LLM systems
When we interact with LLMs today, we’re generally interacting with an LLM system that has a combination of a model, tools and agentic loops.
LLM based AI systems are made up of:
- Large Language Models themselves, which can be broken into:
- Pretraining - processing the internet’s scraped information to build the core token prediction capability
- Supervised fine tuning - giving the model curated examples of good questions and responses
- Reinforcement learning - tuning the model with feedback from a separate neural network
- Tools
- Tools give LLMs access to the world in real time. Think of the LLM like the brain and the tools like arms and legs. Common examples include:
- Web search. If a user asks about something that might benefit from current web information. (e.g. the news)
- Reading or writing files on your computer or a remote system like google drive.
- Tools give LLMs access to the world in real time. Think of the LLM like the brain and the tools like arms and legs. Common examples include:
- Agents
- A combination of LLM calls, tools and specific prompts, often with a non LLM coordination loop, to get to an objective. Examples:
- Open AI’s Deep research
- Any customized AI chat bot embedded in any traditional application
- Vibe coding platforms like Lovable, Bolt.new
- A combination of LLM calls, tools and specific prompts, often with a non LLM coordination loop, to get to an objective. Examples:
How this works in practice with ChatGPT
Ask: “What is the best vegan restaurant in Redwood City?”
Plain LLM response (Response based purely on training data) - 3 seconds
ChatGPT said:
In Redwood City, the top vegan-friendly restaurants (by consistent reputation and quality) are:
🌿 Back A Yard Caribbean Grill (Veterans Blvd)
While not fully vegan, this is a local favorite for flavorful plant-based options like jerk tofu and curried vegetables. It’s one of the few spots where even non-vegans rave about the vegan dishes — spicy, rich, and generous portions.
This is clearly a guess at the first best option from trained memories. Not useful for this kind of question.
With Web Search (Basic tool use: pulling some recent data) - 30 seconds
ChatGPT said:
One of the best options for vegan-friendly dining in Redwood City is Ike’s Lair
✅ Why it stands out
-
The restaurant is identified specifically as a vegan-friendly spot in Redwood City.
-
It has a solid review base (4.4★ from ~385 reviews) indicating many people enjoy the experience.
-
It’s centrally located (2655 Broadway, Redwood City) which makes it convenient.
Clearly still just returning one of the first options from a basic search
With Deep Research (Reasoning / planning + search tool use) - 15 minutes
ChatGPT said:
Redwood City currently lacks any fully vegan, plant-based restaurants within its city limits. Local vegans often note that the mid-Peninsula region has “a lack of fully plant-based restaurants”, especially compared to San Francisco
Given this reality, local vegans and flexitarians often venture to neighboring cities for a true plant-based dining experience. Below are two top-rated fully vegan restaurants near Redwood City that fit the criteria (casual atmosphere, mid-range price, loved by vegans and non-vegans alike). These establishments are not in Redwood City city limits but are within a 5–10 mile radius and are frequently visited by Redwood City residents seeking quality vegan dining.
Much closer to reality. There is no true vegan restaurant in Redwood City
Takeaways
Without tools like web search or agentic features like deep research, any questions to an AI chatbot are going to only be reliable for information that is extremely common on the internet.
Use tools. When you have time, use agents that run multiple searches and cross reference. Find the balance between time and sufficient accuracy for your problem.