Page 17 - AI Vol 1: Foundations of AI
P. 17
scenarios where the model could fail or be 7. CHALLENGES AND MECHANISMS IN
misused. This process allows developers to build AI INFORMATION PROCESSING
in safety features to prevent these issues from Another key concept in understanding LLMs is
occurring in real world situations. the “context window” which refers to the input
data analyzed by the model to generate inference
Constitutional AI, as developed by companies and output. This is analogous to a person’s short-
like Anthropic, is an example of an advanced term memory when having a conversation: the
guardrail strategy. This approach embeds ethical amount of recent dialogue they can recall and use
guidelines and operational constraints directly to inform their current response. In the context
into the LLM’s decision-making process which of LLMs, this “window” determines how much
helps to ensure that the model’s outputs remain of the text the model can “see” and use to make
safe, fair, and aligned with the intended use. 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
For public agencies considering AI solutions, 1.5 can be used with a 1,000,000 token context
assessing the safety and reliability of these window, which is well over 1,000 pages of text.
products is critical. Established AI developers
like OpenAI, Google, Microsoft, and Anthropic The context window’s effectiveness is not just
typically have robust safety protocols in place. about its length. Due to the attention mechanism
However, when it comes to lesser-known AI inherent in these models, not every part of the
startups, agencies should exercise due diligence. input within the context window is given equal
Evaluating the startup’s approach to safety, weight. This mechanism enables the model to
including their use of red teaming and other focus on more relevant parts of the input while
guardrail strategies, is essential to mitigate the generating responses. However, this also means
risks of inappropriate or harmful AI interactions. that even with large context windows, an LLM
FOUNDATIONS OF AI | LOZANOSMITH.COM VOLUME 1 | 17