Generative AI is a type of artificial intelligence (AI), one of many, where it is trained on a very large set of data. After training, if you give it some direction, it generates something for you. It can generate an answer, text, a picture, or it might generate code. It generates things based on your query and what it’s learned previously. This is different than many other forms of AI which for example might look for patterns in large data sets. In order for generative AI to work well, it has to be trained on incredibly large sets of data, such as large language models (LLM).
Not all AI generates something. For instance, think of AI that is searching for drugs that might work against a particular kind of tumor in cancer. What it’s doing is looking for a particular pattern, or for something across a huge swath of DNA and a huge swath of information to say, “Here are eight possible candidates.” It didn’t generate a drug, it didn’t generate anything. So that’s been a common use of AI for the past decade or more. It’s often looking for patterns that match, patterns that don’t match, things like that. Generative AI, which is garnering a lot of public attention lately because of the emergence of ChatGPT (a chatbot) and DALL-E (a deep learning model), isn’t looking for a pattern at all. It’s actually generating something for you, whether it’s a document, or an image, or code, or even jokes.
AI is in the eye of the beholder
The term “AI” is to some extent a marketing term. If it seems incredibly intelligent and incredibly helpful, whatever math was used is arguably immaterial. If you tell an AI it’s an emergency and it gives you a step-by-step solution in five seconds to save a life, you’re going to think it’s wickedly intelligent. At that level, this is AI to you, regardless of what algorithms it used.
Purists have a somewhat different view. Their position is that machine learning (ML) algorithms have been around since about the 1950s. ML is just math that can learn and then accumulate more information that it builds on, and then learns a little more. That information may help it be more right or more wrong, but it’s building up the information about how to respond. AIQ uses many new algorithms as well as some that are from as far back as the 1960s, that were almost unusable then, but because of the cloud now, they’re quite impactful. It’s not useful if it takes a year for some ML to learn your application, and then it takes another year to learn the next build. But if it can be done in five seconds, that’s really interesting. You don’t care if it uses a convolutional neural net (CNN) or it uses an LLM based on a CNN, or it uses one of many other available algorithms. If it seems intelligent to you, it’s providing intelligence for your purpose, that’s AI.
LLMs break AI use cases wide open
LLMs came out of a translational model system that originated at Google about seven years ago to handle the complexities of language translation in a new way. For a long time, automatic translators just translated word for word. For example, a user would think, ”Oh, I know what this word is in French, and I’ll just map it. I know that word too, so I’ll map it.” But it turns out sentence structure is important. In English, the rules are loose about word order, but in other languages, the sentence structure is strict, and it’s the opposite of what we say. Google had a really good idea to actually go learn, essentially supervised learning, from actual real human-translated sentences. So, by itself, it learned that sentence structure is different. Instead of learning word by word, since it was the meaning of the sentence that mattered, it learned to construct a sentence that meant the same in another language. Supervised learning means that it has a label created by a human, for example a cat also labeled as a cat. In time during training a model reads thousands upon thousands (or millions) of pages to train itself and learn the CONTEXT of sentences across languages.
That led to thinking about LLM’s training with unsupervised learning. GPT-4 has read literally a hundred trillion phrases on the web. That’s unsupervised. It doesn’t know what’s right and what’s wrong, what’s true and what isn’t true. It just ingested trillions of words and sentences and context. However, it can weight certain things more than others.
Supervised learning might be, “I’m going to give you the answer to everything, and then you can create a function that gets me from the question to the answer.” Unsupervised is just go learn everything there is but, “I’m not going to be able to label the right or wrong answers.”
The Promise of Generative AI in application testing
It’s an exciting time. There are myriad opportunities and ways to use AI to grow your business, amplify your brand, and improve QA productivity, that includes being able to use generative AI in application quality assurance (QA) testing today. Appvance invented the first generative AI for QA in 2017. Appvance’s patented AI technologies automatically generate scripts using a built-in knowledge base of how UX libraries, how applications in general work, and by learning your application, what’s important to you in your application, and what’s important to your application’s users. With built-in models, it can deeply learn your application guided by QA professionals and a variety of data on user flows. Its generative AI system can generate hundreds, thousands, even tens of thousands of scripts all by itself… and run them by itself. And I mean no scripting, no recording, no maintenance. It simply generates scripts (code) that runs real user flows and validates outcomes. Often 10X the coverage of “test coverage” alone (we say full application coverage).
Appvance is hard at work expanding its 6-year lead in generative AI for QA. Think about what additional generative AI, now with the power of GPT-4, (in other words 100 trillion phrases read), is going to be able to add to software testing QA. In the coming months and years, leveraging this immense database that cost a billion dollars to build. No one has had the money to do this before, but now it can be leveraged to make testing much closer to true natural language, and understanding of applications much closer to the way humans work. This is another level of transformational technology, not seen before, that’s going to change the way that testers, QA people, and developers interact with, achieve more visibility, and address the quality issues in their applications. It is going to add additional capabilities to propel QA firmly from Agile to DevOps, increase visibility, reduce risk, and be 800X+ more productive from a quality standpoint (80 x 10). And because generative AI delivers higher quality, more intuitive applications, your positive brand equity (Link to my first blog) is also secured.
Want to embrace generative AI today in QA? Apply for a demo of AIQ and see for yourself what autonomous script generation can do for your visibility to bugs before release.