In what way? Once the model is loaded into VRAM, very little data has to travel over the bus. You're going to be limited by storage speeds loading the model anyway.

You're probably not going to get any external PCIe port other than thunderbolt. Thunderbolt is fast enough for most Mac use cases anyway. TB5 gets you 80 Gbps bidirectional, or 120/40 Gbps asymmetrical, and there aren't a lot of things in a desktop environment that would really benefit from more than 120 Gbps.

I'm confused, most stores like that already have digital price tags, and have for many years. Even Walmart has been using them already for a while now.

Cuba is bullied because the Cuban community in Miami is numerous and have a loud opinion to doing so. Florida is generally quite important in winning a presidential election.

Amazon has been making a lot of amateur mistakes too in recent months. I expect it will get worse before it gets better.

Hydro-Quebec is facing future capacity shortages due to past underinvestment in new infrastructure. They're facing losing a significant chunk of their installed capacity due to the Churchill Falls deal being at risk of falling apart. That deal would also have included a bunch of new construction, which may not happen any more. They're making big plans about future investments to increase capacity, but are still returning billions to the government in dividends instead of re-investing in capacity expansion.

Do we really have the capacity to export so much more power to the US?

THIS!

Been in the Capital Expenditure Management business for a while Iv'e watched the Southern California office market completely retract and try to reinvent itself into usually residential with little success. I can only imagine all the brokers baiting property owners with data centers, not realizing the practicality of upgrading power, water and fire life safety to data center requirements is impossible for anything less than $300/SF

They did not use a rasberry pi 4, they used an RP2350 microcontroller. If memory serves, there's basically static recompilation happening, where Linden writes SueprFX assembly, which is being translated to ARM assembly at compile time.

I disagree completely with the premise. Prices don't necessarily have to increase. We have yet to reach maturity of chips optimized for inference, in addition to the regular old factor of computing work per dollar going up over time. Moore's law might be "slowing down" but we aren't at the end of the road yet. Keeping today's features running will cost less to deliver in both infrastructure and electricity in the future.

What will more likely happen is that features and functionality will keep expanding to use more processing power. But where is the limit? I say that comes when we can render at near-realtime 8k/240hz a video (or video game if you prefer) with procedurally generated world, characters, and storyline based on the users input via whatever real-world data, UI or sensors you want to use. This might even be possible now if you are a billionaire with access to million dollar server farms. Probably my imagination isn't broad enough in estimating the limit of "personal computing" but additional computing power beyond that seems pointless for any one individual.

In any case the price of an ai product will depend on the features offered and how much hardware is needed. Probably you can run a 500 billion parameter LLM on a smart watch in 2055 but if you want that power today it seems to cost about $20 a month. I doubt anyone will try to ever charge more for today's $20 featureset. The price of this stuff will absolutely decrease, the only unknown is how companies will roll out new functionality and how specific future features fit into the pricing tiers over time.

It's somewhat understandable though. The $5 contributors waited this long, they can wait a little longer for things to be fixed. If that is possible. Ultimately though they may need to verify identity or at least a domestic phone number within the user's region, which may not be what the users were looking for.

Your greybeard has such a crust that the crustie bits have their own wizard beards.

Then make it optional. Let me set the location name language versus the spoken language, either just as a global setting, or on a territory by territory basis. Anybody who lives in a city where the language spoken is different from the language they speak is affected (like an English-speaking person living in Montreal), but also any tourist who asks anybody for directions is going to hear street names that are completely different than Google Maps.

As an English-speaking Quebecker, I don't recognize many of the street names that Google says out loud, because some of them sound completely different than what I expect. I could set the whole language to French, but... French is not my native language, so why should I? I just want the street names to be pronounced correctly.

Do you think we can get someone to make a rom chip containing a private key, securely?

Using Google Maps (or any mapping app) in Montreal is a constant facepalm where it tries to read French names as if they were English words. There are tons of situations in the world where the native language of the driver is not the language of most street names. Is it too much to ask for it to know "I'm speaking in English but I should read all street signs in this territory in French"? There's lots of situations where you might have similar situations (tourists, expats, cities with multi-lingual populations), and using unrecognizable names for streets is a big pain.

And that cuts both ways. The girl's school that the US bombed used to be part of a military base, but some years ago was separated and walked off. It should have been taken off the target list but with thousands of targets in the database nobody is really checking each one for updates.