I would add that I enjoy experimenting by running LLMs locally and using remote Virtual Machines. But at this point I would have to admit however that at this point this is not a truly worthwhile endeavor. This is because the “frontier models” used by major AI providers have been perfected and “fine tuning” by using almost unimaginable amounts of processing power to train them.
But it is also because the major AI providers are still massively subsidizing their products to businesses, governments, and people. I have used these products to the point that it would not be possible to calculate what percentages the paltry sums that I have paid for subscriptions and billing through APIs covered the actual expense that the large providers incurred because of my activities.
I initially assumed that it was the large companies and governments that use AI to keep track of the activities of massive numbers of people that took up the vast majority of compute being used by AI providers. But when you look at the numbers as closely as possible as to how much computing power is being used locally and remotely by virtual machines to do similar work, it is staggering. The experience that I am referring to is generating images and video.
I paid a couple hundred dollars to a company called LTX Studio, https://app.ltx.studio/ to restore and animate old family pictures. The results were quite amazing, but I thought that there must be a cheaper way to do this. So, this motivated me to spend $Thousands on hardware, remote VMs and subscriptions and APIs from the major AI providers trying to duplicate what LTX studios did.
I did have some success, but by and large the only viable alternative for this type of activity were subscriptions to the major AI providers. You can get results that would and did cost hundreds of dollars... from a heavily discounted subscription to one of the major AI providers. Even though I spent a substantial amount on local hardware that I intended to use for this purpose, it just wasn’t enough.
Currently, the hardware available that comes close to being viable for this purpose is from Apple and costs around $20,000. There are specific purposes and reasons why this type of expenditure is worth it for some businesses, but for a hobbyist such as me it would be impossible to justify at this point in time.
I don’t do much with AI excepting the PC issues I mentioned and a laptop using battery overnight when it is supposed to be hibernating. Not some battery, all of it. I suggested to turn off wifi overnight; the AI took that farther with a script that shuts down most processes before I close the lid and my battery now maintains. Windows was the culprit again, with complex crap waking the sleeping device every few minutes or seconds for no good reason.
I discuss & compare w/ the AI such as SpaceX, Blue Origin, NASA programs, goals and timelines and it gives me lots of particulars that I trust, for the most part, if it covers reported facts. Or any random thing about a movie, current events and so forth. It makes stupid mistakes rather frequently; it is a reference source but...trust but verify.