3 predictions for AI in 2024
This last year was a banger for AI as the technology went from niche to mainstream about as fast as anything ever has. 2024, however, will be the year when the hype runs full-steam into reality as people reckon with the capabilities and limitations of AI at large. Here are some ways we think that will happen based on Techcrunch's article.
Agents, generated video and generated music graduate from quaint to experimental
Some niche applications of AI models will grow beyond “eh” status in 2024, including agent-based models and generative multimedia.
If AI is going to help you do more than summarize or make lists of things, it’ll need access to things like your spreadsheets, ticket buying interfaces, transportation apps and so on. 2023 saw a few tentative attempts at this “agent” approach, but none really caught on. We don’t really expect any to really take off in 2024, either, but agent-based models will show their stuff a little more convincingly than they did last year, and a few clutch use cases will show up for famously tedious processes like submitting insurance claims.
Video and audio will also find niches where their shortcomings aren’t quite so visible. In the hands of skilled creators, a lack of photorealism isn’t a problem, and we’ll see AI videos used in fun and interesting ways. Likewise, generative music models will likely make it into a few major productions like games, again where professional musicians can leverage the tools to create an unending soundtrack.
The limits of monolithic LLMs become clearer
So far there has been great optimism about the capabilities of large language models, which have indeed proved more capable than anyone expected, and have grown correspondingly more so as more compute is added. But 2024 will be the year something gives. Where exactly it is impossible to predict, as research is active at the frontiers of this field.
The seemingly magical “emergent” capabilities of LLMs will be better studied and understood in 2024, and things like their inability to multiply large numbers will make more sense.
In parallel, we will begin to see diminishing returns on parameter counts, to the point where training a 500-billion-parameter model may technically produce better results, but the computer required to do so could probably be deployed more effectively. A single monolithic model is unwieldy and expensive, while a mixture of experts — a collection of smaller, more specific models and likely multimodal ones — may prove almost as effective while being much easier to update piecemeal.
Marketing meets reality
The simple fact is that the hype built up in 2023 is going to be very hard for companies to follow through on. Marketing claims made for machine learning systems that companies adopted in order to not fall behind will receive their quarterly and yearly reviews… and it’s very likely they will be found wanting.
Expect a considerable customer withdrawal from AI tools as the benefits fail to justify the costs and risks. On the far end of this spectrum, we are likely to see lawsuits and regulatory action with AI service providers that failed to back up their claims.
While capabilities will continue to grow and advance, 2023’s products will not all survive by a long shot, and there will be a round of consolidation as the wobblier riders of the wave fall and are consumed.
(Source: Article "8 predictions for AI in 2024" by Techcrunch: https://techcrunch.com/2023/12/19/8-predictions-for-ai-in-2024/)
