AI is getting creative and feels more human

Artificial Intelligence (AI) is no longer the shiny new tech. The tech stream is past its sloshy beginnings and is arduously winding its way through complexities. Investors have become picky about funding AI startups. Despite the tightening purse strings, AI had its moments in 2022. It was a year of small but noteworthy breakthroughs. AI entered the phase in a technology growth trajectory when the focus is on performance and accuracy. The technology is simply trying to get better at what it started out to do.

In 2022, AI got better at generating original text, summarizing a lengthy piece of writing, giving more coherent answers. OpenAI’s ChatGPT became rather eloquent at answering questions, having trained on large transformer-based language models that specialize in conversations. In essence, AI made itself useful for specific outputs like college essays.

The breakthrough technology in 2022 was Generative AI, producing impressive text, images and videos based on text prompts. OpenAI launched DALL-E 2, the most prominent among the text-to-art image generators. It connects written text to concepts, to produce images that are more realistic than their predecessors. Google and Meta gave us models that generate videos from text. Startups like Midjourney and NightCafe began challenging bigger rivals with their text-to-art tool. Stock-photo service providers incorporated image-generation models.

The fast-evolving diffusion model became the poster child for Generative AI, eclipsing the older generative adversarial networks (GANs). The model is based on thermodynamic principles. It progressively adds random noise to training data and then learns to reverse the process to remove the noise and rebuild clean data. New research improved the ability to predict what comes next in computer vision to rapidly generate videos after being trained on a video dataset.

At the same time, the open-source movement in AI by platforms like Hugging Face stimulated innovation. Stable AI’s community-based opensource model, Stability Diffusion allowed startups to build Generative AI tools for new kinds of content and art.

All this will broaden the scope of the Creator Economy in 2023. Generative AI has certainly lowered the barriers of our imagination. We can, in minutes, give shape to any concept our brain conjures. Generative AI startups are providing tools that will improve storytelling and brainstorming, write software code, find datapoints, write copy for ads, produce content for legal services, create new music and videos. New tools will make the best of our emails and presentation decks seem old and boring.

Yet there are limits to this rather seductive vision of AI. In 2022, AI felt more human, with flaws and all. While large language models improved, they got confident, even when they were wrong. Meta’s Galactica, showed a high level of confidence, even for the very obvious false information. It is bad enough to be wrong. But being wrong and confident is a bigger problem and frankly annoying. Despite an extensive application of reinforcement learning against biases, it showed basic ones like preferring white male scientists.

We are now faced with the prospects of a vicious cycle of AI getting trained on Internet junk, and AI responses getting added to it, which then gets scraped again and used as training data. But why blame AI. It is trained on the information we put out there. Information by experts available online is written in an authoritative tone of voice. So, AI adopted a similar tone when mixing facts with fiction.

The improvement in the accuracy of confidence level will make AI generated text more palatable. Accomplished economists and scientists tend to have a good understanding of how much they don’t know. They caution against the completeness of their own understanding, offer a caveat, or multiple plausible answers.

• As Silicon Valley retrenches, a tech talent shift accelerates