Generative AI
“Generative AI, also known as gen
AI, is a subset of artificial intelligence that can create original content
such as text, images, videos, audio, or software code in response to user
prompts. This technology relies on sophisticated machine learning models called
deep learning models, which simulate the learning and decision-making processes
of the human brain. These models identify and encode patterns and relationships
in vast amounts of data, enabling them to understand natural language requests
and generate relevant new content”. Gen AI gained mainstream popularity in 2022
with the launch of Chat GPT, which introduced a user-friendly interface in
order to be able to “prompt” the model to get the desired output (text, image,
video generation, etc).
Gen AI starts with foundation models,
which is a deep learning algorithm trained in millions of data. The goal is to
predict the output, for example the next word in a sentence or the next element
in an image. We see clearly now what was discussed in previous articles: neural
networks are used for predictions, and in this case the goal is to data. The
reason that Gen AI achieved popularity only now is that the training requires thousands
of GPUs (Graphic Processing Units) which require intensive computational power
and millions of dollars! Once the model is trained, it must be tuned: that is
refined for a specific content generation task. This can be done by fine tuning
(e.g. nurturing the model with thousands of labeled examples with the desired
output), or reinforcement learning with human feedback (which we have already
discussed as the mathematician providing mathematical equations examples and
giving reward or penalization depending on the answer). In a final step, the
model is evaluated, which can give way to further training and tuning.
These models are referred to as
Large Language Models, referring to their capacity to handle large amounts of data
and that they are based on text. The big innovation came about in 2017 with the
appearance of transformers in the paper “Attention is All You Need” by Vaswani and others. A transformer
is a deep learning model architecture that processes sequences (like text or
time series) using a mechanism called self-attention. This allows the model to
understand the relationship between all elements in the sequence at once,
rather than step-by-step. Transformers can analyze all parts of the input simultaneously
and learn long-range relationships—making them more powerful and faster than
older models like RNNs or LSTMs. This explains why GPT (Generative Pre-Trained
models) had a breakthrough in recent years.
Another important concept is RAG
(Retrieval Augmented Generation). Without it, an LLMs would only base it responses
in its training data. But RAG pulls in information from a new data source (for
example the internet), combining both it´s training data with external data
creating better responses. Meaning the responses that we get are not only based
on the data the model was trained on, but on the whole internet or external data
we provide!
While previous models were
rule-based, meaning for example you provided a set of predefined rules and
expected an outcome; by predicting the next dataset Gen AI is able to generate
new content. These models excel in the generation of text, audio, image, video,
code, etc. Giving an example of text generation, Gen AI can provide you with a
complete and thorough business plan for your new product, with a result
surpassing a human!
Some of the applications for
business include the following:
1) Chatbot
for customer support: remember in the 20th century when a real person answered
each query that came about through media channels? This was costly and ineffective!
A model can be trained with the most typical questions the users ask, and
generate a response right away! If the customer is not satisfied, the
intervention of a human can always be requested.
2) Content
Creation: Marketing agencies are using it to generate content, which can be
automated and launched in a schedule.
3) Content
Summarization: remember when a lawyer would spend hours reading a contract? Large
amounts of text or even video can be summarized to obtain the appropriate conclusions,
reducing human error.
4) Supply
Chain Optimization: SAP has it´s own LLM, SAP Joule, which enables the creation
of quotes, purchase request or purchase orders by a simple prompt. Hours saved
in data processing!
5) Predictive
Analytics: the nature of prediction of the model enables forecast trends and
customer behaviors which allow proactive management strategies.
Consider that, whereas training
an LLM model is costly and time consuming, leveraging on existing LLMs to
create your own AI-powered apps is cheaper than ever. The whole goal is not to
reinvent the wheel, since the battle for LLMs is already fought by US and China.
And even if some larger corporations have developed their own LLMs, the goal of
this essay is to show you that in a post-industrial society knowledge replaces
capital as the most expensive commodity. As an entrepreneur, don´t expect to
create the next Chat GPT: use existing technology to leverage your AI-powered
products and reach the market in a flash!
What
is a Transformer Model? | IBM
What
is Generative AI? - GeeksforGeeks
Resource
Allocation Graph (RAG) - GeeksforGeeks
No comments:
Post a Comment