I’ll start this by saying I am currently working with a small AI research group, I don’t claim to know everything in the field (the more you know, the more you know that you don’t know, and there is always a place to learn more), but this is what I’m seeing around me.
I think a lot of people and companies are looking at AI the wrong way. From companies thinking they can fully replace people to “vibe” coders, and so on. On the other side there are people seeing it all as over hyped and not even a viable tool. Not being able to make hands correctly in pictures or tell how may Rs are in strawberry. I see it more like a new industrial revolution some what to when illustrator and photoshop came out and artist were not replaced, but they were retooled and small groups were able to accomplish a lot more that only big studios were able to do before. I see AI doing the same while effecting more fields if not close to all big companies that doesn’t mean slap AI on everything. While still needing competent people in each field even more to run these tools (artist may need to learn some python).
Some of the large LLM companies are starting to see a plateau and think the answer is keep shoving more of everything including crap. In a hope to make some kinda “cure all” model that they can sell to every industry and don’t see why they are still getting some crazy crap out. This is also akin to users over prompting an AI model bringing less weight to each word. Where distilled models like R1 have had better results in a smaller package. The problem with distilling is you still have the crap but it has less weight, and in this form I don’t think it is a good solution but is on the right path.
Where I have seen the most real growth in AI as a viable tool is in specialized modular and microservices architectures. With multiple small scope monolithic structures at the back. For an example of both structure and needing competent people in each field lets look at taking concept ideas or art to production ready game assets that is already being used to make different skins and weapons in games today. The small scope monolithic structures for this would/can be LoRA models, LoRA is a parameter-efficient fine-tuning method that allows for specialized training of many AI models without requiring massive amounts of data or computational resources. Instead of updating all the parameters of a large model, LoRA introduces a small set of low-rank matrices that are trained to adapt the model to a specific task or domain, This technique is not just in art or games but also starting to be used in healthcare, finance, education, coding and car manufacturing like czingers ai made car. Back to the example the typographers for the game would make a LoRA based on schnell keeping it’s algorithm and replacing it’s reference library with the one they make for the games text then trains it (training can be done on a single workstation). This is repeated for each part such as models being trained only on hands or weapons, trees, faces and so on. This not only fixes the crap in crap out problem it also gives better and consistent result, but the person making it would need to know about both fields and supervise results. These LoRAs are then used in modular pipelines/workflow to be able to none destructively change larger custom AI models. Resulting in something that looks more like unreal engines blueprint system and able to take text or image prompts and give textured 3d models that fit in the game.
In the example above you can replace typographers with medical researcher or programmer and schnell with what ever LLM and they train their models on their research or code base (vibe coders you still need to know coding to not get crap in crap out) to be used in conjunction with others. I am not saying LoRAs are the only way to do this as well.
End point, I see AI being a viable tool that will let small groups get a lot more done, but not in the ways it is being hyped as a cure all, it’s a toolbox going from screwdriver to power drill. If you think I missed something/ got something wrong please let me know. If you are working with AI on a professional level please write what you think is a realistic look at the AI industrial revolution.
