Dynamait

[Edited 2nd time] AI is not yet capable of dynamically creating and running software to produce a result.

The Dynamait framework running on the Taisker platform (task orchestration that AI uses to perform complex tasks and adhere to critical instructions) will change that. Perhaps in the end, making computer programs obsolete.

=== General description ===

Dynamait for AI is what a computer language framework is for programmers. The AI can now "program" its way through to accomplish tasks. There's a Human in the Loop, so it doesn't run amok.

Dynamait has tools and tool components, defined in natural language with structure: the component's goals, parameters that control it, intentions, program flow, instructions, context, and environment. Each component has verifications, tests, and examples.

Dynamait is a natural-language computer-interface framework that the AI, or more correctly, an orchestrated hybrid chat can use to write its own computer programs and run them, while staying transparent and safe for humans.

=== Dynamait - main features ===

The main features that differentiate it from a regular computer language are: The ability to use AI prompts and to target specific models and AI technologies, along with rule-based logic and rigid procedural code. The compiled natural-language package with examples, tests, and verifications. The use of dynamic interpretation, with evolving and dynamic "code" responding to real-time conditions and available resources.

=== Background: Humand.ai Taisker

Before I explain Dynamait, I have two background components that are needed to make the Dynamait framework useful. In the previous version, they were part of the Dynamait idea itself.

=== Humand.ai - the self-aware aligning AI chat ===

Humand.ai (the AI that teams up with you) is a self-aware chat that knows what the discussion is about and constantly aligns with you, gathering needs, leading discussions through stages and iterations, while consulting you. A prototype is in the making, and the code will soon be available on GitHub.

=== Taisker - for AI to manage its own complex tasks ===

Taisker is an AI task orchestrator for complex tasks. The task workflow is created by Taisker's task builder, initiated by Humand.ai, consulting with the user. Each workflow is a series of conditional actions using Dynamait-defined tools and relying on their results.

Dynamait can be bound to accountable, legally identified human authors and users, open to observer scrutiny, by legally identified accountable and authorized administrators.

It has two parts: constructing Dynamait tools and modules, and using them.

=== Summary === 1. You chat with the "AI" (actually with Humand.ai - the traced self-aware human-aligning chat) behind the AI.

2. Humand.ai detects intentions and topics being discussed, and sees the need for off-AI activity, usually due to a complex request, a failing response, or a task unfit for AI. It consults the human and aligns the goals, reporting the current problem and options for solution.

3. Humand.ai calls Taisker - The AI tasker, which finds similar "programs" or suggests a general workflow to achieve the goal. The flow uses Dynamait tools, composed of Dynamait components and modules, which use focused non-AI and AI components as needed.

4. If needed and if authorized or determined as safe, Taisker will oversee and manage the construction of the missing tools in the Dynamait framework, consulting the user along the way.

5. Dynamait uses two very different technologies bringing them together.

5.1 rule-based rigid logical steps with no generative networking AI, and in particular, with no inference language models.

5.2 AI-based inferred steps, including LLMs but mostly SLMs or NLM (nano), Knowledge graphs, RAG Graphs, and other types of AI knowledge bases.

5.3 Hybrid AI and logic, using them sparingly for what they are good at, when needed.

=== A bit more detail (mostly the original post) ===

We can ask the AI to do that for us, sometimes it will, sometimes not, but in any case, LLM based AIs can "create" a program in some programming language, and then brings up an environment where you can run it and use it. (The Canvas preview, for example, Claude and ChatGPT, or the vibe coding websites.)

With Dynamait you help the AI build its own tools so that it can get to your answer.

These tools can mix rule-based logical instructions with AI Generative-quoting statistical capabilities. You tell the AI what to do in detail. (loop till you went through all the data, sort the results by the soandso score, and then construct a new sentence out of the fourth letter of each word.) and it succeeds in doing it. - As opposed to AI failing while claiming it succeeded.

You tell it that you want it to create a Yamaha chord pattern from a song online, and then play it with the chords you set on the phone. For this it will tell you to go to a website like suno where you find out that this kind of function does not exist. Instead of giving up, you go off on a discussion slowly going into details of what is needed to make your program.

You then "program" it with Dynamait. This is a program that the AI runs, which gives the correct results. It is built with components, and modules, and parts. These are NOT prompts for AI. They are Tools with params that the "AI" can use.

For this to work, behind the AI there is a new type of chat: Humand.ai. It responds like an agentic chat, but under the hood, it is very different. It is acquiring the results with rule-based software, using AI components only where that is advantageous.

Humand.ai uses tools defined with the Dynamait framework, for tracking the discussion flow, analyzing the topics, issues, intents, requested tasks, and combined responses. They were mostly written by the hybrid human and AI development team, but some evolve dynamically with the discussions.

The workflow is managed by Taisker, the AI task orchestrator, a state machine that humandAI uses to execute and follow complex or critical tasjs, " Taisker takes the load off the big LLMs and instead, uses rule=based logic modules with no AI, only assisted by AI as needed.

=== From the original definition ===

What is different? Why is it not hallucinating, adding its own thoughts? And why is the usual AI path never able to follow instructions to the Tee?

Humand.ai is a program that seems to be an AI chat, but actually is not. You are running a set of modules in sequence or parallel. Each module has a workflow (similar to a program in a computer language).

The Taisker task orchestrator workflow is defined in a structured, configurable format, with parameters. It is similar to a computer program, only this is goal-based and written in natural language.

During your Dynamait-enabled chat, the need for this "tool" or "module" is discovered, or the option to use it is raised. It can be suggested to the user or invoked automatically, depending on the tool's settings and the chat's rules.

When the module is invoked, it loads two sections: the workflow and the params, some being dynamic prompts for AI, others parameters for the execution of rule-based algorithms and sequences of rule-based actions - the equivalent of function calls, event invocation, or object interactions.

A constructor object for each of these helps assemble the configuration lines, automatically or with human assistance, and sets them up accordingly.

An AI observer or a logical rule-based observer can be set to look at the current state of things (you tell it what that means), and before it does anything, it could stop and decide how to run, or even to change the "code" that it was set out to do. It could also stop and ask you for instructions (if that's what you want it to do).

The scope of the AI in these cases is very small. It does not see the whole picture. Sometimes, AI can be skipped altogether, and a state machine (with some kind of logic or other) can be used.

Each module of this novel programming language, which runs in the Humand.ai app or browser extension, has three parts that are loaded dynamically, as needed.

1. The Taisker workflow - the flowchart. But it's a dynamic flowchart that can change itself!!

2. Each Taisker task stage written with Dynamait - the function body, the instructions. Like functions, they have parameters "to look at" and act accordingly. Then they have an internal flow.

Dynamate has modules, components, instructions, intents, goals, roles, and results. It's like a prompt for AI but its a prompt for rule based behavior run by "logic machines" as defined in the program itself.

But these instructions, stored in the Taisker configuration files and the Taisker knowledge base, can be attained dynamically. Some are marked with an importance score, allowing the program to ignore missing info in some cases.

If the information is essential, it may suggest alternative sources or falling back to default values.

Finalizing the path: It can try itself out on different scenarios. You can supply it with multiple approaches, not staying locked into a single rigid procedure. In some cases, an instruction with only the general direction may sufface, leaving the implementation to Dynamait.

The Dynamait can create dynamic output types, some of which could be used in new ways.

3. Dynamic steps - almost every type of action can be achieved with alternative redundant methods, stacked against each other in parallel with verification methods, that give you better certainty and validity.

Did I just lose my chance to become a billionaire?