AI definitely has been making waves in the world of content creation this year. We took a closer look at what has happened with generative AI in 2022 and picked up the most significant stories that will certainly set the game in the future.
The Skyrocket of The Year: ChatGPT
The first time capabilities of AI as a human-like writer wowed the Internet was in 2020 when the research and development lab OpenAI released their NLP (natural language processing) model GPT-3. The system was able to perform a wide array of tasks, including language translation and text generation — from creating marketing copies and blog posts to designing chatbots. Some said, this ‘state-of-the-art’ language model could be an early glimpse towards AGI — artificial general intelligence — the program at its high-level scale that not only masters tasks just like a human does, but also has a thorough understanding of the real world in its complexity and possesses a human-like mindset.
In a couple of years, multiple AI researchers have been spawning new models grounded on GPT.
The timeline chart mentioned above doesn’t include the brand-new model from OpenAI. ChatGPT, an improved version of previous models, was published in November and went viral. According to co-founder of OpenAI Greg Brockman, the model gained 1 million users in under a week — faster than Instagram or Facebook did. It’s able to communicate in a more conversational way by giving answers to follow-up questions, admitting mistakes, and assisting in writing code. Some compare the event with the development of electricity, the arrival of the printing press, or name it ‘the biggest change to the Internet since crypto’.
In spite of having its pitfalls and weaknesses, ChatGPT is capable to provide accurate and comprehensive data on users’ queries. Many argued that it may squeeze Google out of its number one position at the top of search engines. At least, sometimes ChatGPT beats search engines in some cases — for instance, providing the ultimate guide to Pokemon types instead of burying a user in tons of links as Google used to. Although Google doesn’t sit around — they are also working on their MUM and LaMDA tools which are also supposed to rethink the searching experience by building a conversational dialogue between users and the system.
Anyway, the reinvention of searching isn’t something that is waiting for us out right around the corner. Looks like neither OpenAI’s nor Google’s models are capable to change the way we used to search so far because this process is more complicated than just compiling Pokemon types guide or providing weather for tomorrow. The existing Google’s ‘One true answer’ problem is going to be even more serious for GPT systems.
But what is real in the very near future is a skyrocketing raise of apps and services underpinned by natural language processing models which opens up possibilities in specific niches — customer service, sales, marketing, legal services, and even biology and social science. We don’t even realize all the benefits (and risks either — some professors are already worried about school assignments) are lying inside the magic box. Moreover, the accumulation of NLP and text-2-image technologies would become an early stage of a product with a human-like interface. We’ll be watching.
The Breakthrough Of The Year: Text-to-Image
Nearly 2 years ago AI research and development company Open AI released a deep learning model called DALL-E. It was a successor to GPT-3, a previously revealed text generation model. DALL-E was trained to create images from text prompts and made a buzz in the news fascinating people with coherent but still quirky concepts of armchairs in the shape of avocado or baby daikon radishes wearing tutus.
At that point, DALL-E was already an impressive milestone in the field of generative AI. As with the next OpenAI model DALL-E-2, the tool wasn’t openly available for general users: you could play around with a range of built-in inputs there.
During a couple of years, text-2-image models have been replicated increasingly. Neural networks became capable to make more realistic imagery. Anyone with access to the Internet got a chance to test the water with Craiyon (as known as DALL-E mini), Midjourney, or Stable Diffusion models and ask whatever they want AI to create.
Social media has been drowning in tons of generated images — sometimes realistic, sometimes odd, but still impressive. What becomes probably even more viral over the last few weeks is an image editing app Lensa, hosted by Prisma Labs, the company that built a popular style transfer app back in 2016. Lensa was released in 2018, but the recent feature ‘magic avatars’, based on Stable Diffusion algorithms, hyped up Lensa: it crafts a bunch of unique portraits from 10-20 user selfies rendering into cyberpunk, anime, or fairytale styles. Despite security and ethical debates around the app, it reached over 5 million downloads worldwide in the first fortnight of December and had 1.8 downloads in November, according to Statista.
The First Step of The Year: AI Content Starts Being Pulled Out of a Grey Area
All these years since AI content first arrived stock image market players used to be prudent to accept its legitimacy because of concerns about copyrights as well as reputational risks. But it’s getting harder for them to ignore the fact that AI content has started to change the game. Stability AI, the company behind Stable Diffusion, released their model under a license that allows commercial and non-commercial usage of content extracted with their software. OpenAI offered their customers using DALL-E for commercial projects as well.
A few months later, Open AI and Shutterstock announced their partnership, according to which Shutterstock started selling content generated with DALL-E. Platform contributors whose content was involved in the training dataset are supposed to receive compensation — to keep its reputation, Shutterstock creates Contribution Fund for this purpose. Exactly on the same day, another stock photo agency Getty Images announced the strategic partnership with BRIA — to ‘give users access to state of the art, ethical, Generative AI-driven features that will expand their creativity and improve efficiency’. And finally, Adobe Stock has recently accepted AI-generated content without any limitations on the tools that can be used. Besides, they require their contributors to ensure they have all the necessary rights for images.
Meanwhile, some artistic communities keep fighting with AI content — this is how the homepage of ArtStation, a portfolio platform for digital artists, looked after its contributors’ protest against featuring AI content on the website.
The legitimacy process has its rises and falls, but now it’s official: the market has started to adopt a new way of content creation, though the legal basis for intellectual property certainly needs to be updated. Plenty of questions about using AI-generated content still remain on the agenda so far. One of them is who owns the copyright to pieces of generated content. We’ll probably see the answers soon — at least, a few big steps towards taking AI out of the grey area have been made.
The Problem to Be Solved of the Year: AI Safety
As mentioned above, all the concerns, fears, and risks regarding AI evolution took broad discussion both in the media and professional community. The alignment problem, copyright, and ethical issues as well as other risks are worrying even those who build AI. According to the survey of ML researchers conducted by AI Impacts in 2022, 69% believe people should more prioritize AI safety than it currently does (up to 49% in 2016). Another survey was conducted by scientists and asked NLP researchers about controversial questions in the field. The results show 58% of respondents believe that understanding of benefits and risks associated with AGI should be a significant priority for the community.
Looks like a new specific field of science (and market?) focused on studying AI safety issues is about to be shaped. As of today, dozens of countries have national AI strategies, as well as signed agreements and multinational strategies which refer to AI safety as a serious subject to watch. The rise of interest in AI safety motivates a new wave of non-profit (Center for AI safety) and business organizations such as startup Conjecture which has raised millions from investors including the founders of Github, Stripe, and FTX, according to the State of AI Report. A research analyst Benjamin Hilton estimates that around 300 people (the interval could range from 100 to 1500 people) worldwide work on reducing existential risks from AI, though he admits his estimation may be subjective. There are companies hiring people for AI safety research roles on Linkedin or job search websites like Lever.
Well, what will AI safety pros do at their job? Here are a few key directions that we tapped into this fresh field:
— Researchers — are the people who focus on the safety of novel technologies that are already here: for example, they strive to improve human-machine interaction in using autonomous vehicles. Today, this field is the most approaching to actual business needs.
— Scientists. These talents might be able to interpret and understand the internal mechanisms of AI systems as well as predict possible outcomes of AI actions and measure its moral behavior through experiments. This sort of research, for example, is provided by Anthropic, an AI safety lab.
— AI advocates. Increasing awareness of AI, its actual risks and benefits, as well as expanding people’s knowledge of how AI works, how it could be safely integrated into their life, and how they should be prepared for changes is a crucial part of building the safety future with AI. That is what scholarship and educational programs aimed to do: from AGI safety fundamental courses to Vitalik Buterin PhD Fellowship in AI Existential Safety.
Bonus. The Buzzword of The Year — Metaverse
While running through our long-read piece, you may have thought — what about the Metaverse? Yes, we couldn’t pass by the concept that has been going around during this year, though some argue that hype runs ahead of anything tangible about the Metaverse. At least, the experts think that up until now the Metaverse is still just a part of the Internet, not a specific type of technology. Neither an innovative idea nor a shift in how we interact with the existing VR/AR or physical assets were suggested yet. Well, maybe 2023 will get a sense of whether the buzzword turning into reality or it remains in Metaverse. Notably, Meta is going to devote about 20% of its costs in 2023 to developing Metaverse, according to Bloomberg.