Page 17 - AI Vol 1: Foundations of AI

scenarios where the model could fail or be 7. CHALLENGES AND MECHANISMS IN
misused. This process allows developers to build
in safety features to prevent these issues from
occurring in real world situations.
Constitutional AI, as developed by companies
like Anthropic, is an example of an advanced
guardrail strategy. This approach embeds ethical
guidelines and operational constraints directly
into the LLM’s decision-making process which
helps to ensure that the model’s outputs remain
safe, fair, and aligned with the intended use.
For public agencies considering AI solutions,
assessing the safety and reliability of these
products is critical. Established AI developers
like OpenAI, Google, Microsoft, and Anthropic
typically have robust safety protocols in place.
However, when it comes to lesser-known AI
startups, agencies should exercise due diligence.
Evaluating the startup’s approach to safety,
including their use of red teaming and other
guardrail strategies, is essential to mitigate the
risks of inappropriate or harmful AI interactions.
FOUNDATIONS OF AI | LOZANOSMITH.COM
AI INFORMATION PROCESSING
Another key concept in understanding LLMs is
the “context window” which refers to the input
data analyzed by the model to generate inference
and output. This is analogous to a person’s short-
term memory when having a conversation: the
amount of recent dialogue they can recall and use
to inform their current response. In the context
of LLMs, this “window” determines how much
of the text the model can “see” and use to make
decisions about what to say next. This includes the
user’s inputs, the entire text of the conversation
(i.e., previous interactions in the same chat
session), any documents provided by the user to
the model, and any information the model itself
retrieves for its use in generating a response.
Different models have varying sizes of context
windows, which significantly impacts their
performance. OpenAI, GPT-4 Turbo has a
context window of approximately 100,000
tokens, which can be equated to around five
hundred pages of text. The latest model from
Claude2 has a context window of 128,000
tokens, roughly equivalent to three hundred
pages and Google’s recently announced Gemini
1.5 can be used with a 1,000,000 token context
window, which is well over 1,000 pages of text.
The context window’s effectiveness is not just
about its length. Due to the attention mechanism
inherent in these models, not every part of the
input within the context window is given equal
weight. This mechanism enables the model to
focus on more relevant parts of the input while
generating responses. However, this also means
that even with large context windows, an LLM
VOLUME 1 | 17