Just Ask, Moses!

The AI landscape today (November 2024) is heavily driven by large corporations that subsidize/fund the compute and R&D costs of small companies that have worked for years to train their LLMs.

Therefore, it is hard to miss that these large corporations and funds are looking to recoup their investments by selling the final product (like MS’ Copilot) to large and medium enterprises that are already making good money. Why? Because they can make even more money by cutting costs: streamlining their processes, shortening their internal procedures, and getting rid of employees who were hired merely to check a box, among many other reasons.

There is no doubt that AI will generate revenue for enterprises, sooner rather than later. Thus, given the clear objective for AI development, it is safe to say that the current AI trajectory is driven heavily by the objectives of large corporations, which primarily focus on making or saving money.

Why would this be a problem?

It is not a problem for these enterprises that benefit from the existing route AI is taking. On the contrary, they call it a revolution in technology that occurs only once every hundred years. However, they are cracking walnuts with their laptops.

How?

To put it simply, an increasing number of small professionals are securing high positions in large corporations, and they establish their baseline by thinking solely in financial terms. While not inherently negative—after all, that is the heart of capitalism (with exceptions for small individuals)—the missing, bigger opportunities lie in different kinds of questions: not How can we make more money? but rather How can we use this technology to broaden our humanity? No capitalist has ever asked this (and no socialist either). Carl Sagan did, though.

So, today’s AI is justifiably focused on making money. While this focus is inevitable and not an issue in itself, the problem arises from the fact that this is exclusively what drives AI, turning a great potential into a tool that serves corporate interests instead of an opportunity for humanity as a whole. I believe the AI market is large enough to accommodate not only corporate interests but also those of individuals (aka consumers). A good example is Kagi, a search engine that relies solely on subscription revenue and has become self-sustaining about eight months ago, without external funding—just consumer money keeping it alive.

Existing model

The current AI model is quite simple: you ask questions, and it delivers answers—straight and clear. In theory, the better the AI, the better the answers. The limitation of this process (besides its accuracy, of course) is that it is linear, transactional, and it shouldn’t be. The AI goes straight from your questions to categorizing them lexically and morphologically, establishing the scope, and finding the proper tokens to achieve that scope. This forms a linear pathway. The drawback of this line is that AI perceives any deviation from it as a separate line. The lack of memory (historical context) has exacerbated this issue, causing disruptive and disconnected replies.

On the positive side, both ChatGPT and Dot can now retain much of the discussion history, allowing them to take that into account when responding. The fundamental issue remains: there is still a linear trajectory between billions of documents ingested by these LLMs—straight to making money—by streamlining processes, decluttering procedures, and eliminating non-creative people, as Jensen Huang has stated. This isn’t the way for people to benefit from AI; it only benefits corporations.

Proper model

Consider how people communicate and share information. Conversations are rarely straightforward; people often mishear things, misinterpret others' statements, seek clarifications, adjust their tone of voice, and modify their objectives based on the other person's responses.

These intricate processes are notably absent from AI discussions, but we now understand why. A better design for an AI assistant, one that should genuinely assist human beings rather than corporate protocols, should involve a constant practice of asking for clarification. Whenever an unclear term arises, or a contradiction or void in an argument is detected in users' queries to AI, the AI should pause and request clarification. These queries should not follow an algorithmic pattern, which can be tedious, but rather be triggered dynamically based on the AI's analysis and interpretation of human input.

The first condition is for the assistant to retain some form of memory; the second is for it to cross-reference any new message with this history. Each phrase must be evaluated from not only the AI's actionable perspective but also a clarity standpoint. If the meaning is unclear or contradicts previous interactions, or is vague, the assistant must ask for further clarification.

I cannot stress enough that these clarifying questions are crucial for better understanding an user's intent and also for fostering a deeper engagement between human and machine, creating a middle ground for what I term decent dialogue. There can be no proper dialogue between two systems without meaningful feedback. Current AI assistants lack this feature completely, attempting to simulate it with algorithmic questions, which often frustrates the user. For example:

If the user states, Today is Tuesday; therefore, it’s autumn, the subsequent AI question should be, What is the connection between Tuesday and autumn?, instead of How do you feel then? (This exaggeration aims for clarity).

Fractal communication

In case I haven't emphasized this enough, the process of stopping to ask for clarifications (much like Moses should have stopped to ask for directions in the desert…) is the foundation for a fractal design in AI-to-human interactions.

Retain your algorithms (if they serve your purpose), but drop down a semantic layer whenever the situation demands it and repeat the algorithm. Elevate the semantic level back to baseline or beyond once you've clarified the uncertainty, and be ready to elevate it again whenever the human interlocutor seems satisfied with your responses.

For example: Human: I believe today is red. AI: What do you mean by today is red? (Down-semantic) Human: I mean the sky is red today (straighten the semantics). AI: What is the significance of the red sky for you? (Up-semantic, compared to baseline).

People would not feel annoyed by these questions; more often they enjoy explaining themselves, especially to machines, which they generally perceive as somewhat dumb. In contrast, people are irritated by redundant or pointless follow-up questions, feeling as though they are being forced to continue a conversation for no meaningful reason.

I call it fractal communication because this is how our communication operates at its core, and it should be readily replicated in AI interactions using a fractal model (by detailing specifics) instead of a linear one, which aims for swift actions. You can dive as deep as required, while still maintaining the algorithm, and rise as high as the vocabulary allows, keeping the same algorithm.

Ultimately, the specific algorithm matters less than the AI's capability to engage deeply and broadly during conversation with a human. This is not a change in technology; it's a shift in how we perceive the goals of that technology—focused on achieving proper human-to-AI interaction.

Consequences

First, there’s the benefit of feeling heard and understood when interacting with an AI, which brings users closer to technology. These clarifications allow for better and deeper (fractal) understanding. This process creates a virtuous cycle of grasping human intentions and meanings that cannot be achieved quicker or more effectively in any other way.

Without this approach, the dialogue between humans and machines becomes just machine-to-machine communication. People must simplify their meanings and strip their language of nuance, resorting to machine talk, which leads to frustration and dehumanization. Neglecting to foster a fractal dimension in dialogue risks turning this technological revolution of AI into a regression of the human speech.

What is required

When designing AI assistants, it is essential to include, alongside technical and management teams, artists and philosophers. An artist can identify when the language has broken down and suggest how to rectify it, while a philosopher can help determine how to meaningfully connect with human emotions. Both can bring insights into nuances and meanings that engineers and managers may overlook due to their formal training.

Many AI projects are engineered by technical teams, which is good, we love engineers. However, these initiatives often miss some important elements of arts, style, meanings, and life touches. Consider the process of building cars: no matter how skilled the engineers, a designer is always needed. This principle holds true for many products that interact with or are purchased by people: think about houses, gardens, devices, furniture, etc.

Key point of sales

Presently, AI assistants are primarily purchased by early adopters (whether corporate entities or consumers) who are accustomed to compromising to benefit from new technologies.

To engage the average person, AI must present itself as a relevant, competent partner in dialogue. Today, there is no genuine dialogue occurring between AI and humans; it resembles a search legacy devoid of meaning, depth, or warmth.

Consequently, current AI solutions cater solely to corporate needs, not to the dialogues of everyday people. Yet, even corporations are composed of individuals. If a proper dialogue is not established with white-collar workers, and AI becomes a mere tool imposed upon them—sometimes even replacing them entirely—these individuals will never engage with AI products in their personal lives.