Yeah, people tend to go crazy with over-rationing when lost. Like people who die of dehydration in the desert who are found with water still in their water bottles because they over-rationed it.

Also, most people, stranded without cell signal, don't just leave their cell phone on futilely trying to contact a cell tower until the battery runs out. They shut it down and then turn it on at regular intervals to see if they then can get a signal (at a later time or different location)

Assuming the lost person is carrying a cell phone that's turned on and hasn't already drained its battery. As I understand it, a cell phone battery will drain faster when there is no cell service since the phone will increase the transmitted power in order to try to make a connection.

This is why I have a Steam Deck. I literally can carry it very easily and plug it into my 55" OLED HDR TV. Connect N Bluetooth controllers and off you go.

Word don't exist in a vacuum. Words are a product of the universe that leads to their creation. To get good at predicting words, you have to get good at modeling said underlying universe.

Q* (A* Q-learning, aka iterative tackling of possible routes to solve a problem via perceived distance from the solution)?

Massively MoE, with cos distance between hidden states feeding back to provide the model with an assessment of its own confidence?

More "distributed" models where portions are more decoupled from others and don't need as much bandwidth, to make it easier to train and run on cheaper hardware?

Will we finally see a lower O() context memory model? Could we at maybe close in on O(logN)?

Non-Transformers based (Mamba, etc)? Doubt it, but possible.

PCN-based with neuromorphic hardware? I *really* doubt that, but it would be an utter game changer.

Will surely be highly multimodal from the start. I wonder what the full modality list will be, though...

Sure, it "looks" better. But it's inaccurate in delayed and that's why nobody who is competitive uses it.

Reverse engineering detection is also pushed as a legit use for circumventing privacy

East Germany by itself wasn't able even in principle to sustain a costly industry that depends for its survival on narrow specializations in many areas and quantity shipments

It was MUCH simpler to be a semiconductor manufacturer in that era, and overall global volumes were much smaller.

Besides, the real goal there, and elsewhere in the Eastern Bloc, was always one - supplying the military, so commercialization was never the important driver.

This is simply false. East Germany sought to be a leading global commercial supplier.

Finally, judging the effort as "total failure" is a bit of a stretch

Bankrupting the nation to produce laughably small quantities of inferior chips absolutely IS as bad as a failure can get.

Dresden today is the center of the so-called "Silicon Saxony"

Yes. Today. You may perchance notice that your maps no longer show an East Germany there which led this multidecadal mission to turn around the disaster that was the East German semiconductor push.

This has nothing whatsoever to do with "training", or even the AI. This AI is a RAG summarization model. It's only tasked to summarize the text presented to it. Google is passing it the top search results. It's accurately summarizing them. It's not tasked with evaluating them.

The AI is doing its job. The problem is the stupid task Google set for it: "summarize the top Google results together without any attempt to evaluate them"

I've seen maybe a hundred screenshots of "TeH gOoGlE aI sCrEwEd Up!!!" since the feature was released, and there were like maybe two that were the AI's fault (one with a math error, one with oversummarization losing context - both being fixable). The rest was just the stupid idea of the Google execs to do this in the first place.

Just to add to this, most rental car agencies run small satellite locations the size of a gas station in tight urban locations. Others satellites are located on the premises of a single big corporate customer. Having the option to skip or delay building infrastructure in these locations is potentially very economical.

Or, it could go down the East Germany route, who also decided it was going to become a global powerhouse in semiconductors; poured tons of money into it and employed the best industrial espionage that the world could offer to get ahold of entire foreign production systems...

. ... and ended up basically bankrupting the nation, because they just couldn't deliver on the strict chains of quality needed at every step of the long production chains, and could neither produce nor innovate fast enough to keep up with foreign peers.

How do you propose we deal with a race-to-the-bottom food industry determined to feed us shit? Yes, as individuals we can start seeing away from foods with added sugars and the like, but how do you improve nutrition systematically?

> Nobody cares the protocol, as long as the plug is standard.

Everyone cares about protocol, because if the plug is standard but the protocol isn't, then you'll plug in and it won't work.

But since the protocols are the standard, you can at least get an adapter if the plugs are different.

> Yeah, old model S have an incompatibility ahead

If by "ahead" you mean they've had that problem for several years now, sure.
=Smidge=

> Nope. 3 million people recharge at home today.

3 million people have gasoline pumps at their house? (I think you missed the context...)
=Smidge=