My Expectations for AI in 2024

A few modest predictions and personal hopes of what AI may bring in 2024

The end of 2022 and the first few months of 2023 were defined by the launch of ChatGPT and its improved GPT-4 version. The rapid innovation and evolution of such technologies, including their almost immediate incorporation in many core digital productivity tools and workflows, seemed to suggest that even more advanced technologies would come our way by the end of 2023. For me personally it all felt like a whiplash - exciting, disorienting, concerning. However, the second half of 2023 brought more of an evolutionary change and consolidation. Nonetheless, the improvements of GenAI tools across the board is still incredibly impressive. However, there are still many gaps in their capabilities, both perceived and actual. It’s possible that most of these gaps will be definitively resolved with the arrival of AGI, but at this point I am fairly confident that we’ll manage to sail through 2024 without this epochal event transpiring (at least not externally).

Probably the single biggest AI development that we can expect some time in 2024 is the release of GPT-5. OpenAI has gone back and forth about suggesting whether such model is in the works and/or when it might be released, but all the smart money is that it’s coming some time in 2024, not least because Google has already announced their upcoming release of Gemini Ultra, which, according to the original product announcement, is more capable than GPT-4.

However, regardless of whether GPT-5 and similarly capable LLMs get released, my sense is that we are approaching the end of the road of what LLM-based GenAI tools can accomplish on their own. I still believe that we can get a few more qualitative leaps forward from their capabilities, but we’ll probably need to marry LLMs to some other tools and techniques for the next level of truly capable breakthrough.

My own hopes and aspirations for what AI can bring us in 2024 are primarily practical. I want to be able to do more, much more, with AI than I was ever able to do before on my own. I want AI tools to significantly increase my productivity, and enable me to do things that would have previously been out of reach. In Andrej Karpathy’s formulation, I want AI that is all about IA - Intelligence Augmentation. Here are a few particular things that I hope we get in 2024:

1. Fixing the hallucinations. Hallucinations keep plaguing the AI apps. Chatbots give us links to nonexistent articles, make up news stories, write code that doesn’t work, give solutions to math problems that make no sense. Image generation tools have come a long way, but they still end up giving us some weird artifacts, and are a constant source of jokes and amusement. IMHO, fixing hallucinations will be one of those things where we need to go beyond LLMs. Finding statistical correlations is not the same as thinking and reasoning, although those two approaches can feed off of each other.

2. Consistent, reliable and enduring AI generated content. For all AI systems the outputs of the queries are limited to a single session. And in many cases a single inquiry. It would be great to get personalized chatbots with your own chat/query history, and leverage that for all your future sessions. In visual realm it would be great to get consistent visual style, and even greater to get consistent characters. This would open up a whole new realm of creative work.

3. 3D GenAI systems. I believe that all of the current image generation tools are ultimately limited by their 2D underlying structure. There is only so much of fiddling that you can do to make that more robust. 3D AI and spatial understanding would push the envelope of what can be done with visual GenAI systems.

4. Even more advanced AI coding assistants. This is one area where I’ve heard a lot of effort is being put not only into the algorithm, but also in generation of more robust training data. No, GenAI will not be taking away any coding jobs (and may in fact create even more demand for them), but we’ll see a proliferation of tools that will make all coding jobs significantly more efficient.

5. Getting proper reliable AI agents. AI agents are still in many respects the Holy Grail of the current stage of AI development. Short of getting the full AGI (whatever that means), enabling people to reliably get a significant portion of their own work done by autonomous systems will enable countless new personal and professional opportunities. This is where I think that fixing the hallucionation problem will be crucial, as well as going beyond the LLM-only approaches.