✍️ ChatGPT & Co: Text and learn
The landscape of text generation tools has witnessed a remarkable surge, but unfortunately, many of them appear strikingly similar without offering substantial innovation...
Attack of the clones
Every day, I keenly observe the proliferation of AI text generation tools, including AI conversational agents, with a skeptical eye. In fact, the sheer abundance of these tools has led me to abandon my reliance on Excel and instead turn to sites like Futurepedia, Future Tools or Ai Tool Hunt to keep track.
As I reflect on the matter, I find myself contemplating the sustainability of these tools. After all, when it comes to general-purpose solutions, is there anything that can truly surpass ChatGPT, especially now that it is enhanced by its extensive ecosystem of plugins? Moreover, it's worth noting that our everyday tools, such as Office Suite and Google Workplace, are embracing the integration of the GPT language model or comparable alternatives by default."ChatGPT is the new 'Intel inside'," proclaimed FastCompany, a telling statement, isn't it?
Certainly, a second group of tools is emerging, providing features that cater to more specific needs. Their advantage? These specialized models incorporate specific knowledge, terminologies, and rules relevant to each domain, enabling them to produce more accurate and relevant outcomes.
Let's consider an example: Jasper leverages the power of GPT-4 and fine-tunes its model specifically for marketing purposes. The key advantage lies in its user-friendly interface, equipped with pre-defined templates for various content types like social ads, blog articles, product descriptions, and more. This feature proves invaluable for individuals who may not possess advanced prompting skills. However, let's play devil's advocate: are we not at risk of being inundated with templated content, leading to a homogeneity that lacks quality in the end? Moreover, should "prompting" not become an essential skill? It seems evident that by mastering the art of strategic prompting using versatile tools, we preserve greater flexibility and the capacity to create content that breaks away from standardization.
Fine-tuning for distinctive data
However, fine-tuning applied to proprietary and private data may potentially make a difference. It is in this regard that I firmly believe in a parallel path that is likely to gain importance in the coming months and years: the B2B sector with AI tailored to the needs and data of businesses. It is from this perspective that I consider LLaMA, ChatGPT's rival backed by Meta, the parent company of Facebook, as a game changer.
Allow me to elaborate: LLaMA offers a unique approach as a non-commercial open-source model. It is a smart choice by Meta, as it promotes adoption and collaboration around LLaMA while improving its tarnished image due to numerous privacy scandals. However, this endeavor is not entirely altruistic. By making it open source, Meta benefits from collective intelligence that can contribute to enhancing the model's performance.
In this case, businesses can now access this model and use it as a foundation for fine-tuning, enabling them to create AI models specific to their needs, industry, and private data. The motto is: sin-gu-la-ri-ty!
Furthermore, LLaMA offers a key advantage in terms of privacy. Unlike many existing solutions, its open-source approach allows businesses to maintain tighter control over their sensitive data. Traditionally, when companies use AI models provided by external vendors, their data may be processed outside of their secure environment, raising concerns about data confidentiality and protection. With LLaMA, companies have the option to deploy the model locally, on their own servers or systems, reducing the sharing of their data with third parties.
Source-grounded AI
In recent years, we have witnessed a frantic race to increase the size of language models in order to enhance their performance and text generation capabilities. While large-scale language models (LLMs) certainly have their usefulness, other approaches are emerging, such as source-grounded AI.
The concept: Instead of solely focusing on the size of the model and the amount of training data, this approach aims to give users the ability to define a set of trusted sources to guide the interactions of the AI. This is precisely what the Tailwind project explores, as showcased at the Google I/O 2023 conference, by allowing users to define a set of source documents that serve as a "fundamental truth" or foundation to shape the model's interactions. The advantage: By giving users control over the data sources used by the AI, it strengthens trust and transparency in artificial intelligence systems. Additionally, this approach addresses certain pain points associated with large-scale models, such as the high cost of utilizing cloud services or privacy concerns related to data transmission and analysis in the cloud. Moreover, by running locally on devices, these models offer increased speed and responsiveness.
Artificial Creativity
Beyond these fine-tuning or source-grounded AI models, there are also outsiders aiming to disrupt the existing paradigm by pushing the boundaries of machine creativity and exploring new perspectives in text generation. Unlike traditional approaches focused on perfection and precision, these models embrace errors, hallucinations, and inaccuracies as catalysts for creativity.
As a reminder, hallucinations in the field of AI refer to the generation of results that may seem plausible but are either factually incorrect or unrelated to the given context. These outputs are often the result of inherent biases in the AI model, a lack of understanding of the real world, or limitations in the training data.
By experimenting with selective and intentionally flawed mechanisms, these models can produce results that are as surprising as they are inspiring, opening up new possibilities for innovation and artistic expression.
DreamGPT, for example, explores this approach by favoring the generation of content that may seem absurd or incoherent at first glance. However, by pushing the boundaries of logical reasoning and embracing ambiguity, it opens up new avenues in content generation.
Let's illustrate this with an example. Take the famous phrase by Descartes, "I think, therefore I am." If I were to ask ChatGPT to complete “I think, therefore…”, any response other than "I am" would be deemed incorrect. However, it is precisely the playfulness and wordplay surrounding this quotation that become intriguing in an artistic context, sometimes even forming an author's unique style or signature. In a previous article, instead of using the title "Shopping According to Your Values" to discuss the growing trend of websites that enable shopping aligned with personal beliefs (local, ethical, etc.), I chose the title "I shop, therefore I am." While some might argue that by adequately briefing ChatGPT, we could achieve the desired result, it would require constant intervention and guidance from the user, thus limiting the freedom and spontaneity of creativity.
These outsider models represent an intriguing evolution in the field of AI as they challenge established norms and explore unconventional approaches to artificial creativity. While enterprise tools continue to focus on efficiency and productivity, these approaches pave the way for a deeper exploration of artistic expression, and divergent thinking. They also raise fascinating questions about the boundaries between human and machine. However, they do generate some concerns. In December last year, I highlighted the significance of embracing our imperfections and fostering contrarian skills in a world shaped by AI. But if AI starts harnessing strength from its imperfections, where does that leave us?
MD