The US is still fixable if young people will just fucking vote. Irreparable damage has been done, which means the old structures are gone, but they can be replaced with new ones -- a hypothetical Democratic president would need to clean house top to bottom. Obama's greatest flaw was that he didn't see the need to do that after Shrub.
I don't fall in the ring category, but I will be voting Democrat for the first time. That's how bad this is getting for me. I have real reasons to hate the Democrats as a party, but I'm scared of them less than the modern Republicans .
In several states we have moved well beyond "fixable". For instance, in Texas, where they are outright saying that they are redrawing districts with the explicit goal of preventing Democrats from being represented. Not only that, they're saying they campaigned on the promise and are therefore obligated by their constituents to institute a single party state through the denial of representation of their political opponents.
I genuinely don't believe the US can be fixed anymore. We're on a slip-and-slide and far too much damage has been done to stop it. There aren't even adults left at the table to try.
That speed is significantly lower that what can be attained with streaming from SSDs.
Not only in big desktops, but even in most recent mini-PCs, it is possible to read concurrently from one PCIe 5.0 SSDs and one PCIe 4.0 SSD, at a total sustained reading throughput of around 20 GB/s.
With an optimized inference implementation, it should be possible to overlap completely the computations with streaming weights from SSDs.
This should improve the inference speed to around at least 1 token per second, on a cheap computer, under $2000 even at the current super-inflated prices.
There are enough tasks where this would be useful. Obviously one should use for most tasks a fast small LLM and use the big one only when this actually saves time.
I'd rather not be thinking about tomorrow's meeting all night! Just trust "the robot has got this"... and scramble in the morning when the robot has failed. But at least I got a good night's sleep!
For values of "usable" that include "14.6 seconds/token". It's a cool accomplishment! And newer hardware would speed it up some. But I think I'd want something a bit faster before declaring it usable in practice.
"Better things aren't possible" is a terrible outlook. There have been real improvements in this space, such as passkeys, and recognition that some of this stuff, like frequent password changes, is counterproductive.
In what way are Passkeys, as implemented (not theoretical benefits), better than passwords?
Better: not in a single metric but rather as a complete measure of both preventing unauthorized access to a resource and also _enabling_ authorised access to that same resource.
They're better in some ways. They don't rely on a secret with low entropy which is really brute forceable. They can't be used on a phishing site because the URL is part of the secret. Even when you authenticate to a fake site you don't give them the ability to authenticate as you until you change the secret (like you do when you give them your password)
They also have 2fa built in. No need for a separate app, entering codes whatever.
That's why I mentioned "and also enabling authorised access to that same resource". Passkeys are great at preventing unauthorized access. But that comes at the expense of preventing authorised access. For example, using another device or even moving to another device. Replacing a stolen or damaged device is also nearly impossible with a reasonable quantity of Passkeys.
Well that's why they can sync between devices. I don't really see the problem.
Even if you don't like to rely on big tech (google/apple), I don't either, there are many options now for full FOSS implementations like bitwarden and KeepassXC.
If you use a yubikey as a passkey then yes, that's not a great option also because most services don't allow you to enroll more than one passkey. But with bitwarden that doesn't matter.
> Well that's why they can sync between devices. I don't really see the problem.
How do you sync passkeys to someone else's phone to authorize them to act in your name?
This is the basic use case, very common in the physical world, that security industry refuses to accept exists.
Passwords have this capability by nature.
> because most services don't allow you to enroll more than one passkey
Which is dumb and part of what makes passkeys not just useless, but dangerously so. Same story with 2FA, and the many services that only allow you to have one registered authenticator app at the time.
What Passkey implantation syncs between devices? I've only ever seen "cloud sync", e.g. syncing with someone else's computer. Can a user sync iPhone passkeys with her Boox E-ink tablet (Android)? Can either sync with a Debian desktop?
> There have been real improvements in this space, such as passkeys
Passkeys are not an improvement. The way they're being implemented entrenches middlemen into auth flows in a way that reduces users' security in a broader sense, while being just as susceptible to compromise as any other form of credential.
It's actually a little harder than people outside the UI framework space might think (there's really not a lot of time per frame; it's easy to put too much work where it doesn't belong; it's easy to regress if you're not monitoring for it after code changes).
It's core to usability at this point, though. They should make the effort.
Governments don't have a business model. They don't need to make a profit. (Also, debt doesn't work the same way for the federal government that it does for a business or even a US state.)
>The real reason we are using Claude is for the SaaS aspect of it. It has a toolchain, a friendly interface, and a bunch of integrations with business applications.
That is a very thin moat, though. There's nothing you can do with, for example, Claude Code + Opus 4.8 that you can't do with your own custom harness running API-level Opus 4.8, which means that if you can afford the hardware (the moat for running any SOTA model) you don't need to pay Anthropic anymore.
I'm not saying they shouldn't, but I understand why they don't.
