What is Generative AI, and Why Should We Care?
tia-aremuTable of Contents
ChatGPT’s launch in November 2022 sparked global fascination with generative AI (GenAI). Millions of people were excited that they could engage with a bot to compose poems, emails, codes and more without any programming. According to Aaron Levie, this art of generating synthetic content was “a glimmer of how everything is going to be different going forward.”
Today, there are thousands of generative AI tools apart from ChatGPT. Some notable examples include ClaudeAI by Anthropic, Bard by Google, and Bing Chat by Microsoft. There’s a tool tailored for every need. Originally designed for generating text, these applications have evolved to cover the creation of sounds, images, and videos.
As genAI becomes widely accessible, prompting conversations on the rise of AGI, there is a need to understand what it means—its capabilities, limitations, and societal impact—as well as what you, especially, stand to benefit from each one of them.
Understanding Artificial Intelligence (AI)
If you’ve watched movies like ‘Her’ or read books like ‘I Robot’, you would recognize AI as a human-like technology that can think and act by itself. For instance, in ‘Her’, when Samantha composed piano pieces reflecting emotional experiences, that sophisticated computer program exhibited artificial intelligence.
Beyond sci-fi, you’ve probably interacted with AI using your phone to instruct Siri or Alexa. You tell these virtual assistants what tasks to perform and how to respond. According to Andrew Moore, the Dean of Computer Science at Carnegie Mellon University, AI is defined as “the science and engineering of making computers behave in ways that, until recently, we thought required human intelligence.”
Fun Fact: Alan Turing conducted the first substantial research work on artificial intelligence, which he called “Machine Intelligence.”
Classification of Artificial Intelligence (AI)
AI can be classified into two systems. One is capability-based, dividing AI into narrow, general and super.
- Narrow AI focuses on specific tasks (e.g self-driving cars).
- General AI manages more complex cognitive functions (e.g Siri handling speech and vision).
- Super AI surpasses human-level intelligence (theorized).
The other classification is functionality-based, which is divided into four classes.
- Reactive machines respond to the present based on data (e.g Facebook feed).
- Limited memory systems accumulate knowledge over time (e.g algorithmic ads).
- Theory of mind enables understanding different perspectives (e.g AI assistants).
- Self-awareness allows for conscious experience (theorized).
Introduction to Generative AI
6 in 10 US adults are familiar with ChatGPT, indicating that you’ve likely tried instructing it to create something. Whether you’re asking it to write an article or a story, this process of artificially generating original content is generative AI.
Generative AI traces its roots back to chatbots from the 1960s, which were limited to responding to prompts and couldn’t create content. Recent advancements, like generative adversarial networks (GANs) involve algorithms competing against each other to synthesize realistic data.
Built on transformers and large language models that scan billions of parameters, modern systems such as DALL-E and ChatGPT demonstrate remarkable progress in generative creation.
Fun Fact: The first generative AI tool was Eliza, a chatbot created by Joseph Weizenbaum in the 1960s.
How Does Generative AI Work?
Generative AI works in five ways and can build one yourself. Suppose you and I are avid book readers and want to create a generative AI tool for writing novels, here’s how we’ll do it:
- First, we build a model and then gather lots of books to train it. Let’s call our model BookGPT.
- Then, we design two contrasting networks: a generator and a discriminator.
- The generator predicts fictional books based on real examples, and the discriminator decides if it’s real or fake. Creating a GAN.
- Next, we train the generator to deceive the discriminator with fake book content while the latter works hard to spot the deceit. Over multiple rounds, the discriminator evolves to better spot fake content.
- Lastly, we assess our GANs performance (the back-and-forth between generator and discriminator). If the discriminator creates plausible novels even as our generator tries to create an adversary, our model, BookGPT, is successful.
That’s basically how to build generative AI. Now that we have a working model, we can generate novels continuously by giving it prompts. The only thing left is constant supervision.
Still not sure how generative AI works? Watch this video:
Pro Tip: OpenAI (ChatGPT) allows you to build your custom GPT and get paid for it. Log into their website. Look for the “Create” tab to start building. To train it, look for “Configure” section.
Types of Generative AI
Generative AI models can be classified into three categories: input, output, and multimodal (a combination of input and output).
Input-Based Generative AI
Starts with text prompts to produce text, images, audio, video or code.
- Text-to-text systems such as GPT-2 and our BookGPT generate written responses from textual inputs.
- Text-to-image models, exemplified such as DALL-E create original digital art.
- Text-to-audio tools like AudioLM provide customizable voices and music.
Output-Based Generative AI
Synthesizes content from images, voices or sensor data inputs.
- Image-to-text models generate written descriptions from visual inputs.
- Audio-to-text transcribes speech to text.
Multimodal Generative AI
It combines inputs like images, text, and metadata to boost the quality, relevance, and diversity of generated outputs. As inputs and outputs continue to expand, so will generative AI’s creative production abilities.
Prompting Generative AI
Prompting is when you type a question into a GenAI tool like ChatGPT to generate a response. Prompts leverage the model’s capabilities, whether asking it to pen a poem (creative), explain physics theories (informational), solve maths problems (instructional) or discuss favorite bands (conversational). Basically, a prompt presents a query, the AI processes it, and then answers with synthetic content.
AI prompt engineering is crucial for quality results. Poor prompts produce nonsensical outputs. Effective prompts:
- Clearly set the context and necessary background
- Provide precise instructions
- Include examples of ideal inputs and desired outputs
- Anticipate and address potential ambiguity upfront
- Use respectful, inclusive wording
For example, vague prompts like “I’m hungry” as seen in the image below, pose a risk of yielding poor results. Thoughtful prompting enables generative AI to responsibly create content.
Pro Tip: You can get free prompt templates from around the web and on social media. Additionally, you can prompt ChatGPT to generate a series of effective prompts for your specific need.
Importance of Generative AI
Generative AI holds promise. According to a McKinsey report, AI has the potential to contribute up to $13 trillion to the global economy by 2030. With $2.6 to $4.4 trillion stemming from generative AI models like DALL-E 2 and GPT-3.
By automating and augmenting a wide range of tasks, GenAI can do it all. GPT-3 swiftly generates accurate first drafts of content within seconds, saving hours of human effort, a corporate report shows 25% of companies using ChatGPT have saved over $75,000. Meanwhile, DALL-E 2 excels in creating product images and marketing visuals, achieving this at a fraction of the typical creative costs.
“These tools and technologies we have developed are the first few drops of water in the ocean of what AI can do,” remarked Fei-Fei Li, co-director of Stanford’s Human-Centered AI Institute. However, they must be “human-centered to benefit people in positive and benevolent ways.”
Why should you care? Humanistic AI
Outputs from DALL-E 2 and GPT-3 showcase intelligence and creativity. Some experts believe advanced models could potentially achieve artificial general intelligence (AGI), essentially, human-level cognitive abilities.
According to Elon Musk: “The pace of progress in artificial intelligence is swift, growing at a pace close to exponential” . Unless we all care, “the risk of something seriously dangerous happening is in the five-year timeframe, ten years at most.”
Without control and governance, we risk being the fuel for generative AI rather than its beneficiaries. As these models become more capable, they demand ever-expanding datasets to improve performance. If proper safeguards aren’t in place, the personal data and digital traces we generate could be used to train generative systems without permission, oversight, or shared benefits.
Fun Fact: Humanistic AI prioritizes human values, needs, and ethical considerations when developing AI. Tom Gruber, the co-creator of Siri, INTRODUCED IT.
Impact of GenAI on Society
Generative AI has numerous benefits and is gradually reshaping society. However, concerns linger around legal, environmental, and information issues, as well as “fake” generated content. Furthermore, these advanced systems could disrupt industries and eliminate jobs. The IMF has issued a warning, projecting that AI could replace 40% of jobs within the next 15 years.
“As with any transformative technology, generative AI comes with promise and peril,” said economist Erik Brynjolfsson. “It calls on us to shape policies and governance to amplify the benefits while minimizing unwanted impacts.”
GenAI Ethical Considerations
Critics emphasize that generative models can inherit biases from training data. Images of minorities and women generated by early systems sometimes contained hurtful stereotypes.
Models also raise concerns around creative copyright and attribution. Who owns the output: the human user, the algorithm creator, or both?
Fun Fact: UK Prime Minister, Rishi Sunak, held the AI Safety Summit to address issues in GenAI. It took place at Bletchley Park in November 2023.
Generative AI Requires a Human-Centered Approach
As generative AI continues to advance, it presents both promise and risk. Models like DALL-E 2 and GPT-3 showcase the potential to enhance countless industries through automated content creation. But biased or misleading outputs could also negatively impact society if development outpaces ethical safeguards.
Ultimately, a balanced, human-centered approach is necessary to ensure that these tools benefit people in broadly positive ways rather than causing harm. Responsible governance and AI safety research will play a crucial role in realizing the full potential of this generative technology.
References
Buttazzo G. (2023). Rise of artificial general intelligence: risks and opportunities. Frontiers in artificial intelligence, 6, 1226990. https://doi.org/10.3389/frai.2023.1226990
Google for Developers. (n.d.). Overview of GAN structure. https://developers.google.com/machine-learning/gan/gan_structure
McKinsey Global Institute. (2018, September). Notes from the AI Frontier: Modeling the Impact of AI on the World Economy. https://www.mckinsey.com/~/media/McKinsey/Featured%20Insights/Artificial%20Intelligence/Notes%20from%20the%20frontier%20Modeling%20the%20impact%20of%20AI%20on%20the%20world%20economy/MGI-Notes-from-the-AI-frontier-Modeling-the-impact-of-AI-on-the-world-economy-September-2018.ashx
Resume Builder. (n. d.). 1 in 4 companies have already replaced workers with ChatGPT. https://www.resumebuilder.com/1-in-4-companies-have-already-replaced-workers-with-chatgpt/
Interesting Engineering. (2024, January 15). International Monetary Fund warns, AI to threaten 40% of global jobs. https://interestingengineering.com/culture/ai-to-threaten-40-of-global-jobs
