Demystifying LLMs and Generative IA

El samuelo
3 min readDec 7, 2023

--

Intro:

LLMs, or Large Language Models, is THE technology beyond what we call Generative IA. It is based on machine learning and trained on trillions of parameters. This enables the generation of human-like text through neural networks that analyze and mimic language patterns.

Technology Core:

At their core, LLMs are a manifestation of advanced machine learning algorithms. They absorb enormous amounts of text data and learn to predict and generate language based on the patterns in this data. Based on a prompt or context given, the LLM will predict what would be the next word coming one after another… and create human like sentences…. Impressive? Yes as LLM coulb be used to summarize, translate instantly large amount of data and facilitate the human-machine interactions.

Revolutionary? In a way, yes! But the more we dig into it, the more we see their limits.

I see LLMs as an immense tool needed to be shaped, configured in order to create real value.

No Magic, Just Data:

There’s no mystical aspect to LLMs. I understand you could be amazed by how they operate and the interactions you can have with them. But, don’t forget that they operate based on the information they’ve been fed with … which is basically the information found in internet.
The quality of data can be and should be questionable, what do we know about the training data set of most of the LLMs ? Nothing. And this is a major concern, especially in a digitized world where fake news or news created by bots (including through LLMs are flooding).

There is and will be an intense race on LLMs and yes, after playing with a few of them, they are different. LLM could become a soft power weapon and legislations on them will come.

Limits of LLMs

I’ll try to expose the main limits that I observed on LLMs.

  1. Expertise Boundaries:
    While proficient in processing vast amounts of data, LLMs lack expertise in specialized fields.
    They might generate content that seems accurate but can lack the depth and precision required in complex or niche subjects. Human oversight remains crucial in such contexts.
  2. IA Hallucinations:
    LLMs, despite their sophistication, may generate text that seems plausible but lacks factual basis. These “hallucinations” occur due to inherent biases in the training data or misinterpretation of context, potentially leading to misinformation or inaccuracies.
  3. Past Data Limitation:
    Being trained on historical data, LLMs struggle to create entirely novel content or ideas. They often replicate existing biases present in the data they’ve learned from, making it challenging to break away from the limitations and biases of past information.

To sum up

Alright, enough of b*****g on LLMs. I am convinced that this technology could bring a ton of positive innovations to our world.

I think it is time to understand what does and don’t do the LLMs and use them only for what they do best : Generate content.

I guess what we all want is to use this technology without the limits expressed, meaning having deep, profound interactions with it and create/bringing value through them.

Plenty of innovations are being created to effectively use the LLMs in secured and specific way.

My next article will about the RAG (Rerieval Augmented Generation) a concept/tech which has the power to unlock the LLMs potential on domain specific context.

Thanks for reading,

++
Samuel Poutignat

--

--

El samuelo
El samuelo

Written by El samuelo

I vulgarize the blockchain & IA ecosystem's key notions as well as my learnings, my experiences and discoveries...

No responses yet