2022: The rise of Generative AI

What are the top development in AI we experienced in 2022?

1. Generative AI is a reality: it works. It's a Cambrian explosion of new, ultracapable AI tools.

The Global AI investment market is projected to reach $422.37 billion by 2028. Over $2B has already been invested in Generative AI, up 425% since 2020 (Financial Times/ Forbes).

Generative AI refers to the new type of artificial intelligence apps, trained on vast amounts of data, that can create new meaning from text, images, code, and other forms of content. Leading generative AI tools are DeepMind's Alpha Code (GoogleLab), OpenAI's ChatGPT, GPT-3.5, DALL-E, MidJourney, Jasper, and Stable Diffusion, which are large language models and image generators. This year, OpenAI (text-generating AI ChatGPT is built on it) and the makers of Midjourney and Stable Diffusion released public-facing products that set off a Cambrian explosion of new, ultracapable AI tools, allowing millions of people to experience generative AI for themselves.

Some argue that these apps will destroy millions of jobs, while others say they'll be a boom to human creativity. But whether you're an AI Apocalyptic or Integrated ( a fancy Umberto Eco's way to say, pessimistic or optimistic...), this year's advances mean that we are no longer debating theoretical benefits. It's a matter of fact that the results are real.

1a. Large Language models finally delivering results: ChatGPT &friends.

Many LLMs were released this year: they can compose phrases, sentences, paragraphs, and whole essays in a few seconds and are just prompted by a single question. Others can sufficiently translate from many languages to many languages. The advances are so impressive that someone even says that LLM landed what Mahiro Mori called the uncanny valley. "A place in which a technology modeled after a human is close enough to human-like behavior that human observers can be fooled sometimes but are creeped out by what appear to be oddities in the technology's behavior at other times." For instance, ChatGPT can confidently compose responses, grammatically correct, relevant, and well-paced…but simply wrong or socially not acceptable. Bias, not transparent sources, or not being able to distinguish between real and fake news. The sources point is debatable: is that inevitable and by design? If you train an algorithm with thousands of sources, you won’t be able to identify which source was more or less responsible for a particular result-. All of this is precisely what Mori predicted the uncanny valley would feel like…

Will LLMs replace Google? IMHO we are far away from that. And we are talking about different tools. Google is a search engine; humans usually generate the content it searches for. ChatGPT is generative text AI ( we may extend the comparison to large images AI vs. Google) and a fast and good interface ( a much more human-friendly way compared to a keyboard where humans can interact with machines! ) simulating to "understand" the human language and simulating meaningful replies. It's pre-trained. It means that its responses are not updated (it was trained in 2021), and in some cases, by design, it will offer you a too politically correct view of the world ( try asking, "if you were the President of the USA, what would you do with Ukrainian war).

Besides that, the cost of this everyone's accessible super, fancy demo is quite high. The cost of running ChatGPT is $100,000 per day or $3 million per month.

1b - Generative AI art …everyone is an artist!

The advances in text-to-image AI art generation have been terrific. AI image generators have been trained on millions of combinations of existing images and captions to produce new images based on text prompts. The AI doesn't know what it is doing; it only knows what imagery-specific descriptions are associated with. The speed and ease of use allow users to let their imaginations run wild, creating imagery by simply typing what they want to see. And that's led to some weird and wonderful creations... still not perfect… DALL-E seems more comfortable with Cat-heads than pig-heads and is still quite confused when imagining a pig-head human running after a rabbit….while clearly Google is quite accurate when searching for "cat head human". Anyhow, as highlighted previously, we can't compare Google with Generative AI since they are different tools. As of today, humans ( Google is searching content created by humans) seem much better capable of drawing Cat head humans vs. the ability of DALL-E to generate Cat - head humans when prompted with a specific request...

Also, salmons swimming down rivers is not the most straightforward concept to be represented...

Besides some inaccuracies, AI art generators are no longer an amusing curiosity for a few bunches of experts. It's a powerful technology everyone can access. This new reality is driving a sensitive debate: what does the commercial use of DALL-E's AI-powered imagery mean for creative industries and workers? Will it replace them? According to OpenAI, the answer is no. DALL-E is a tool that "enhances and extends the creative process.". I wrote about the relationship and intersection between AI and human creativity a few years ago...my article is here.

An artist using DALL-E ( or a writer using StoryLab or NovelAI ) would look at different artworks for inspiration. Generative AI can help an artist develop creative concepts…and unleash the artist in you. Probably you've already seen Instagram posts and stories produced by AI. If you are curious to get some fancy and weird selfies, try: Lensa AI, MyHeritage AI Time Machine, MidJourney, DreamBooth AI, Wombo Dream, NightCafe AI.

2. Machine Learning for everyone: No-code/ low-code solutions

Low-code and no-code development tools and platforms are on the rise worldwide, becoming increasingly popular and accessible. Thanks to them, it's possible to develop apps more quickly, with smaller teams and much lower budgets using custom software development tools. They typically operate in two ways: via a drag-and-drop interface, where users choose the elements they want to include in their application and put them together using a visual interface, or through a wizard, where users answer questions and select options from drop-down menus. Now also, ML can benefit of low-code tools...

Some of these solutions are designed for people who need experience. At the same time, some are most useful for people with a background in ML but want to tweak and fine-tune the results to create applications that operate more specifically or reduce the tedious and routine element involved with preparing data and designing algorithms. In a similar manner for writers and artists, those tools can help you to draft the code and to minimize your mistakes.

Some of the most popular tools currently used that will see further development in 2023 are:

Akkio: it promises you can start deploying AI in 10 minutes without any coding or data science skills. It enables the creation of AI-powered workflows that focus on allowing them to be quickly deployed and assessed. It also boasts a robust suite of integrations, including industry-standard data platforms.

Apple CreateML: simple drag-and-drop functionality that makes it simple to create iOS applications involving recommendation, classification, image recognition, and text processing. Data can be collected using your iPhone camera and microphone, and if you have a Mac computer with a GPU, you can use its power to speed up and enhance the training process.

DataRobot: cloud-based platform with tools to automate data prep as well as building and deploying algorithms, with dedicated models for industrial use cases ranging from banking and retail to healthcare, manufacturing, and public sector bodies. One exciting feature is its focus on explainable AI, which aims to inspire trust in the insights and decisions it produces by making its methods understandable to humans.

Google AutoML: this is not for beginners, and some understanding of ML is recommended. But you can get started with a simple graphical interface and jump into experimenting with its computer vision and natural language processing capabilities.

Google Teachable Machine: more beginner-friendly than AutoML: tutorials guide you through the process of training algorithms to classify and categorize data. Most useful as a teaching aid to learn the basics.

Microsoft Lobe: a simple tool for training image recognition algorithms and learning the basics. The platform automatically selects the most likely successful models, depending on the user's workload. No coding experience is required.

Nanonets: AI platform designed to automate and speed up the process of extracting structured or semi-structured data from documents. Thanks to its implementation of ML, it learns from its mistakes to become increasingly accurate at finding the information you need.

ObviouslyAI: a platform that aims to let anyone plug in their data – in whatever format they happen to have it – and immediately reap the benefits of AI-powered analytics. It offers templates for time series analysis (predicting the value of variables at a given time based on known past performance), predicting churn, risk scoring, fraud detection, and identifying cross-selling opportunities.

PyCaret: a library for Python, so it requires some technical knowledge. It can be considered low-code, though, since it provides several pre-configured functions and wrappers that vastly simplify the task of data preparation, analytics, and model training.

Over the last two years, one of the most common ways for organizations to scale and run increasingly large and complex artificial intelligence workloads ( like OpenAI) has been with the open-source Ray framework. Ray is an open-source unified compute framework that makes it easy to scale AI and Python workloads — from reinforcement learning to deep learning to tuning, and model serving. Ray enables ML models to scale across hardware resources and can also be used to support MLops workflows across different ML tools. The tool's next major milestone Ray 2.0 extends the technology with the new Ray AI Runtime (AIR) that is intended to work as a runtime layer for executing ML services.

3. AI in medical context: Setting up the basics for new levels of accurate, equitable, effective patient care.

A quieter inflection point is happening for applied AI in medical fields like radiology and electronic health record analysis. The advances over the past year are precisely the type of incremental, basic foundational steps we would expect to see. There has been work to create atlases of proteins, genomes, and brains. There has been investigation of bias in the underlying data, stemming from sources such as racism and sexism, that impede the ability to unproblematically pool data to get the size of datasets required. For instance, if doctors were historically less likely to react the same to reports of pain by women and people of color, this would be baked into data, making it non-poolable in frustrating ways.

More and more evidence shows that training AI algorithms on various datasets can improve decision support, boost population health management, streamline administrative tasks, enable cost efficiencies and even improve outcomes. 

Diagnostic Robotics received the World Future Awards for its AI-Powered Population Health Management solution. It helps health plans and providers identify, prioritize and segment clusters of future-risk members based on historical utilization patterns, comorbidities, and other factors. It identifies the most impactable members and provides actionable insights through regularly updated lists of members with the most significant risk of preventable health events. The solution has shown demonstrable benefits to patient outcomes, operational efficiencies, and ROI in case studies with regional and national customers.

But there's still much work to ensure accurate, reliable, understandable, and evidence-based results that provide patient safety and account for health equity.

While large language models made tremendous advances, applications of AI in medical contexts did a significant amount of the basic science required to get out of the uncanny valley and produce the tools that can help human clinicians reach new levels of accurate, equitable, effective patient care. Medical care will not necessarily be cheaper with AI, but it will eventually be better. It takes a long time to find those non-comparabilities in data, so we will probably see much of the same slow, incremental work in medicine and AI next year.

If you were asking yourself: did she write this article using ChatGPT? Unluckily no. I've tried, but it refused to write it for me.

Besides my own opinions and knowledge, I consulted ca. 20 articles... but like an LLM, I will only report some of the sources I consulted. In the future, I plan to use LLM to draft my pieces.

Below is what ChatGPT would reply to you about the latest trends in AI (it doesn't know it is itself a trend:-)!)