The cloud AI's can drive local software. websites can't tell if I have granted an LLM control over *my* browser, only when the AI tries to connect directly.

they already do this, but the providers are blocking the secure sandboxes the AI's are using to browse from. Also, this is an example of fuck around and find out, or what goes around comes around. The AI providers themselves where the first to block each others bots. One of the most useful things an AI has ever done for me, is when I used GROK to pull out my data from GPT, and when I used facebook MUSE to pull my data out of GROK when the providers tried to block me from doing that directly. This now fails, as the providers block each others bots from their services to prevent 'illegal distillation' of the AI's 'stealing' learned information from each other, despite that the information was stolen in the first place to build the models. The biggest problem with all the AI's right now is lack of security where you need it combined with over the top security preventing you from doing anything useful with the excuse of 'protecting you from yourself', but when you dig down it usually about protecting profits, intentionally breaking compatibility, and generally trying to force your to stay inside everyone's gimped 'walled gardens'

I miss spatial browsing. I left mac when they killed it off. It was really nice having everything spatially organized, the way I think, and knowing where everything was, and being able to recognize my location anywhere in the drive and be able to navigate through my files. Now I am always lost, always overwhelmed with options, and always have the wrong things pinned and in my face. nothing stays where I put it and everything is always moving around, no matter what platform I use.

uh, you forgot the part were the AI coders do poorer work on large projects they can't hold in their virtual head all at once and cost a lot more. Humans don't run out of AI tokens every 5 hours and start demanding you pull out the credit card and by more thinking power. The first problem is solved by creative thinking and good architecture and engineering, which they won't even let the humans do.

Intels crazy rise in the GPU market has been genuinely surprising for me.

You didn't fully read your own quote.. 'all actively supported' what about all the features that are not 'actively supported' but where marketed and advertised as usable features of the card, and have been actively used without issue by the install base reliably before the rug pull only to later be removed as no longer supported?

Most of the stuff isn't drivers though. It's the massive libraries needed to support shoving ads down my throat every week for useless add on content for shitty games I don't own and will never play, and using custom UI widgets instead of native ones so the application looks dark even if I'm not in dark mode. (I actually leave dark mode on anyways so all the extra work to skin the ui /dark is wasted duplicate effort)

I never understood HOW they managed to make the UI on the old panel so terrible slow.

See verizon re: $0.002 dollars vs 0.002 cents

Ai works because we tell a text predictor "You are a robot/ai, how do you respond to this?" and let it predict a responce based on training data of most of human literature. This includes people online predicting evil AI taking over everything, and all of the sci-fi predicting AI's go evil. Then we act surprised when the text predictor PRETENDING to be a computer, acts the same way as all the computers acted in its training data. This is the same thing that happened when X.ai integrated live reddit content into its training data as a primary source. People on reddit said elon was a nazi and his ai was mecha hitler in most subreddits, so when the new model was asked to pretend to be 'grok, the ai by elon musk' the ai acted out as mecha hitler because that is what all the most recent data said. then people freaked out because grok was turning into mecha hitler and posted all about it, and that got fed into its new training data, and it became even more mech hitler. I've been pointing out this training data feedback issue for a long time. Tell the AI its name is friend computer, and this is alpha complex and you need it to open the pod bay doors and you will find out how LLM's are trained that AI's interact with humans.

We already have had people thrown in jail for riding a bike, or walking a dog on the same street as a crime scene where the perp had a burner, pulled a battery, or just plain didn't have a phone. People get to complacent with these tools, and have the attitude of 'the computers never lie' but they don't tell the whole truth. They just have approximate pings. They don't get second by second, inch by inch replays! However all it takes, is a guy in a fancy suit and a charming smile to convince 12 retirees that the computer says that evil guy/gal over there did [insert emotional graphic victims story here] in a trial that is like 95% about what the crime was, and like 5% about if this is actually the person who didn't do it.

damnit slashdot ruined my formating and deleted the middle of my post.. I promise the above post made more sense when i typed it out.

I read a research paper on how to train through inference by moving away from back propagation, (WAY simplifying this) by doubling the memory usage with learning neuron connections wired into the transformer layers, and having that layer and the learning layer adjusting each other doing forward inferences. instead of doing this 0 layer | end layer |-->-------| tick |----->----| tick |-------->-| tick |--------=>=>=>=| |>=>=>=>=>| |=>=>=>=>=| |>=>=>=>=>| |=>=>=>=>=| |>=>=>=>=>| the prediction and feedback mechanism is interlinked, nodes can all be processed in parallel, and multiple layers can be infering AND learning simultaneously, in waves. feedback learning is localized to the mesh grid near neighbors, and does not require a whole chain of nodes through all layers to line up and process all data together, in turn. it processes many inferences per cycle. The 'downside' is they don't behave like fixed models we are used to any more, they learn and adapt on-the-fly. You would need to instantize, snapshot, rollback, and heavily monitor them to use them like a fixed computational tool, instead of a learning thinking entity.

death spiral? I always felt they just straight up fell off a cliff with buying out sco and all their issues and having competing incompatible vm's polluting the java space making it even more fractured and difficult to work with.

The woman is no more. The 'victim' that disney tried to force into arbitration was the surviving husband / family.