Im definitely on the side that over using AI and using it commercially seems to be bad. On the other hand, it seems like a tech that has huge potential upsides. I’m not sure we can achieve a post scarcity society with all labor being done by humans. This is where I see AI becoming a massive tool. Assuming we can pair it with mechanical means of work, not strictly digital. I know it’s a touchy subject but I want to hear your opinion. As always, if you’re just going to tell me to read more, recommend literature.
Machine learning by itself is already paying off. LLMs the way normal people use it are fine, but not the way bosses want to use it - you can’t rationalize away employees with this, you can only give them them a tool that empowers them; there will be a lot of heads rolling in management in the corporations where this hasn’t been realized yet.
Code Generation might get better over the next years, but looking at the current trajectory i would say with current tech there will never be a point reached where you can simply replace a dev with an agentic AI without the generated code being full of inefficiencies, bugs and security issues.
It also might open up the first therapeutic LLM (without the current fuckups) if there is a focus in development - THIS would be something that is labor intensive, priced so that exactly the people that need it can’t afford it on a regular basis, and definitely possible to attain without much of a technological limit.
ImageGen can also pay off in some areas - creating tons of “stony wall” textures isn’t fun, but implementing a tiny model that creates as many “stony walls” as you want in your game with differing amounts of stones or dirt as a variable might be worth the effort.
Prototyping is also a big thing in both of those areas (Code Gen/ImageGen), but i think that’s no secret anymore.
Videogen in the current way is a waste of energy. The models need something that anchors them, to make sure that the Coke truck in one camera angle stays the same Coke truck in the next angle; currently it’s just ImageGen *25/second, which causes those issues - and the massive energy consumption. This is the only area where the generation process itself chews through more electrons than the global energy bill allows for. Someone smart will probably crack that nut too - i believe that the solution to those 2 issues (energy consumption and missing object permanence) might be linked.
Edit: The missing permanence might also be a reason for many of the issues of LLM’s, some kind of “self” with a sense of the passage of time to return to. I’m pretty sure i’m not the first one who thought of that, and there are probably a lot of people with even more PhDs at work here.
There are also serious gains to be made in science on the back of AI models, just not the flavors most people are familiar with.
https://pmc.ncbi.nlm.nih.gov/articles/PMC8633405/
Image generation uses a concept of a LORA that is a bolt-on model to augment a base model. It can provide support for additional tokens (those map more or less to words and concepts), or bias the base model on existing tokens. For now, that’s probably as close as you’re going to get to anything resembling long-term-memory on an LLM.
Of course, i was just fixated on the GenAI-Aspect of the question. Climate modeling, Medicine (especially diagnostic and neurosciences), physics, chemistry and even social sciences can benefit a lot. In the manufacturing process it can be used for control of industrial processes (e.g. balancing of a chemical process; i know it’s used in QC in production of electronics). This IS the next big thing, but not in the way the corpos try to sell it to us.
Yeah, for the moment, this is it. We will need more research in how to actually read the data inside of a neural network to be able to a) reroute pathways that are problematic and b) enable a process to hook inside of control points to modify it on the fly.