Level 1: Standard Large Language Model (LLM) Prompts
Using Coda AI columns with text prompts to ask the LLM to do something, and getting either a text response, or an image.
More advanced usage; using Coda formulas to inject data into the prompt and extract data from the response.
Level 2: LLM plus Live Web Search
Using Perplexity or Gemini packs to search the web as part of the LLM’s response, for immediate and up-to-date information.
Level 3: Retrieval Augmented Generation (RAG): Upload Your Business Data & Documents
Uploading your company’s own information (data tables, documents, media, etc.) so they are included in the responses and using the AI Assistant Magic pack to execute RAGs.
Level 4: AI Agents & Agentic Workflows
Building AI Agents: autonomous objects (Rows in Tables) that combine business logic with AI prompts and RAG to perform specific business tasks.
Then chaining several AI Agents together into an Agentic Workflow using Action Formulas to automate complicated Standard Operating Procedures (SOP).
Level 5: AI Agentic Workflows with User Interfacing (UI) Apps
Integrating these AI Agentic Workflows into existing Coda Documents which have User Interfaces (Forms or Dialogs) for human interactions.
Level 6: Agentic UI Workflows with Structured AI Languages (SaiL)
Advancing beyond imprecise ‘Prompt Engineering’ trickery to using ‘Structured AI Language’ (SaiL) prompts that are more like Coda formulas with for-each loops and if-then branches.
Business experts can easily define complicated logic and computations, beyond the ability of purely text-based prompting.
Level 7: Expert Systems that Learn from Experience
Building AI Expert Systems that use complicated Rules with ‘fuzzy-logic’, just like human experts, to solve problems autonomously.
These systems learn as they go, and refine the Rules or add new ones based on experience.
If there is any interest, I can post some “How-To” articles to this forum for any of the above.
Max
I’d love any insights your willing to share. In particular, what would interest me most is how you envision Level 7, ie which process you are looking to explore to achieve this agentic self-training.
As always, love following your thoughts and the fact that you share your findings with the community - much appreciated!
Is this a progression?
Or are some of these level optional/ in parallel?
For example, what is the difference between Expert Systems that Learn from Experience, and RAG enhanced systems? Is L3 data focused while L7 has some additional functionality?
E.g., I hope to get my company to the stage where (some) of their SharePoint data forms the basis for a L3 setup. Do we need to go from L3 to L4, L5 and L6 to get to L7?
This is not theoretical at all, but a solid framework.
I have built docs using levels 1 thru 6 for clients.
Level 7 is a work in progress, but looks promising.
It is based on work I did using ‘inference engines’ in LISP many years ago.
Note that I do not refer to AGI anywhere in this taxonomy.
I dont believe it will be available any time soon.
And we have a zillion business automations ready for current AI models right now.
I am on vacation and travelling across Europe a.t.m.
But when I return to my office, I will post examples, and how-to videos here.
Fears that AI is a bubble about to burst, or has (yet again) been over-hyped, are not true in my oppinion.
But getting a solid ROI from AI investments needs a very specific set of skills:
the ability to design good data structures (ie 3rd normal form DBs)
the ability to analyse business processes and build automation workflows
new skills around prompting, RAG engineering, AI agent building, and knowledge graphs
excellent UX and mobile UI designs that work with AI agents to empower the human in the loop
I believe that Coda provides an extremely versatile and capable platform for exploiting those 4 processes.
And that is why I am going all-in on using Coda AI to achieve the 7 levels of business AI mastery I outlines in the OP.
The SaiL prompt is a formula that tells the LLM to do the following;
store the colors of the rainbow as LIST
take each color in LIST and do the following
output the color as a level-3 heading
output its red and blue components as bullets
if the green component is zero, say it has no green component
otherwise show its green component
then repeat the loop with the next color
The purpose of SaiL is to be able to provide a complex set of instructions in a formula-like manner and NOT to have to try and define the logic in natural language.
Most no-code makers are well able to define logic in this formula-like way.
It is more concise, structured, precise, and understandable, than the long-winded, verbose, imprecise, hard to write, traditional natural-language prompts.
I totally agree. Artificial structures are more legible than prose in so many cases. Is SaiL a language some AI tools use? Or is it a technique you can use in any AI?
Yes, this intended to be a progression.
We began using AI at LEVEL 1 with just simple prompt/response tasks.
Then we discovered we needed to have the AI make use of the web in real-time to give better answers - which is LEVEL 2.
And so on… we kept adding more complexity to our AI use-cases.
So for us (my clients and myself) it was a progression. But you could skip to any level you like.
This level gets the LLM to search for answers within your companies own ‘Corpus of Knowledge’, instead of using its knowledge of the whole internet (its training data).
For this we began by building a Coda Table with one row for each chapter of the Technical Manual that we needed the AI to use when solving a particular problem.
We ran a process where the chapter was summarized, row by row, by Coda AI, into a single paragraph for each row.
We then concatenated these summaries into a single big text that showed the chapter numbers and their summary. This is only done once. The resulting summary text will fit into the LLM’s context, whereas the whole Technical Manual would not.
Then, each time we needed the AI to solve a problem, we had a three-step process.
We asked Coda AI to match the problem against the big summary of the document and choose which chapters were likely to contain the answer.
Then we asked Coda AI to look at the full chapter text to see if it contained the details needed, and respond with those details. This is done in a loop; once for each chapter selected in the first step.
Then we concatenate all the responses as the final response to the user; ie - a list of details retrieved from the manual that the Coda AI thinks are relevant to the problem.
This process is a version of RAG known as RAPTOR (Recursive Abstractive Processing for Tree Organized Retrieval).
However, we can now do all of this much better using the OpenAI Assistant Magic pack from @Troy_Larson. With this pack we can upload the whole manual as a single PDF and then query the resulting RAG Assistant in a single step.
This latter process is much faster because uploading the documents causes them to be converted into embeddings and stored in a vector database on the OpenAI server; and that is searched by the LLM much faster than our previous technique.
Update on LEVEL 7 - Expert Systems that Learn from Experience
Originally I was basing this on the work I did way back in the 1990’s using ‘Expert Rule-Based AI’ systems - specifically the XCON/R1 product from Digital Equipment Corporation.
This was used by card companies to detect fraud, by doctors to diagnose patients, and by aerospace companies to build airliners.
It uses a list of several thousand ‘rules’ that can be used to solve a complex problem. As the system proceeds, it discovers new rules and tweaks existing rules to improve the outcome.
It is not a neural network, but a higher-level set of ‘fuzzy logic’ rules that modify the resulting parameters when they are ‘triggered’. Back then it was very slow and tedious to use.
Today we can build a similar system as an Agentic Workflow, where each ‘rule’ is now a complex prompt and the ‘actions’ are further prompts that contribute to the final result.
I had tried to implement this in Coda using a GPT-4o pack I had built. But I must admit it was not going well.
Then OpenAI released previews of its latest GPT-o1 model. This model uses a similar process to the expert system by executing many prompt/response cycles, evaluating the interim results, and modifying the prompts. It executes a series of loops and branches involving many prompt/response cycles to handle larger, more complex problems.
So now we should be able to achieve the same LEVEL 7 system using GPT-o1 to represent the rules and the actions.
However, I must experiment further with this approach to see if it is possible.
So meantime, I must place LEVEL 7 outside the realm of practical business AI approaches.
Most recent LLM releases understand YAML - a human-friendly notation for presenting structured information (Its like JSON but much easier to read and write).
SaiL is a syntax for defining instructions and predicate logic using YAML.
So yes, all recent LLM models should be able to understand and execute SaiL instructions.
SaiL is still under development, I continue to refine the syntax.
Now that I am returned from my summer travels, I will publish the SaiL specification here shortly.
