As others have noted, the article is analysing the actual financial markets angle.
For my two cents on the technical side, it is likely that any Western-origin shakiness will come from Apple and how it manages to land the Gemini deal and Apple Intelligence v2. There is an astounding amount of edge inference sitting in people’s phones and laptops that only slightly got cracked open with Apple Intelligence.
Data centre buildouts will get corrected when the numbers come in from Apple: how large of a share in tokens used by the average consumer can be fulfilled with lightweight models and Google searches of the open internet. This will serve as a guiding principle for any future buildout and heavyweight inference cards that Nvidia is supplying. The 2-5 year moat top providers have with the largest models will get chomped at by the leisure/hobby/educational use cases that lightweight models capably handle. Small language and visual models are already amazing. The next crack will appear when the past gen cards (if they survive the around the clock operation) get bought up by second hand operators that can provide capable inference of even current gen models.
If past knowledge of DC operators holds (e.g. Google and its aging TPUs that still get use), the providers with the resources to buy new space for newer gens will accumulate the amount of hardware, but the providers who need to continuously shave off the financial hit that comes with using less efficient older cards.
I’m excited to see future blogs about hardware geeks buying used inference stacks and repurposing them for home use :)
>when the numbers come in from Apple: how large of a share in tokens used by the average consumer can be fulfilled with lightweight models and Google searches of the open internet
is there any reason to expect that this information will ever be known outside of apple?
Accuracy wise we won't know the exact numbers, but insiders and industry experts usually are able to find ballpark figures that they share with the press. The alternative is the usual find out the estimates through competitors' lost MAU numbers in apps like ChatGPT for iOS.
I have never before felt pressured about what I can or cannot protest about in Europe by China, but I can’t say the same about our most powerful ally, who has threatened every sector of our society – political or non-political – with consequences if we do not act and speak as they do. China absolutely does not care about our society the same way as that.
I respectfully disagree. Even if you told me that you felt like working at the UN was a waste of time, I’d still tell you that at least you contributed to a historically unique global institution which at least strives to bring people across the world together.
Not a lot of people here work for Meta, which is why you had to lump in "Random SaaS" like that's remotely comparable. I doubt most people here are working on anything harmful, let alone a fraction of what Meta does.
Unless you think Todo list apps cause Ethnic Cleansing.
>Cooperate with their fellow citizens and fix their banking system
That's exactly what Bitcoin was for.
I don't know where you live, but most places it's the banking system which fixes the citizens, not the other way around.
Only way I imagine being able to fix a bank is with a brick through the window. It's why one level up from where I'm at banks are closing their physical locations, keep no cash, and only have people for 1-2 hours in the morning when only the most obedient are awake and not occupied.
Definitely. The fact that they inject it into Google Search means that even fewer people who have never used ChatGPT or just used it as a "smarter" Google search will just directly try the search function. It is terrible for actually detailed information i.e. debugging errors, but for summarizing basic searches that would have taken 2-3 clicks on the results is handled directly after the search. I feel bad for the website hosts who actually want visitors instead of visibility.
As a Finn, rather than bore you with a 2846 bullet point list, I'd say that technologically not a lot, but we do have a lot more to lose, so it is easier to bargain with our industry compared to Sweden's. Our population is not always big enough to compete head-to-head with some sectors Sweden is also a part of.
They'll perhaps never be faster due to weight limitations from the battery. Gasoline is just so light compared to batteries, and the car becomes lighter as the lap goes further and gasoline is used.
It's the KERS (kinetic energy recovery system). It's battery power that's collected from braking (perhaps some other additional sources) that can be released by the driver when they choose. Similar to F1 car systems in the early 2010s.
In this current geopolitical climate, this move is a disaster for anyone who does not fully trust the US or other nation-states with data surveillance.
That problem is solved, but unfortunately the bottleneck has been for the longest time the larger, more complex electrical equipment that is used to connect chargers to the grid. Companies like the Finland-based Kempower produce the best charging equipment on the market, but their problems start where their equipment ends: the grid.
For my two cents on the technical side, it is likely that any Western-origin shakiness will come from Apple and how it manages to land the Gemini deal and Apple Intelligence v2. There is an astounding amount of edge inference sitting in people’s phones and laptops that only slightly got cracked open with Apple Intelligence.
Data centre buildouts will get corrected when the numbers come in from Apple: how large of a share in tokens used by the average consumer can be fulfilled with lightweight models and Google searches of the open internet. This will serve as a guiding principle for any future buildout and heavyweight inference cards that Nvidia is supplying. The 2-5 year moat top providers have with the largest models will get chomped at by the leisure/hobby/educational use cases that lightweight models capably handle. Small language and visual models are already amazing. The next crack will appear when the past gen cards (if they survive the around the clock operation) get bought up by second hand operators that can provide capable inference of even current gen models.
If past knowledge of DC operators holds (e.g. Google and its aging TPUs that still get use), the providers with the resources to buy new space for newer gens will accumulate the amount of hardware, but the providers who need to continuously shave off the financial hit that comes with using less efficient older cards.
I’m excited to see future blogs about hardware geeks buying used inference stacks and repurposing them for home use :)
reply