Using Online Models

Lesson Overview

In this video, we explore the concept of online language models (LLMs) and how they can be used to access up-to-date information from the internet. We discuss various techniques for grounding LLMs in different data sources and compare the capabilities of several online LLMs, including GPT-4.1, GPT-4.0 Search Preview, and Perplexity's Deep Research.

• 00:00: Introduction to knowledge cutoff dates in LLMs
• 00:29: Overview of online LLMs and their capabilities
• 02:04: Demonstration of GPT-4.0 Search Preview
• 03:18: Introduction to deep research models and Perplexity's Deep Research

Key Concepts

Knowledge Cutoff Dates

LLMs have a knowledge cutoff date, which is the point in time when they stop having information about the world. This means that LLMs may not be able to answer questions about events or developments that occurred after their cutoff date. For example, GPT-4.1, released in April 2025, has a knowledge cutoff date sometime in 2024.

Online LLMs

Online LLMs are language models that can query up-to-date information from the internet in a single call. These models allow users to access current information without the need for additional steps or techniques to inject context at runtime. Examples of online LLMs include GPT-4.0 Search Preview and Perplexity's Deep Research.

Deep Research Models

Deep research models are extended thinking time research agents that can spend more time exploring a specific problem, following links, and even reading PDFs to gather unique insights. These models can provide more in-depth and fully cited responses compared to standard online LLMs. Perplexity's Deep Research is an example of a deep research model.

Key Takeaways

1. Online LLMs can provide up-to-date information by querying the internet in a single call, overcoming the limitations of knowledge cutoff dates.
2. GPT-4.0 Search Preview demonstrates the ability to pull information from recent sources and provide concise answers to specific questions.
3. Deep research models, like Perplexity's Deep Research, can spend more time exploring a problem, resulting in more comprehensive and fully cited responses.
4. Users can create their own extended thinking time workflows by stacking multiple LLM calls, dividing problems, and synthesizing the results to achieve high-quality outcomes.
5. Experimentation with different online models and architectures is key to finding the best approach for a given task or workflow.

‍