SUMMARY

Insop Song from GitHub Next presents a webinar on agentic AI, framing it as a natural progression of language model applications. He covers language model basics, limitations, methods for improvement, and design patterns.

STATEMENTS

Language models (LMs) predict the next word in a sequence based on input text, leveraging probabilities derived from large training datasets.
Training LMs involves pre-training on vast amounts of text and post-training, including instruction following and reinforcement learning from human feedback to align with user expectations and preferences.
Post-training includes instruction following to train the models to respond to specific styles and align to human preference by using reinforcement learning.
Cloud-based APIs facilitate the use of language models by allowing developers to send natural language inputs and receive generated outputs.
Effective prompting requires clear, descriptive instructions, examples, and context, encouraging the model to reason through the problem before answering.
Breaking down complex tasks into smaller stages and chaining prompts can significantly improve the quality of the generated output.
Hallucination, knowledge cutoff, lack of attribution, data privacy concerns, and limited context length are common limitations of language models.
Retrieval Augmented Generation (RAG) mitigates these limitations by providing models with relevant context from indexed data, reducing hallucinations and improving attribution.
Tool usage, also known as function calling, enables language models to interact with external environments for real-time information and computational tasks.
Agentic language models interact with the environment through tool usage and retrieval requests.
Agentic models reason and act, breaking down tasks, planning actions, and using external tools to gather information, leveraging methods like chain of thought.
Planning, reflection, tool usage, and multi-agent collaboration are key design patterns in agentic language models.
Reflection involves the model critically evaluating its own output to improve subsequent generations, enhancing performance.
Multi-agent systems split tasks among different agents, each with specific prompts and personas, to handle complex automation scenarios.
Agentic language model usage is a progression or extension to existing language model usage methods.
The core of agentic language models involves iterative calls by the model, reviewing tasks, and making calls to external tools.
Experimentation is very important with different patterns to find out what's the most capable model that is relevant can push the boundaries and assist the development.
The reason agentic language model usage is getting more widely used is because if you have the same model, if you ask just direct request, model may not be able to handle it.

IDEAS

Language models can be hosted locally, even on mobile devices, if they are compact enough, expanding their accessibility.
Systematic tracing and logging are essential for debugging and auditing in language model-based development.
Automated evaluation, using either human evaluators or language models as judges, ensures continuous improvement and adaptability.
Employing a prompt router, which directs inputs to specific prompt handlers based on intent, improves operational efficiency and output relevance.
Hallucination is a well-known issue that models can sometimes oftentimes generate incorrect information, particularly if it's related to some computation.
RAG addresses the citation problem by providing the right references.
Graph RAG could also help to extract the more relevant information.
Reflection is a pattern that is quick to implement and then leads to a good performance.
There are companies that are trying to provide these services as decision on the right side, and research and analysis that gather information, synthesize it, and then provide a summary for the users.
Prompt usually consists of helpful AI agent. You could change that into a different persona to a different agent.
Multi-agent is an interesting way to implement or accomplish your complex task.
The agentic language model usage is a progression or extension to existing language modal usage method.
Most best practices that have been used in language model for simple cases, most of them are applicable for new use-cases.
Hallucinations could be concerning to the content that is being output.

INSIGHTS

The evolution of language models from simple next-word prediction to complex agentic systems demonstrates a significant leap in AI's capacity for problem-solving and interaction.
Effective prompt engineering is not just about asking the right questions but also about structuring the interaction to guide the model's reasoning process.
The integration of external tools and real-time data transforms language models from passive information generators into active problem solvers.
Retrieval-augmented generation is important for mitigating hallucinations and keeping the model truthful.
Reflection is a great pattern to implement and leads to good performance.
Ethical considerations, especially regarding hallucination and misuse of generated content, are paramount in developing and deploying language model applications.
Phased rollouts of models in different patterns improves output.
Agentic frameworks allow more complex tasks.
The ability to start simple, with a playground, has significant implications on development and research.
The use of evaluation, ethical guardrails, and mitigation techniques during the model design process helps safe interactions.

QUOTES

"A language model is a machine learning model that predicts the next coming word given the input text."
"Model cannot read your mind so that means you need to describe what you want the model to generate the output for you."
"So instead of asking direct question, you ask model to think through it or come up with its own solution."
"Without clear evaluation, it is hard to make forward progress or even hard to change the model, change the different type of models."
"Agentic language model usage can be defined as it could reason as well as it could action."
"By combining these reasoning and action, model can do a lot more complex task than simple input and output type of interaction."

HABITS

Systematic tracing and logging for debugging and auditing model-based development.
Automated evaluation from the early stages to track progress.
Breaking down complex tasks into smaller, manageable stages. In other words, thinking through the tasks requested.
Using a playground environment to quickly experiment with prompts. Is experimenting to make sure the tasks being output are on point.
Follow experts in Twitter or YouTube channels to get more updated information.

FACTS

Language models are now used in various applications, including AI coding assistance, domain-specific AI copilots, and conversational interfaces like ChatGPT.
Even with advancements, language models still face limitations such as hallucination and knowledge cutoff.
The field is rapidly evolving, meaning that models and methodologies are quickly improving and being deprecated.
Retrieval Augmented Generation involves pre-indexing data into smaller chunks and converting them into an embedded space using an embedding model.
Agentic language models can be used in decision support, coding, and research.

REFERENCES

ChatGPT
Chain of Thought (COT) method
Retrieval Augmented Generation (RAG)
ReAct (Reason and Act)
Python code
GitHub Next
Stanford Online's professional AI program
Twitter
Youtube

HOW TO APPLY

Write Clear and Descriptive Instructions: Provide detailed instructions to guide the model in generating the desired output, as models cannot read minds and need explicit guidance.
Include Few-Shot Examples: Supply the model with example inputs and outputs demonstrating the desired style and format to ensure consistent and relevant responses.
Provide Relevant Context and References: For tasks involving factual information, provide context and references to prevent hallucinations and ensure more accurate, reliable outputs.
Allow Time for Reasoning: Encourage the model to think through the problem and develop its own solution before providing an answer, using techniques like chain of thought.
Break Down Complex Tasks: Instead of asking the model to perform multiple tasks in a single request, split them into smaller, simpler stages.
Implement Systematic Tracing and Logging: Maintain detailed logs for debugging, auditing, and tracking the performance of language model-based applications.

ONE-SENTENCE TAKEAWAY

Agentic AI extends language models by enabling interaction with environments through reasoning, planning, tool usage, and iterative refinement.

RECOMMENDATIONS

Implement automated evaluation processes to continuously assess and improve the performance of language model applications.
Use a playground environment to start and test new prompts to avoid hallucinations and wrong doing.
Employ prompt routers to direct input queries to the most appropriate language model and prompt handler for efficient and relevant responses.
Start with simple language model usage and incrementally introduce agentic patterns to tackle more complex tasks.
Explore and apply design patterns such as planning, reflection, tool usage, and multi-agent collaboration to enhance the capabilities of agentic models.
Implement ethical guardrails to check generated output and filter out inappropriate or unsafe content.
Use a reflection approach for evaluation, asking the LLM to give a reference before responding to the question.
Begin exploring by following and referencing expert sources and courses to inform and deepen knowledge.

MEMO

Insop Song's discourse elucidates the concept of agentic AI as an evolutionary step in language model applications. He emphasizes the foundational role of language models in predicting subsequent words, a function achieved through rigorous training on expansive datasets. This training is bifurcated into pre-training, which imparts basic language prediction skills, and post-training, which refines the model's ability to follow instructions and align with human preferences through reinforcement learning.

The practical application of language models is facilitated through cloud-based APIs, enabling developers to craft natural language prompts and receive AI-generated responses. The effectiveness of these interactions hinges on the art of prompt engineering—the creation of clear, descriptive instructions that guide the model's reasoning. Techniques such as "chain of thought" and the provision of relevant context are crucial in eliciting accurate and coherent outputs.

Despite their advancements, language models are not without limitations, including the propensity for hallucinations, knowledge gaps, and concerns regarding data privacy. To mitigate these issues, Retrieval Augmented Generation (RAG) is employed, which augments the model's knowledge by retrieving relevant information from indexed data sources. Additionally, integrating external tools allows language models to access real-time information and perform computational tasks, broadening their utility.

Agentic language models epitomize this evolution, exhibiting the capacity to interact with their environment through tool usage and retrieval requests. Characterized by their ability to reason and act, these models adeptly break down tasks, formulate plans, and leverage external resources to achieve complex objectives. Key design patterns, such as reflection and multi-agent collaboration, further amplify their capabilities, enabling them to critically evaluate their own outputs and distribute tasks across specialized agents.

The agentic language model paradigm represents a significant leap in AI's problem-solving prowess. By combining advanced reasoning with real-time data and external tools, these models transcend the limitations of simple input-output interactions, unlocking new possibilities for automation, decision support, and knowledge synthesis. As the technology continues to evolve, ethical considerations and responsible implementation remain paramount to ensure its safe and beneficial deployment.

The development must explore open sources and experiment with different patterns to push