16•Article
Beyond the Hype: Understanding AI's Evolution, Applications, and Limitations
Adham OmranOctober 3, 202520 min read
Article Summary
From Turing and the AI winters to transformers and GPT-4, this piece traces AI’s real arc—then gets practical: accessibility tools, OCR and multimodal document understanding, rapid UI prototyping, and code docs. It also lays out the hard limits (context windows, stale training data, echo-chamber effects, hallucinations, and voice dilution). A clear-eyed guide to what AI can—and can’t—do right now.
Beyond the Hype: Understanding AI's Evolution, Applications, and Limitations
Adham Omran
Senior Backend Developer, iData
The field of Artificial Intelligence (AI) began in the early 1940s with works from McCulloch and Pitts’ design for "artificial neurons" and Turing's 1950 paper 'Computing Machinery and Intelligence', which introduced the now famous Turing test that our modern day AIs are being tested on.
Things progressed over the next two decades, in the 1950s and 1960s, when work on artificial intelligence began in laboratories, primarily within the field of computing. Early efforts aimed to create Artificial General Intelligence (AGI), but they underestimated the difficulty of this challenge. As a result, funding for the project was cut in the 1970s, ushering in what became known as the "AI Winter."
The AI Winter is a term used to describe the period of reduced interest and financial support that followed initial hype cycles that contained excitement and generous funding that were followed by disappointment when expectations were not met. This happened twice, between 1974–1980 and between 1987–2000. It is of particular interest that we are now in a massive hype cycle, the biggest in the history of AI. Does the past tell of the future? Or is this an eternal spring for AI?
Advancements of AI
To evaluate the effectiveness of artificial intelligence as a cognitive technology, one must present its historical development until the present day
- In 2010: Deep Learning techniques became feasible due to hardware improvements, data availability, and funding.
- In 2012: Image recognition, the technology now used in many gallery applications to distinguish pets, faces, and places, made its first breakthrough with AlexNet.
- In 2014: Facebook published their work on DeepFace, a technology for facial recognition that was able to identify faces with 97% accuracy.
- In 2016: Google's AlphaGo became the first AI to beat a human in the game Go.
- In 2017: Google's AlphaGo Zero became generalized to Chess. Also, the /transformer/ architecture was developed by Google and published in the seminal paper "Attention Is All You Need."
- In 2018: OpenAI introduces the first GPT (GPT-1) model, trained on 7000 books, based on the /transformer/ architecture.
- In 2019: OpenAI introduces GPT-2, trained on 8 million documents and 45 million web pages.
- In 2020: OpenAI introduces GPT-3, trained on the entirety of English Wiki, various web texts, and two book corpora.
- In 2022: OpenAI introduces GPT-3.5 with undisclosed training data. Moreover, the Stable Diffusion, a text-to-image model, was released.
- 2023: Meta releases LLaMa, the first GPT-3.5 competitor with open weights; OpenAI also introduces GPT-4 with undisclosed training data.
AI and its applications were specialized for a long time, but it was not until ChatGPT and co. that we got generic models that were trained on enough data to be useful. Previous models and tech were specialized to solve one problem and one problem only, but the current landscape has shifted to generic solutions that leverage speciality when needed.
Is Artificial intelligence limited to ChatGPT?
The recent hype has revolved around generative text and imagery with ChatGPT, but I invite the reader to contemplate other uses for AI that benefit us. Tools for accessibility have been on the rise that improve text-to-speech and speech-to-text interfaces, which help many people who are afflicted with impaired vision or speech. These advancements come from many companies, such as OpenAI's Whisper, Google's WaveNet, and ElevenLab's products.
Use Cases of AI
Let's dive into use cases for modern AI, starting from general use cases to more developer-specific use cases:
General
Large language models (LLMs) can be used to answer emails in corporate & professional environments, where it can be helpful to keep the tone of the conversation professional and reduce the time taken to perform this task. The only concern with this usage case is that it would diminish personal tone and individual expression as everyone would adopt a similar tone.
One fantastic use case is taking text and summarizing it or rephrasing it to fit a particular style of output. The LLM is unlikely to make mistakes here as it is confined and bounded by its input.
Moreover, Optical Character Recognition (OCR) is a technique used to parse text in images. It is widely used by everyone, from students who want to parse their handwriting into digital notes to people who require accessibility, such as those who read menu items when the restaurant does not have braille.
Multi-modal LLMs are capable of parsing text in a similar way. Still, the new advantage that they offer is the ability to interact with that document and understand it rather than plainly extract its text. This enables advanced use cases such as interpretation of tables which is difficult with regular OCR, explanation of diagrams for subjects that make heavy of diagraming, web development to convert a layout created in a diagraming application or even by hand into a working code prototype, imagine designing your website on a piece of paper, uploading an image of that to an LLM and requesting it gets converted into code. That is possible now!
B. Developer
Developers can bootstrap new projects fairly quickly if they are familiar with the language and the toolchain; they can develop new features for small projects or extend code where it fits in the context window. Developers would benefit from the ability to develop robust user interfaces quickly with tools like Vercel's v0, which creates a text-to-application pipeline. Of course, this interface lacks logic to drive it, but it is an interesting tool that can reduce the time consumed for projects.
AI can be utilized to generate code documentation when given accurate and clear instructions about the project’s code and coding style. Reducing time and human error. However, reviews are necessary to ensure the documentation of actual functionalities.
AI Uses Limitations
Context Window
The context window for an LLM is how much it can understand in a session. One may notice this after talking to ChatGPT or Claude for a while. They seem to forget information from the start of the session. This is the context window limit being hit. It stops us from dumping many PDF documents into an LLM and getting a summary. The largest context window was archived in Google's Gemini 1.5 in February 2024, which was up to 1 million tokens; this can represent 1 hour of video, 11 hours of audio, 30,000 lines of code, or 700,000 words. In contrast, other models only get up to 1/5th of this capacity.
This limit, even with the larger models, presents issues in software development; many code bases, especially in industry settings, span way beyond the context window of many of the current models, which makes them prone to failure when trying to extend that code with a new feature. This means we would not reach the level of automation to "feed everything to the machine and request new stuff to be added" anytime soon.
New Information
Resources that are newly published are beyond the scope of many models, including libraries (code made available by other developers to aid in development) in development, and have received significant updates. One example of this is the React web development library, which got a release in late 2024, making it unavailable to use for many models, such as Claude Sonnet 3.5.
Examples also include newly released research papers, books, academic articles, and advancements in various fields such as medical research and architecture. Attempting to circumvent this by feeding the LLM the material makes you run into the context window limits even sooner.
The Great Echo Chamber
One consequence of creating content with LLM and posting it to the internet is that the same crawlers used to initially train the model will now feed it back its output; this often creates degraded outputs and performance.
Accuracy and Quality of Information
One of the current problems facing LLMs is data to train on; grabbing all of the public internet and using it to feed the model might seem like an okay idea at first, but its consequences have been creeping up. The concept of garbage in, garbage out (GIGO) in computer science states that flawed, biased, or poor quality ("garbage") information or input produces a result or output of similar ("garbage") quality. When you use a lot of data without vetting, quality control, or auditing, it results in many incorrect pieces of information making their way into the system, not to mention problematic content by extremist forums, which can align the model in an undesired direction. One of the ways Llama (A model by Meta) gained traction is that it used a smaller subset of data that was vetted and audited to ensure the best quality of input; this resulted in better performance than its contemporaries at the time of its release.
Learning New Topics
LLM models are not efficient tools for learning a subject, especially if one has no prior background in that topic to evaluate the information. Many LLM models are subject to hallucinations. In addition, this kind of learning is not too dissimilar to being spoon-fed information in college. Clever prompting might get it to ‘explain’ things, but nothing guarantees those explanations are at any level accurate or correct.
Another aspect is that learning and development require difficulty; there is always a difficulty slope to climb. Otherwise, it means you already know what you are trying to learn.
Personal Expression
As alluded to previously, using LLMs can dilute your personal voice and unique writing style. It will condense your voice to what it thinks sounds average, diminishing your personal expression and creativity. In addition, with the popular use of LLMs for this particular case, we are bombarded with a great amount of content that lacks substance, value, and a personal touch. Depending on AI completely for writing tasks will eventually hinder one’s ability to write and think. LLMs can be great tools for making minor edits and tweaks without relying on them to produce entire pieces that sound similar to everyone else's, even if that is no longer your voice.
Conclusion
AI presents us with remarkable capabilities and concerning limitations. As we navigate this technological landscape, we must approach AI with informed optimism - leveraging its strengths while acknowledging its boundaries. The future of AI depends not just on technological advancement but on our thoughtful integration of these tools into society.
Article Information
Details
- Author:Adham Omran
- Published:October 3, 2025
- Issue:16
- Read Time:20 min
Need Custom Analysis?
Contact our research team to discuss your specific business landscape analysis needs.
Contact UsMore from 16
