In the broader SF Bay Area, our recent PG&E bills for a 50 year old single-family home without air-conditioning is under $150/mo, with a couple fridges, electric clothes dryer, and a half-dozen laptop class computers. That's averaging about 8-9 kWh per day (250-280 kWh/month).

Last winter, our bill ramped up to over $400/mo for a few months, due to heating with natural gas.

My biggest career horror was realizing how much the medical informatics concepts have been structured around billing and insurance rather than scientific, biomedical requirements.

It isn't a monotonic decline with memories disappearing forever. It is like wave upon wave of changing capacity at different time scales. The general trend is deterioration, but there are frequent periods that can almost seem like remission.

There is a well known daily cycle referred to as "sundowning", where the sufferer tends to come unraveled later in the day. The next morning, they'll be more functional.

Later in the progression, you can see much higher frequency variations. Like periods of disorientation and confusion interspersed with periods of lucidity all within a single sitting or conversation.

In those periods of greater lucidity, recall of the past can be more accurate. General listening comprehension, speaking, and logical thought also seem more normal.

Edit to add: I sometimes wonder if the belief in terminal lucidity is one of those logical fallacies which support lots of superstitions. Are we just fixating on the final wave in this chaotic wave train, and forgetting all the other waves that happened before it? Or is it that more caretakers are engaged and observing these waves towards the end, e.g. because the patient is known to be in the terminal phase..?

It made sense that interrupt in Emacs could get into a controlled state of receiving the next command. It's a little bit like the SAK (secure attention key) concept, as seen with Windows use of ctrl-alt-del.

Edit: Ironically, as a long-term emacs user, I don't really remember any commands that start with ctrl-c! For me, the most common sequences start with ctrl-X or meta-X. Or the prefix search commands ctrl-S and ctrl-R.

1. Use a password-based encryption method (not tied to hardware identity) if you prioritize moving the disk around. Then it is just as readable in a spare machine.

2. Use an easy to remember password/passphrase and write it down somewhere you keep paper documents, if you prioritize recovery.

This still provides meaningful protection when you need to discard the drive. The random downstream recipient of the hardware will not know your password, even if you skip the step where you "crypto-shred" the drive by setting a new random password.

And people who get a sense of accomplishment from hitting the jackpot on a slot machine.

Operating an LLM is a strange combination of the two.

In theory, they could be made to strictly compartmentalize their "memories" and exercise some kind of robot-client privilege. There could even be built-in, task-appropriate data-retention limits and anonymization algorithms to reduce the risk of leaks.

But instead, we should assume they will be made to remember too much and to send abusive levels of reconnaissance and other telemetry back to their parent companies. Because anything less is "leaving money on the table" and that has become the greatest sin of all... :-(

In other words, when will we really see a transition from "yet another token generator" to something that appears to coherently observe, perceive, form intent, plan, and act in a way that is compatible with an existing, long-running human context?

(And, also, do this with enough determinism to be a viable product and not some gaping liability...)

I realize there is a full, multidimensional continuum here.

On one end are directly-aimed weapons that do their damage while still being aimed by the operator. Their risks include collateral damage limited to things like aiming errors, effect radius, or continuing down-range beyond the target.

Further out are messy things with more active guidance that can turn and seek the target and potentially go off course. But their time to target is still quite limited and more or less being observed by the one who fired it. The risk expands with its potential "cone of maneuvering" and travel range.

Then you get into these things with long dwell times and autonomy where the eventual targeting event happens without supervision and is greatly affected by things happening in the environment which the operator cannot have really predicted nor controlled for. The longer time in operation increases the risk not only from wandering/guidance but from how much the environment can change before it performs its final targeting event.

Another example in this category could be chemical and biological weapons. There is a lot more uncertainty in the targeting effects due to the way it disperses in the environment.

These newer drones have just gained locomotion instead of having to wait for victims to come to them.

Just like people trying to time the market with "technical analysis". It is extremely easy to find whatever pattern you are looking for. A lot harder to accurately distinguish predictive power from fantasy.

Meanwhile, like I think you suggest, I would assume everyone can generate similar outputs themselves. The idea that you can claim priority on your dream prompt and lock up the market on prompt responses sounds delusional to me. It's not novel invention when you're spit-balling at the same level of abstraction as every fantasy/scifi writer who ever was.

So I also have doubts about the sustainable business model. How long will it take for this fantasy to unravel, as people discover they cannot monetize their AI outputs as much as they dreamed, and in turn cannot afford to pay the AI services they use?

My absolute nightmare is that this becomes a "too big to fail" thing and oppressive/fascist governments decide to back full regulatory capture. That instead of letting it unwind, they grant and support enforcement of an increasingly absurd and arbitrary copyright/patent regime to support this monetization scheme.

It sounds more like you are saying YIMBY but imagining MIMBY (maybe in my ...) where you've just replaced those other guys making the decisions with your own cabal?

There actually used to be this "empty suburb" feeling at many times per day when a typical bedroom community had sent its kids to school and parent to work. Particularly when they were not wealthy enough to have paid laborers around doing things during their work day. If anything, they got busier since COVID as people have more varied schedules.

Then there is the new but incomplete development, e.g. with graded lots and some subset of streets and walkways. If work is suspended for some reason, it may be decorated with idled earth movers, piles of building materials, or partial foundations or framing.

Or it might turn into the next type, which is an aborted subdivision build or after a severe wildfire, which is basically a moonscape of graded lots with no buildings nor vegetation.

Then there are the abandoned neighborhoods that were once vibrant. Old, decrepit buildings, and wild vegetation, e.g. around dead industrial towns.

In unified memory, all the memory is host memory and data can go from program to GPU with zero copy movements. The addresses of buffers can be shared via appropriate MMU translation support, so that the application and graphics subsystem are communicating effectively through the basic RAM cache coherency protocols over the same buffers.

Edit to add: Aside from the zero copy transfer potential, it also means dynamic allocation strategies can shift the balance between host and graphics allocations on the fly. Individual image and message buffers can be allocated on the fly instead of setting a static split between the two worlds.

That is one place where an analytical solution is a benefit, even if it is a bit less realistic. You just have a position(t) parametric function you can evaluate when rendering sporadically.