md_

I'm confused. The article from September 1 linked to here is strangely future-tense ("But the firm’s latest investment in 10,000 of the company’s H100 GPUs dwarfs the power of this supercomputer....This AI cluster, worth more than $300 million, will offer a peak performance...").

It links to a Tom's Hardware article (https://www.tomshardware.com/news/teslas-dollar300-million-a...) from August 28 that says "Tesla is about to flip the switch on its new AI cluster, featuring 10,000 Nvidia H100 compute GPUs") and says "Tesla is set to launch its highly-anticipated supercomputer on Monday..." (presumably the September 1 event).

So, like, does Tesla actually have 10k H100s? Or do they have an order for 10k H100s? Or an intention to buy 10k H100s?

Is the sole source for these articles this (https://twitter.com/SawyerMerritt/status/1696011140508045660) random Twitter post by some guy who runs an online clothing company?

I don't mean to snipe, but this article doesn't seem to rise to the extremely high editorial standards of such tech-press luminaries as "TechRadar" and "Hacker News".

> What makes you think that Telsa, a company with far less AI workers and knowledge, an far less money than the above companies can out design and out build them?

Presumably because Elon himself will be involved in the design, and Elon, as we all know, is one of the world's great thinkers. ;)

The really obvious thing, if you watch the Freeze "best of" clip, is that most people have terrible running form. In most cases, I suspect I could tell who has a chance and who doesn't simply by looking at one stride from each competitor.

There are so many factors that go into biomechanics--I struggle to improve my own running form--but it's interesting to me that it comes together in such an obvious-at-a-glance factor as "looks fast."

What I understand the author to be saying is that the SE will unwrap locally stored synced credentials and then encrypt them to the previously-established iCloud sync key.

It doesn't really matter if the storage of the actual wrapped keys is in the secure element or on disk, of course--what matters is if the SE exports unencrypted credentials, or if it fails to validate the identity of the key to which it encrypts credentials before exporting.

It doesn't seem from that description like it does either, but it's a bit unclear to me from the docs. Do you know?

I'm not sure how to relate that to the slashid article, which notes,

> In other words, the private key used to encrypt kSecAttrSynchronizable keys in the Secure Enclave is backed in iCloud. As such, when a new device needs to restore the keychain it can reconstruct the private key and thus decrypt the keypair needed to access all the private keys marked as kSecAttrSynchronizable, which include passkeys.

So presumably the iCloud sync pubkey is in fact available to the TEE to do the necessary encryption before exporting the synced credentials. Where that pubkey is stored only matters if you can trick the TEE into encrypting to some other pubkey, which presumably you can't--the implication of the docs seems to be that it pre-establishes which keys it will trust at iCloud enrollment time.

Is that not the case? I think if the TEE happily encrypts to any pubkey, there's a super obvious bypass here, so I doubt that's how the system works.

I don't think the distinction is meaningless, though, as I said, I do agree that exportable keys are more vulnerable than non-exportable.

What I understand the iCloud scheme to guarantee is roughly:

1. The TEE will not export keys except encrypted to the previously-established sync keypair. (IOW, it will only export keys after first encrypting them with the public key of the iCloud-sync keypair.)

2. The iCloud sync keypair will not be divulged by the iCloud server-side HSM except upon presentation of the correct auth credentials (your iCloud password or recovery key).

I think this provides meaningful security beyond keys sitting in plaintext on (FDE-encrypted) disk, namely, that malware running on the device cannot itself simply export the keys from the TEE. It must instead join a new keychain to the iCloud trust circle, as described at https://support.apple.com/en-hk/guide/security/sec0a319b35f/..., and to do this, it must be able to authenticate to iCloud as the user.

That's certainly possible--malware running on the device could phish the user's iCloud password, for example. But it's no longer a zero-user-interaction access to the synced key.

Again, clearly weaker than non-exportable keys, but still stronger than keys sitting in plaintext on disk.

I don't think most services are requiring specific vendor attestation for Passkeys. I think most services are requiring authenticators that support user verification. Does your USB security key support uv?

E.g. https://www.yubico.com/products/yubikey-bio-series/

If someone has code execution in your browser, they can just steal your authentication cookies. They don't need your Passkey.

It's true that unsynced/hardware-bound credentials can help mitigate such attacks, but they cannot prevent it. The primary threat model of Passkeys is password theft via phishing, reuse, server-side compromise, etc.

If you want to use a USB security key, you still can--Passkeys do not take away from that. But most people are unwilling to go buy a USB key or carry one around, which is where Passkeys--arguably a bit less secure, but far better than the passwords they replace--are still a great step forward.

I think you're the one who misunderstands.

Synced Passkeys are, in the Apple world, still stored in a Secure Element; they just are exportable. https://www.slashid.dev/blog/passkeys-deepdive/ describes this fairly clearly.

Syncing does certainly open up attacks that are not possible against non-synced credentials, but you're incorrect to say that synced Passkeys are not stored in the Secure Element.

At least, that's my understanding. Tell me if I'm wrong!

> And, if things really are completely collapsing, nobody wants gold. They'll want something much more immediately useful--food, clothing, ammunition, etc.

I'm not sure I buy that. Barter for an intermediary is fundamentally useful. In a case of societal collapse, people with food/clothing/ammunition--or people who can provide useful services (like driving me to the nearest border) aren't going to want more food/clothing/ammunition--they're going to want something fungible that they can trade for what they need!

So, sure, there's probably some rare scenario where gold (ideally in small, easily verifiable discrete units, not big bars) is useful.

I didn't find it sensationalist. Obviously, like, being born in a country with a poor healthcare system might have a bigger impact on your morbidity/mortality in some cases than dietary changes, but it's hard for people to change the country they are born in.

Huh? Isn't the entire point of matched controls in an epidemiological study to isolate the effect of that single variable?

There's some research that suggests group alcohol consumption increases social cohesion. I can't remember the studies I originally read which suggested this, but a few from a web search:

https://www.psychologicalscience.org/news/releases/moderate-...

https://www.sciencedaily.com/releases/2012/06/120629211854.h...

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5462438/

Whether other psychoactive substances have similar effects I don't know, but it seems reasonable to assume alcohol's somewhat-unique social role reflects its actual impact on group bonds.

I was struck while recently reading David Grann's "The Wager" by how little people understood scurvy at the time. Grann writes about how some doctors believed it to be a reaction of the body to being away from solid land for too long, and attempted a treatment by burying patients up to their necks in soil.

In the shadow of encroaching regulatory despotism, the luminous innovation of the iPhone was threatened with bureaucratic shackles by the European Union. The relentless march of progress was halted, as the central planners decreed that every charging port must bow before the altar of uniformity. No longer could Apple's vision for a sleek and efficient Lightning connector reign supreme; instead, the heavy hand of Brussels demanded compliance with the USB-C standard. The champions of individualism and choice saw their liberty erode, replaced by the stifling straitjacket of conformity. The spirit of innovation, once ignited by the entrepreneurial genius of Silicon Valley, flickered in the face of such top-down directives, leaving a world dimmed by uniformity, where the art of technological diversity was sacrificed at the altar of bureaucratic convenience.

Or, like, you know...it's just a charging cable.

A core argument the post makes is that TPMs are insufficient for verifying full stack integrity and thus ineffective for FDE. (Eg by exploiting vulnerable drivers, an attacker can dump the disk encryption key from kernel memory.)

But in such a scenario, an attacker can also use such an attack to bypass any remote attestation/DRM/etc!

I guess you could argue that such attacks are too much work for consumers, and that low fences control big dumb animals…but I think, fundamentally, the same argument applies to consumer security functions like FDE!

Tl;dr: I think it’s hard to argue that TPMs are both useless for practical user security and a threat to free computing. It’s gotta be one or the other!

I suspect the number of people who believe in total unilateral disarmament and military aid to Ukraine--yes, I agree that's an internally inconsistent point of view--is vanishingly small, and not, as @yinser snarkily wrote, "a large block of US citizens."

No doubt there are a far larger number who find weapons development distasteful or would not personally want to work on weapons, but nonetheless recognize it as an unfortunate but pragmatic necessity.

I suppose, if one were given to inane Internet snark, one might say, "Haha, you refuse to become a proctologist, yet you go for your colonoscopy. WHICH WAY WESTERN MAN?" But, frankly, that would make you look stupid, just like @yinser does.

I'm struggling to understand your point.

Is it that someone who a) believes the best path to peace and stability in the region is a Ukrainian defeat of Russia and b) also opposes Donald Trump's, say, tax policy has a "dilemma" in whether to support military aid that goes to Anduril (and thus, ultimately, potentially to Trump's re-election)?

If so, this seems like a pretty facile point.

Like, life is full of pros and cons. I want to leave my umbrella at home so my bag is lighter, but I want to have it with me if it rains. WHICH WAY, WESTERN MAN!?!

In a democracy, you sometimes have to do business with people with whom you disagree. You also might try to avoid doing business with them, all else being equal. WHICH WAY?!

I dunno. I feel like you thought you had some sort of deep philosophical point, but I'm not getting it I guess?

> So people continue to predict doomsday and most of the time it doesn’t happen.

Well, the same is true for any prediction of anything unlikely. People continue to predict world-changing innovations, and most of the time they are wrong. That's the reality of any low-probability/high-impact prediction, in the positive or negative direction.

As a specific case of that: most startups fail, so predicting a startup will fail is safe.

I'm pretty prone to exactly that prediction. But I like this quote from Erik Davis (https://www.google.ch/books/edition/High_Weirdness/Rcq2DwAAQ...): "In the court of the mind, skepticism makes a great grand vizier, but a lousy lord."

Amusingly, Kevin Roose and Casey Newton made a ChatGPT-authored cocktail recipe on "Hard Fork" a while back: https://www.nytimes.com/2023/06/30/podcasts/hard-fork-ai-poi....

I think the review was basically, "I’ve had worse cocktails. I’ll be honest."

So...meh?

In contrast to the parent post:

When I was an undergrad, I interviewed at a then-very small, now-very-prestigious quant firm. In the last interview or so, I asked the quant head something like, “but what value do you provide to society?”

He said, “we provide liquidity.”

I turned down the job they later offered and went to work at a software company that sells a thing people buy.

Now I'm jealous. I've been working on croissants off and on for a while--off, right now, because of the hot weather--but frankly was never happy with my results compared to a decent boulangerie. Whereas I do feel my baguettes and pizzas are both decently competitive.

Croissants are damn hard.

Dunno why nobody has said anything like this, but I make some awesome bread (baguettes, mainly) and pizza.

What's great is not only that you get to eat the results, but whether it turns out well or badly, you get to repeat it all over again--and again, and again, and again.

Half agree.

Successful founders get to remain CEO (or CTO, or whatever). Unsuccessful ones don't: their investors push them out, or the company tanks.

Along the way to obtaining any position, there's a lot of partially correlated signals you have to send, and they're different between different paths. To become CEO at an established company, you might have to show prior experience, know how to wear a suit and tie, etc.

To become CEO of your own startup, you have to show investors or customers something else, but I'm not sure those signals are less bullshit: you might find that wearing shorts or playing League of Legends during investor calls somehow convinces people you're a super genius.

Basically, there's always an element of bullshit, because we're humans, and our judgment is tempered with all sorts of irrational biases.

I don't think that means that an essay on "what it takes to do $job" is inherently bad, though.

> Why launch such a teaser without announcing the (actual) goal ?

We're talking about it, right?

Musk has all the money in the world, but it seems to me that what he really wants, if he cannot have your love and respect, is your attention.

My money is on shipping the OpenAI API, wrapped with a prompt like,

"Below is a conversation between an intelligent AI assistant which deeply understands the true nature of the universe, and a human."

Wasn't X.ai a calendar-scheduling assistant back when non-LLM-based chatbots were, like, a hot thing in the Valley?

There's some fucked up irony here that their domain name probably ended up being their most valuable piece of IP, and it gets bought up by Elon Musk to tout his...whatever the fuck this is.

You seem really emotionally invested in this subject, even though it's obviously not one you know much about.

Good luck with that.

> I am sure you are not genuinely claiming that it's impossible for them to be influenced by PR firms or external marketing spend? That you sincerely believe they are never influenced by external marketing of people? How might you suppose they find people to write profiles on? I personally know that it isn't chance.

"PR people pitch profiles to writers" is very different from...oh, what did you write? "From everything I personally know about the media, articles like that are usually paid for..."

The former is normal journalism. The latter is an ethical breach. So, yes, let's talk about tone for a moment: in your post, you accused multiple people of ethical or legal breaches, on the basis of, oh...your feelings?

Very respectful of you! Yes, my "tone" is direct. But I didn't accuse you of anything unethical, did I?

> I do use my real name for whatever it's worth, I have absolutely no problem wearing these statements on my sleeve

Hahah, OK then.

> I'm really happy you took the the time to respond to me though

Oh, but I did. As far as I can tell, you made one, and only one, technical claim: that Signal's reproducible builds aren't useful because they don't cover the NDK code. I asked you, twice now,

"As I said, in which codebase does the key verification and encryption live--is it in the native code, or in the Java code? I suspect you don't know, or you don't know why that matters."

You have twice now failed to reply to that. Care to try for a third time?

Edit: Seems like Signal's reproducible builds now do cover native code, though I haven't tried this myself: https://github.com/signalapp/Signal-Android/blob/main/reprod....

So, like...what was your point again? Other than that you don't know what you're talking about?

No, I wasn't talking about Telegram. I was making a general statement about crappy e2ee startups with bad designs and accountability. There was a recent one using Sealed.io whose name I forget.

Re Telegram, sure, use what you like. https://www.pwnallthethings.com/p/russia-is-spying-on-telegr...