Introduction
In an increasingly AI-assisted world, decision-makers have a vast array of tools at their disposal to help them navigate complex landscapes. However, the introduction of AI-powered tools has opened up unprecedented opportunities for enhanced decision-making, enabling businesses to leverage machine learning models to gain deeper insights, streamline their operations, and make more informed decisions. In this blog post, we will explore a five-tier framework of AI tools that illustrate the potential of AI in transforming business processes.
This framework ranges from base language models to advanced systems, illustrating the spectrum of tools that AI offers. Each tier represents different capabilities, from gathering and interpreting extensive information, extracting valuable insights from complex data, customizing outputs to fit specific needs, to streamlining workflows. These tools, when leveraged appropriately, can aid in more effective decision-making.
We will delve into the specifics of each tier, discussing tools such as GPT-4 accessed via API using Magai, Anthropic's "Claude" model for enhanced context understanding, document vector embedding tools for mining contextual information, advanced fine-tuning tools like Microsoft's Azure, and the integration of AI into existing workflows with tools like GPT for Sheets for Google Sheets.
We will also explore the differences between accessing GPT-4 through ChatGPT and pure API, discussing behavioral controls, customization, required expertise, and potential use cases.
Whether you're an expert who knows exactly what you want from your AI model, or you're just starting to explore the benefits of AI for your business, this blog post will provide you with valuable insights. Stay tuned as we unravel the potentials of each tier, the technical specifics, and the practical considerations for implementing these tools in your business.
Here’s how we currently see things. With our outstanding questions and all! We hope it helps!
First, some terminology
- Raw model access: Users now have the ability to easily access raw language models through services like Magai, providing more precise and reliable responses and allowing the easy use of system prompts specific to user needs.
- Context Awareness: Tools like Anthropic's "Claude" can handle significantly more context (up to 100,000 tokens of about 70,000 words) than GPT-4 (about 4,000 tokens, or 3,000 words). This capability opens up new use cases, including using extensive information about clients and their business problems, or analyzing lengthy transcripts or interviews.
- Including hidden gems from your documents: Tools that allow users to upload documents and perform semantic searches within the embedded content provide a new way to uncover and draw upon relevant information within large datasets to enhance model responses. This can be particularly valuable when you need to find and use insights hidden in extensive or complex document sets.
- Fine-tuning: More advanced tools, like Microsoft's Azure, offer document upload, vector embedding, semantic search, and additionally allow model fine-tuning based on the uploaded content. While currently requiring a significant investment of planning and expertise, these systems are expected to become easier to use in the future, opening up new possibilities for AI-enhanced decision-making.
- Workflow Integration: The ability to embed AI in existing workflows, such as using the GPT-4 add-in for Google Sheets, offers powerful possibilities for automating routine tasks and streamlining processes. This level of integration can greatly enhance productivity and efficiency.
Five tiers of tool
The five tiers detailed represent a range of AI tools that enable more efficient and informed decision-making in business settings. From base language models to advanced systems, these tools use AI to gather and interpret extensive information, extract valuable insights from complex data, customize outputs to fit specific needs, and streamline workflows, all to aid in more effective decision-making.
- Tier 1 - Base Language Model (Technical: API Access to GPT-4 via Magai): This tool involves using GPT-4, a top-performing language model, through API access via a service like Magai. The API access allows the user to leverage the pure, unmodified model, enabling precise and reliable responses tailored to the user's specific needs. Additionally, Magai offers a customizable system prompt, enabling users to give the model a role, language, and behaviors suited to their business. This tool is essential for experts who know exactly what they want from their AI model.
- Tier 2 - Enhanced Context Understanding (Technical: Claude's 100K Model): The second tier utilizes "Claude", a model by Anthropic, which can handle an extensive amount of context (up to 100,000 tokens or 70,000 words). This capability opens up new possibilities, such as embedding vast amounts of context about the client and their business problems, or massive amounts of interviews or transcripts, providing the model with a much richer source of information for generating answers.
-
Tier 3 - Contextual Information Mining (Technical: Document Vector Embedding Tool): The third tier introduces a tool that permits users to upload multiple documents and then uses vector embedding of the content. When a model like GPT-4 is used in this environment, it draws on the uploaded documents using semantic search to help enhance the response, adding any useful context found in the documents to the prompt and response. This tool is excellent for finding "hidden gems" in client documents, adding a new dimension of context to AI responses.
Outstanding questions: What types of documents that are most effectively used, how it determines what constitutes a "hidden gem", and how much additional context can be used in the prompt. What are the constraints on the volume and types of documents that can be uploaded?
-
Tier 4 - Advanced Fine-tuning (Technical: Azure from Microsoft): The fourth tier involves a more advanced system like Microsoft's Azure, which not only allows document upload, vector embedding, and semantic search, but also fine-tunes the model based on the uploaded content. While currently requiring a significant investment of planning and expertise, these systems are expected to become easier to use, allowing users to further personalize their AI tools.
A key question here is how the model fine-tuning process works and what skills are required to effectively use it. What are the constraints on the volume and types of documents that can be uploaded? What kind of planning is involved, and what kind of expertise is currently needed? Then: more clarity on the types of documents that are most effectively used for fine tuning, how well it draws on this new context, whether it reduces the impact of the initial model’s training (this is likely!).
- Tier 5 - Workflow Integration (Technical: GPT for Sheets for Google Sheets): The final tier introduces GPT for Sheets, an add-in for Google Sheets that allows users to run queries against GPT-4 in individual cells in spreadsheets. This tool is especially useful for structured inputs and outputs, enabling the completion of tables or running prompts against multiple inputs. It's a critical part of the workflow, integrating AI capabilities directly into commonly used software. This category will massively balloon in the coming days and weeks!
APPENDIX: What are the ways in which ChatGPT access to GPT-4 is different to pure access to GPT-4 via the API?
ChatGPT and pure GPT-4 accessed via API do utilize the same underlying model (GPT-4), but the manner in which they're used and the experience they provide can differ in several key ways:
- Behavioural Controls and Safeguards: ChatGPT typically includes several behavioral parameters and controls that are put in place to ensure that it can be used safely and effectively by a broad range of users. These controls might prevent the model from generating inappropriate or harmful content, or they might guide its behavior to be more helpful in general conversation. On the other hand, accessing GPT-4 directly through the API usually involves fewer of these controls and modifications, meaning that the outputs are more directly determined by the underlying model and the input it is given.
- Customization and Flexibility: When accessing GPT-4 through the API, users generally have more control over the behavior of the model. For instance, they can modify the temperature and max tokens parameters to control the randomness and length of the output, respectively. Users can also input a system level prompt to instruct the model about the specific role it should play or the language style it should use. This level of customization might not be available or might be more limited when using ChatGPT.
- Expertise Required: Using GPT-4 via API might require more technical knowledge and understanding of AI models than using ChatGPT, as users have to construct their own prompts and manage the model's parameters directly.
- Use Cases: Due to the high level of customization possible with API access, it can be better suited for specific or niche use cases where a high level of precision is required. On the other hand, ChatGPT is designed to be a general-purpose conversational agent and might be better suited for a broader range of everyday tasks.
Remember, these differences might not be applicable in all cases and the exact differences can depend on the specific API or platform being used to access GPT-4.