Where will AI be end 2025?
Happy New Year to you!
Happy New Year to all my subscribers! I want to thank everyone for reading this for their support, be it sharing your views, liking the issues that resonated with you, or sharing my newsletter with your community. All kinds of support are very well appreciated! :)
Thought I write down my thoughts on where AI will be through 2025 and it can serve as a reflection to myself when I come back to this issue in Jan 2026.
Agentic AI Workflow
2025 will be the year of “Agentic AI Workflow”. Many organizations such as Microsoft and Salesforce have already hopped onto this gravy train, even to the point of announcing the death of Saas (Software-as-a-Service). However, I do hold a different viewpoint. How so? In a previous issue (Successful Agentic Workflow 2025), I mentioned that the tasks at hand need to be broken down into elemental tasks i.e. tasks that cannot be broken down further as the first step. This is not that straightforward. Think about it, how will you break the task of “Logging to your computer” into smaller tasks? Double-check it with another friend of yours if you have the same breakdown. There is a good chance that you’ll have a large overlap but still have elemental tasks that aren’t.
We still have a long way to go on this for Agentic AI Workflow, but I very much look forward to it as it will open another front on how human intelligence works which keeps me very excited! :)
World Model (Spatial Intelligence)
If you are a seasoned AI practitioner or have been monitoring this space, you will know that Prof Fei-Fei Li is another person to watch for what is upcoming in the AI space. Her idea of building a World Model is very interesting! So interesting that a startup is spun off from it and has raised $230M. In a nutshell, what Fei-Fei Li is trying to do is to create a 3D Environment for machines (including robotics) to train in first and then be deployed into the real world. The research horizon for this should be a few years but once it hits maturity, we can have robots that can operate in the physical environment, and if you are a gamer, a possibly less buggy and closer reality game environment! Not forgetting some of the possible use cases on Metaverse such as retail, hospitality, and healthcare. Definitely another space I am paying attention to although I find it quite surprising that the hype from it is not as much as Agentic AI Workflow.
Has LLMs (==AI?) Plateaued?
Recent news has it that LLMs training has plateaued and they are now being trained in synthetic datasets rather (article here). Has it? I am skeptical actually. If you look at the current development of LLMs, most of them are still neural network-based, and very likely transformers are used. While I do agree that whatever accessible data has been used to train the major LLMs, I feel this is an opportunity to now focus on the architecture of the neural network models. The appearance of RAGs and Knowledge Graphs at the start of the LLMs era says a lot that we need more work on the model architecture side, to reduce false regurgitation of facts, and not forgetting the number of “r”s in “strawberry (i.e. logic). Training it with all the available data is just a low-hanging fruit for building LLMs. What we would like to do going forward is to see if we can:
Use different model architectures to improve accuracy, reduce hallucinations, and work in better logic (or causality).
Train up an LLM faster - LLMs still need a relatively long time to train one up. It takes weeks to train one from scratch which means it will not be able to incorporate the latest information inside its parameters. This is a downside that is very important especially if we want to commercialise LLMs.
Smaller, deployable models - after training, we cannot have LLMs deployed just on the cloud alone. It creates a whole other set of issues such as computation, energy usage, privacy, etc. There is a need to have smaller models for deployment, again going back to model architecture.
Given the above, I am hoping we see a breakthrough in a different architecture that will allow for a better accuracy, deployable model come 2025. :)
Conclusion
All right! So these are some trends I will be looking out for come 2025. Let us see where the above will be come early 2026! Regardless, I am still very excited for what is to come for AI! I will love to hear your thoughts on this so please share them in the comments or PM me on LinkedIn!
Recommended Issues
Successful Agentic Workflow 2025
Good AI Governance is NOT the End-Game
Support my work, share my newsletter with your community, and network! :) Feeling generous? Drop me some “books” please! It will help me to get more materials for my learning! :)
All the best for 2025 Koo, and congrats again for the great articles always so content rich and relevant despite their conciseness. Keep it on!