LLM, AI agents and services

The release of ChatGPT to the public in November 2022 sparked a wave of hype around AI technologies that shows no sign of slowing. The unveiling of new AI products is frequently juxtaposed with anxious discussions from tech influencers and thought leaders about the possibility of human enslavement. However, listening to some so-called "experts," one gets the impression that they don't fully understand what they're talking about.?

This article initially was a collection of information about technologies, products, and careers connected to AI. While the content might seem obvious to some, I hope it will be helpful for those looking to get a handle on this new and fast-changing field.

Types of Artificial Intelligence?

To understand AI, we must first examine its different classifications.

Artificial Narrow Intelligence - ANI:?This type of AI is characterized by its focus on performing a single, specific task. While they can achieve impressive results within their defined areas, these systems lack the general intelligence and the adaptability to apply knowledge gained in one area to another. Examples of this "narrow" AI include everyday tools like ChatGPT, Gemini, and virtual assistants such as Alexa.

Artificial General Intelligence - AGI:?AGI refers to systems that possess intelligence and cognitive abilities that are comparable to, or approaching those of, humans. Although no pure examples of AGI exist today, the creation of AGI is a major focus for research and development by leading organizations. The primary question is not whether AGI is possible, but rather when it will be achieved.

Artificial Super Intelligence - ASI:?This level of AI represents a theoretical future in which artificial intelligence surpasses humans in all aspects, including creativity. ASI, at this stage, is still a purely hypothetical concept often found in science fiction.

From the preceding points, it becomes clear that the more alarming scenarios described by figures such as Yuval Noah Harari are primarily focused on the potential dangers of AGI and ASI. A different, more grounded discussion, as presented in many articles and analyses, explores how many and what types of jobs may be displaced once general AI moves from theory to practical deployment. The debate around DeepSeek and its impact on IT markets, in contrast, is rooted in the competitive landscape of "weak" AI solutions.?

Indeed, “weak” AI is essentially a tool for significantly boosting individual productivity. Consider, for instance, the process of translation. Years ago, an experienced translator would need to dedicate hours to translating from one language to another. With the adoption of services like Google Translate, a classic example of "weak" AI, this process has become dramatically faster, even though the results still require human oversight. Currently, the output from Google Translate can be enhanced through Google AI Studio, providing users with multiple translation options. While some refinement is still required, the time needed to complete the task has been dramatically reduced from hours to mere minutes thanks to "weak" AI.

Thus, “weak” AI is an empowering assistant, enhancing human productivity and efficiency without completely replacing the human. In this evolving partnership between humans and AI, the human, and particularly the human expert, maintains the leading position, directing the AI for specific purposes and critically examining its work.

"Strong" AI, however, is anticipated to entirely displace many professions. The timing for this transition is uncertain, but the expectation is that it will occur over the next several years.

We now turn our attention to examining specific practical examples of "weak" AI.

LLM, AI agents and AI-based services

When you visit platforms like Google AI Studio, ChatGPT, or DeepSeek, you're engaging with a service built upon Large Language Models (LLMs) – or, more precisely, with a choice of several LLMs.

A Large Language Model (LLM) is essentially a neural network that has been trained using vast quantities of data. While the exact number of operational LLMs remains unclear, it likely numbers in the thousands. We have seen the emergence of LLM "families" developed by major tech companies, for instance, ChatGPT by OpenAI, Gemini by Google, and LLaMA by Meta. Alongside these, there are also specialized models from various companies and research institutions.

This naturally leads to the question: which LLM is best suited for your specific task? LLMs vary significantly in their underlying architecture, their network depth, the number of parameters they employ, their training methodologies, and a multitude of other factors. Specialists in fields such as machine learning (ML Engineers), natural language processing (NLP Engineers), and AI research (AI Researchers) can provide expert guidance in this area. However, you can also assess various models on your own by evaluating their responses to your questions. In my personal experience, for general information queries, I tend to rely on Gemini 2.0, as its responses are consistently detailed and insightful.

However, it is crucial to acknowledge the inherent limitations of LLMs. One of their key constraints is that they are trained on extremely large datasets, but these datasets are nonetheless finite. While LLMs can successfully handle queries within the scope of general knowledge, they often struggle to provide accurate answers to highly specialized or niche inquiries.

To mitigate the limitations of Large Language Models, consider the following approaches:

1. Fine-tuning:?This process involves curating a relevant dataset for a specific area of interest and transforming it into a suitable format (e.g., text files, CSV). Then, employing tools such as Hugging Face Transformers and PyTorch, you can integrate this dataset into a chosen LLM, effectively modifying its parameters. The outcome is a customized model that can answer your niche-specific questions with increased precision. However, fine-tuning demands considerable computational power. For those using medium-sized models (e.g., LLaMA 7b), a PC equipped with a high-performance graphics card like the RTX 4070 (or better) is crucial, since the processing primarily takes place on the GPU, rather than the CPU. It's also worth noting that fine-tuning can lead to a model "forgetting" some general information it had previously learned.
2. Retrieval-Augmented Generation (RAG):?This technique also leverages a pre-indexed dataset. RAG extracts relevant pieces of information and provides these contextual fragments, alongside the initial query, to the LLM. For working with the LLM+RAG configuration, you do not necessarily need to possess programming skills. You can utilize ready-made solutions like Verba or the Poe service. Furthermore, these tools frequently offer the ability to select from a variety of available LLMs.
3. Prompt Engineering:?This approach sidesteps the need for programming but instead focuses on the art of constructing well-crafted prompts and requests, providing the model with detailed instructions.

Clearly, LLMs require a certain level of expertise to leverage effectively. The positive takeaway is that, for now, human intervention is essential, particularly from experts in relevant subject areas. At a minimum, someone must prepare the data, assess the reliability of the responses, and skillfully engage with the “machine.”

Shifting our focus, let's explore AI agents. AI agents are systems where an LLM operates as the core engine, or as an element that shapes process logic or action sequences. These agents perform specified tasks, such as researching topics online and crafting articles, or creating summaries and then distributing them via platforms such as Telegram or email. A wide range of tools for creating AI agents are already available, empowering users to build agents that suit their specific functional needs. Frameworks such as n8n, Abacus.ai, Bootpress, Rasa, and LangGraph offer robust visual interfaces to create agent workflow logic. While these are not out-of-the-box solutions and may come with associated costs, they provide a high level of control and customization.

In contrast, services represent solutions that require minimal programming skills and avoid detailed logic design. These tend to be packaged with simple, intuitive interfaces, allowing you to create, say, an image generator in just a few clicks.

In conclusion, while AI will not displace humans in the near future, it will fundamentally transform how we tackle various challenges. Just as the introduction of CAD did not lead to the disappearance of architects and designers—they simply incorporated new tools into their workflow—subject matter knowledge will continue to be critical. While digital photography might have rendered the traditional roles of developers and photo enlargers obsolete, the artistic ability to capture the right shot remains highly sought-after. And while translators may encounter challenging times, human language expertise remains necessary to ensure the accuracy of the final "machine translation."

The past forty years have seen countless examples of how technology has transformed professions. The key takeaway is that as long as AI remains in a "weak" and "narrow" state, the key success factor will be the human being who can effectively interact with it.