2023 Recap of AI
Reviewing the past wild year in AI
What a year it has been! This past year of 2023 has been the year of generative AI. While generative AI has been around for a few years, it exploded this year with the rise of ChatGPT and the speed at which AI models have been created. Below is a recap of the main happenings of the year, followed by emerging themes.
Widespread usage of ChatGPT - If you haven't heard of ChatGPT, you've probably been living under a rock. It's a large language model (LLM) that started as the best chatbot available and moved on to much bigger things, such as being able to talk about images, surf the web, have an agent marketplace, and make function calls. This past January, ChatGPT set the record for reaching 100 million active users two months after launch.
Release of GPT4 - GPT4 was released in March, and it is still the best LLM that has been released. Most open-source models compare themselves to GPT3.5 because they cannot match the performance of GPT4.
Release of multimodal models such as LLaVA and GPT4-vision - Multimodal models allow for interpretation of multiple modes of data, such as images, text, and audio. This is powerful and came about as people wanted to interact with more mediums than just text. Being able to ask questions and interpret other mediums, such as images, is very powerful. Expect these abilities to become commonplace next year.
The beginning of AI agents - Projects like LangChain, AutoGPT, and BabyAGI started paving the way for using LLMs to accomplish a variety of tasks when given a goal. AI agents will transform how we interact with LLMs as we can use them to do more. The best way to interact with AI models will likely not be telling them explicitly what to do but by giving them objectives and letting them solve it for us. Currently, LLMs tend to be very hands-on in their usage, but AI agents provide a path forward that changes this.
The start of hyper-personalization - Hyper-personalization is modifying content so that the wording or imagery changes according to the individual viewing it while still retaining the core message. Companies and individuals understand that this level of personalization creates huge lifts and is incredibly powerful. While this first emerged in marketing contexts, it extends to how websites display themselves, political discourse, and framing many different types of daily conversations.
A huge release of video models - At the end of this year, we saw a huge dump of video generation models from Stable Diffusion, Pika labs, and Meta. While these are still shorter-form videos, the results are exciting for what the next year will hold.
AI regulation - The EU signed its AI act, and the Biden administration issued an executive order on AI. As the public begins to grapple with the power of what can be done with AI, the call for regulation will continue to rise.
Studies on work improvement - Ethan Mollick, a professor at Wharton, released a study along with BCG about the 'jagged frontier' of how AI assists individuals in their workflows. The use of AI lifts the performance of most professionals, with under-performers seeing the biggest gains.
Retrieval Augmented Generation (RAG) - This technology is something that BrandGuard has been doing well before it was called RAG with our release of BrandGPT earlier this year. RAG is a way of extracting relevant content to a question to provide a more informative and correct answer. For instance, BrandGPT takes a brand's style guide and provides correct answers and page numbers to questions related to a brand's style.
Model shrinking - As the power of AI is realized, many people want to use it for IoT or edge computing type devices. The problem with most AI models is that they tend to be too large to fit on a single device. A lot of work has been done this year to find ways to shrink models and reduce requirements so that they can run on devices while still remaining performant. Expect this to be an ongoing process, but it is unclear if elements of Moore's law will help with this.
Themes
Verification - Verification is emerging as a prime interest in companies looking to adopt generative AI. Despite individuals loving LLMs, corporations have been resistant to adopting new generative AI models at scale because of their inability to trust and control the outputs.
Open Source vs Closed Source models - There's a war going on between models that are proprietary and those that are openly available to the public. Meta gave a great boon to the open-source community when they released their LLaMA LLM model. It was quickly iterated on by the community to create many variations along with multi-modal capabilities. At the moment, no open-source model can fully compete with GPT4, but time will tell if open-source models will be able to meet or exceed the capabilities of closed-source models.
GPU hardships - Getting access to GPUs has been a challenge for AI products and researchers. This has caused researchers to figure out how to run AI models either on CPUs or with much lower hardware requirements. It has also caused Nvidia's stock price to skyrocket, and competitive pressures have arisen to find other ways to add to the GPU supply.
An explosion of ideas - The world of possibilities feels much greater. The world has a better understanding of what AI is capable of. People are iterating and trying new things with AI. Expect to see lots of obvious, non-intuitive, weird, and incredibly helpful things materialize into the world due to applying AI on the right problems.
AI Catch Up - Every big tech company is looking to catch up to OpenAI in terms of model ability. Google, Meta, Amazon, Apple, and X are all releasing their own LLMs. At the same time, other LLM companies have emerged, such as Anthropic, Mistral, Cohere, and Inflection. As the bar for performance continues to rise, we are seeing dispersion between companies and models as time goes on.
Increase concerns of security - People began to understand the widespread use of their data. A common question we hear at BrandGuard is how we store brand data (we keep each brand siloed) because companies and individuals are beginning to understand just how important training data is for models, and where that data comes from has various implications.
As you can see, a lot has happened this year in 2023. It has been a bewildering pace for those of us in the midst of the AI explosion. At the beginning of next year, we'll dive into what we believe is upcoming for AI in 2024, partially pushed by the emerging themes above.