Module 2 : Generative Artificial Intelligence
...
...
No. Human languages follow predictable patterns and structures that LLMs can identify and mimic with enough exposure. sophisticated pattern recognition, NOT AN UNDERLYING INTELLIGENCE that captures, retains and applies concepts. That is why people get nervous about feeding AI more and more because these tools will become more intelligent.
Do GenAI tools understand the human language?
Most GenAI tools are engaged with a "prompt", with the user requesting the tool to provide them with an output that meets the parameters they specify
How do we interact with GenAI?
With sufficiently large data sets and pattern exposure, neural Networks can mimic these patterns, producing "original" content that follows the conventions it has learned to identify
How to "train" the AI
ChatGPT and Eden AI
Name popular text-based GenAI
embeddings
What are tokens encoded into to create relational links?
embeddings are analyzed to understand the prompt
What do embeddings do?
AI is software and all software is unstable. Models are constantly improving, but small, seemingly insignificant changes can lead to unexpected behavior outcomes of these unexpected behaviors can have potentially major consequences
Why is Instability a concern with GenAI?
"jailbreaking" to free models to remove guard-rails and carry out illegal commands using AI models as a tool in cyberattacks enterprise models being compromised and being a "one-stop shop" for attackers to gain insight into critical company data, processes and decision making
Why is security a concern with GenAI?
A lot of GPTs are controlled by a select number of people. The transparency into training data, decision-making parameters and "guard-rails" is growing increasingly poor We cannot have clarity into how models are arriving at their answers.
Why is transparency a concern with GenAI?
tokens
a user enters a prompt to the AI tool of their choice. GPT breaks the prompt down into SMALLER INPUTS. What are these inputs called?
Neural networks
algorithms that can scan huge amounts of data, analyze it and identify patterns within it
GenAI
an algorithmic data processing and learning model designed to be able to dynamically create original content (content can be text, images, code, audio, and more)
its generated by calculating probabilities for most likely next tokens. this process is repeated until the response is considered complete. The response is de-tokenized back into human language. the response is subject to post-processings to validate quality/formatting the response is displayed to the user
how is a response generated from the AI tool?
large language models (LLMs)
language-based neural networks are known as..
Midjourney and Dall-e
name popular image-based GenAI
Fliki and descript
name popular video-based GenAI
- They do not "think" as we conceive of thought. it only PREDICTS based on large-scale pattern recognition - AI has no internal state representation of the world. (there is no notion of "truth" in an AI model) - All content AI generates is dependent on the data it is trained on - outputs often include falsehoods and/or "hallucinations" that can 1. include information that is inaccurate 2. create citations from nonexistent sources 3. respond irrationally -currently limited in both producing and summarizing long-form text - bad at keeping up with moving data (i.e. weather and news)
what are the limitations of AI?
