It does not. The $1 of electricity can be simplified to just a $1
However, the way the parent worded it ie $1 of compute electricity - requires the compute to exist and temporarily be loaned. Otherwise “just a $1” would require a massive capex to buy/build compute.
Self driving cars have existed for at least a year now. It only took a decade of “1 years away” but it exists now, and will likely require another decade of scaling up the hardware.
I think AGI is going to follow a similar trend. A decade of being “1 years away”. Meanwhile, unlike self driving the industry is preemptively solving the scaling up of hardware concurrently.
Having both worked at Amazon and still in the cloud industry, to me this seems like a strange take.
Apple leases these computers from Amazon like it would from any other colo. Why wouldn’t these servers be considered Apple servers?
Barring a major privacy violation by AWS (which doesn’t seem likely), or some other sort of 0-day hack the data on these servers is entirely private to Apple.
This means trusting AWS’s privacy policies too, which users aren’t always aware of or expecting when Apple promises full control. It’s about honest transparency.
We pay a premium for Apple's privacy restrictions and then end up being tricked to rely on Amazon's. It is false advertisement to say the least....
When Apple says your data doesn't leave their servers, that doesn't mean those servers have to be in their own datacenters or that Apple doesn't have other vendors that help them deliver their service. It also doesn't mean those companies have access to your unencrypted data. That data also, by necessity, likely traverses other networks in encrypted form on its way between you and Apple.
Apple claims that data they are in full control of your data from Safari, Maps, and Spotlight etc.... Data center or not, this contradicts both the spirit and letter of that promise.
If you have any evidence that Amazon is accessing the data of their leased tenants, that would be an earth shattering indictment of the entire cloud industry
It would be akin to accusing Datapipe or any other provider of pulling drives out of any client with racked servers in their data centers.
And that’s exactly why high security applications don’t use cloud like this; because that can never be guaranteed by some policy compliance certificate. That’s the only thing stopping them from pulling drives, etc.
If Apple is leasing physical servers from Amazon, without Amazon further involved in running them (other than dealing with hardware issues), then this argument holds water. Otherwise it doesn’t, even if they’re using Amazon VMs or some higher level services. A VM hosted on someone else’s physical platform is not an “own servers” in this context (which is who has access to the data).
I know HN is very imbued with the cloud approach, and maybe from that perspective running your own servers is just so unthinkable it may as well not exist at all, you don’t get to change how language works. If someone says they’re running on “their own servers” that always means the whole stack including physical and up.
Ownership is determined not only by who pays for it, but also who has direct access to the actual devices.
Yes-ish. When you're making broad "trust us" claims to consumers - who don't know the industry or its practices, let alone the technical details - then the really honest approach is to follow those consumer's understanding of your promises.
Otherwise, they might end up feeling that they were duped by the weasel-words of a sleazy lawyer.
Maybe some might feel that way at first, but it’s also an opportunity and responsibility to educate.
This problem is why enterprise contractual agreements and large compliance systems exist for companies at this scale. Large hosting providers like AWS, Azure, GCP, etc. provide an ability to scale and assurances about risk mitigation, privacy, and availability that are much more viable than each company having to maintain their own private in-house fleets just to create an additional illusion of privacy/security that’s actually no better than tight contractual controls to begin with.
Maybe they need to explain this properly, but servers don’t magically have a lower level of risk just because they’re behind your four walls. In fact, if you lack the experience and expertise, the risk is almost certainly higher depending on your threat model. (And for Apple, their threat model is at the nationstate level. They don’t choose their hosting providers lightly.)
I think one of the most important lessons in life (even as a healthy person) is realizing “motivation and action are cyclically causal”
Of course “action follows motivation” but even when not motivated “motivation follows action”.
For example, even as a healthy person I am not always motivated to go to the gym after a busy day at work which I am “so tired from”. I go dispite the lack of motivation. Unsurprisingly, I walk out of the gym feeling re-motivated and “with more energy”.
Acting with neither intrinsic nor extrinsic motivation is technically impossible, no? Otherwise, i wonder how this mysterious third force compelling you into a gym absent motivation relates to your personal psychology / environment, and how executive dysfunction (both genetic, and technology-induced) fits into the picture.
You can take advantage of this phenomenon to snowball small actions into bigger ones, too. To get into the habit of working out, I started going to the gym everyday. I didn't work out everyday, but the act of going out of my way to be at the gym lent me the motivation to actually work out more often than not.
I now get to the gym (or some form of exercise) 6 days a week. That was entirely because I made the decision to go to the gym and watch some YouTube.
Then I’d end up staying 90 mins but I’d get my 50 min workout in with a lot of long breaks! Then things started becoming a habit but I still have many days where I just watch YouTube at the gym lol
I turn a lot of real photos into cartoons. I love the feature!
Most recently I took a photo of my grandma and me, asked Gemini to make it a cartoon, asked Gemini to make the new variant into a birthday card.
My grandma loved it! I was happy to make her something custom. Buying people cards just never felt right to me. Writing was also never my strongest suit - so this new form of expression for me has been enhancing :)
The only remaining thing I need to do better is getting the card printed! I wish that also was only 12seconds of work.
FWIW, I think you might be better off with immutable rows and lamport clocks.
Everything is a full new row because it is “a message” including read receipts. Some messages like read receipts just don’t render in the chat.
Edits can work the same way by rendering over a previous message, even though the local and remote DB have multiple rows for the original and edited messages.
A month of development can easily save a day of research.
I was very surprised (or perhaps disappointed is a better word) when I didn’t see Lamport, paxos or raft mentioned at all. At least crdts made an appearance, although almost in the post scriptum.
What you’ve described is the foundation of Lucene and as such the foundation of Elastic Search.
FSTs are “expensive” to re-optimize and so it’s typically done “without writes”. So the database would need some workaround for that low write throughput.
To save you the time thinking about it: The only extra parts you’re missing are what Lucene calls segments and merge operations. Those decisions obviously have some tradeoffs (in Lucene’s case the tradeoff is CRUD).
There are easily another 100 ways to be creative in these tradeoffs depending on your specific need. However, I wouldn’t be surprised if the super majority of databases’ indexing implementations are roughly similar.
At least in theory. If the model is the same, the embeddings can be reused by the model rather than recomputing them.
I believe this is what they mean.
In practice, how fast will the model change (including tokenizer)? how fast will the vector db be fully backfilled to match the model version?
That would be the “cache hit rate” of sorts and how much it helps likely depends on some of those variables for your specific corpus and query volumes.
This can’t be what they mean. Even if this was somehow possible, Embeddings lose information and are not reversible, I.e embeddings do not magically compress actual text into a vector in a way that a model can implicitly recover the source text from the vector.
LLMs can’t take embeddings (unless I’m really confused). Even if it could take embeddings, the embeddings would have lost all word sequence and structure (wouldn’t make sense to the LLM).
Those are two reasons why C is less tedious than assembly.