AI, but they constitute the fringe part of AI, and not the core of AI. In other words, LLM’s are like dessert and not dinner.
LLM's could induce a new crop of problems for social media platforms, since moderating automatic content generation (which is as good as humans), could pose many challenges.
- Can LLM’s make AI fly?
The answer or the opinion of the author is a big “NO”.
LLM’s are not data modelers or ontology builders. What this means is they could fail resoundingly in banking, insurance, retail, real-estate, other financial applications, and almost everywhere, where there is critical analysis and processing.
LLM’s could be making a grand attempt to learn ontologies from raw data, but for most organizations domain knowledge is not something that could be revealed publicly, and unlike a Shakespeare novel, process ontologies could be based on a dynamic runtime evolution.
In essence, to make AI fly, even in the simplest academic formulation, an LLM should spit out something like a petri-net, after processing natural language statements.
- Can LLM’s replace human Workers?
The answer or the opinion of the author is “Yes”.
The saddest part of this technology is that it could eliminate low paying help-desk jobs, and therefore it could make poor people even poorer.
- What could have been the goal of AI?
In the opinion of the author, AI has a great role to play to reduce the stress on mid-level workers, who build the nuts and bolts of systems.
What the author realized in his 30+ years of experience in the software industry is that “the younger generation is supporting a much larger economy than what the author and his peers did, when they were young”.
As economies expand, the tooling too should get more sophisticated, and to that extent “AI could be leveraged” to build a better and efficient society.
The typical workflow of most organizations, may be depicted as shown below.
Requirements → Domain Models
→ Product Dev.
→ Technical Support
→ Helpdesk
The bulk of the effort of any organization is absorbed in product development, and technical support.
AI could leverage “Ontologies” to build next-gen tools and technologies, which reduces turn-around time, and increases productivity.