The new lawsuit said Li began working as an engineer for xAI last year, where he helped train and develop Grok. The company said Li took its trade secrets in July, shortly after accepting a job from OpenAI and selling $7 million in xAI stock.

I must say that it’s going to be a bitch-and-a-half to retain core engineers if people are walking away at what amounts to $7 million/year in effective compensation.

kagis

https://www.lxuechen.com/

Looks like he only started working in industry in 2023, too (though was doing relevant work as a graduate student prior to that).

If I recall correctly, at least for non-group chats they do use end-to-end encryption. That being said, obviously there are some practical limitations on the impact if you think that WhatsApp would actively try to be malicious, since they’re also providing the client software and could hypothetically backdoor that.

kagis

According to this, they do use end-to-end encryption for group chats too.

Maybe I’m recalling some other service or a default setting or something. Some service had non-e2e-encrypted-group messages for at least some period of time.

$3-10k…not getting the speeds and quality

I mean, that’s true. But the hardware that OpenAI is using costs more than that per pop.

The big factor in the room is that unless the tech nerds you mention are using the hardware for something that requires keeping the hardware under constant load — which occasionally interacting with a chatbot isn’t going to do — it’s probably going to be cheaper to share the hardware with others, because it’ll keep the (quite expensive) hardware at a higher utilization rate.

I’m also willing to believe that there is some potential for technical improvement. I haven’t been closely following the field, but one thing that I’ll bet is likely technically possible — if people aren’t banging on it already — is redesigning how LLMs work such that they don’t need to be fully loaded into VRAM at any one time.

Right now, the major limiting factor is the amount of VRAM available on consumer hardware. Models get fully loaded onto a card. That makes for nice, predictable computation times on a query, but it’s the equivalent of…oh, having video games limited by needing to load an entire world onto the GPU’s memory. I would bet that there are very substantial inefficiencies there.

The largest GPU you’re going to get is something like 24GB, and some workloads can be split that across multiple cards to make use of VRAM on multiple cards.

You can partially mitigate that with something like a 128GB Ryzen AI Max 395+ processor-based system. But you’re still not going to be able to stuff the largest models into even that.

My guess is that it is probably possible to segment sets of neural net edge weightings into “chunks” that have a likelihood to not concurrently be important, and then keep not-important chunks not loaded, and not run non-loaded chunks. One would need to have a mechanism to identify when they likely do become important, and swap chunks out. That will make query times less-predictable, but also probably a lot more memory-efficient.

IIRC from my brief skim, they do have specialized sub-neural-networks, which are called “MoE”, for “Mixture of Experts”. It might be possible to unload some of those, though one is going to need more logic to decide when to include and exclude them, and probably existing systems are not optimal for these:

kagis

Yeah, sounds like it:

https://arxiv.org/html/2502.05370v1

fMoE: Fine-Grained Expert Offloading for Large Mixture-of-Experts Serving

Despite the computational efficiency, MoE models exhibit substantial memory inefficiency during the serving phase. Though certain model parameters remain inactive during inference, they must still reside in GPU memory to allow for potential future activation. Expert offloading [54, 47, 16, 4] has emerged as a promising strategy to address this issue, which predicts inactive experts and transfers them to CPU memory while retaining only the necessary experts in GPU memory, reducing the overall model memory footprint.

Oh, wait, yeah, you’re right, and in fact a number of packages do take that when binding to an address. Sorry, that’s on me.

cannot bind to local IPv4 socket: Cannot assign requested address

inet 169.254.210.0

Yeah. That’ll be that you’re needing an interface with that address assigned.

ifconfig

Going from memory, I believe that if you’ve got ifconfig available and this is a Linux system and you need to keep the address on the current interface to keep the system connected to the Internet or something, you can use something like ifconfig enp7s0:0 10.10.10.3 to use an interface alias, use both addresses (169.254.210.0 and 10.10.10.3) at the same time. Might also need ifconfig enp7s0:0 up after that. That being said, (a) I don’t think that I’ve set up an interface alias in probably a decade, and it’s possible that’s something has changed, (b) that’s a bit of additional complexity, and if you aren’t super familiar with Linux networking, you might not want to add more complexity if you don’t mind dropping just setting the address on the interface to something else.

Probably an iproute2-based approach to do this too (the ip command rather than the ifconfig command) but I haven’t bothered to pick up iproute2 equivalents for a bunch of stuff.

EDIT: Sounds like you can assign the address and bring the interface alias up as one step (or could a decade ago, when this comment was written):

https://askubuntu.com/questions/585468/how-do-i-add-an-additional-ip-address-to-an-interface-in-ubuntu-14

To setup eth0:0 alias type the following command as the root user:

# ifconfig eth0:0 192.168.1.6 up

So probably give ifconfig enp7s0:0 10.10.10.3 up a try, then see if the TFTP server package can bind to the 10.10.10.3 address.

I haven’t done anything with OpenWRT for a lomg time, but…

I have the IP of the server set to 0.0.0.0:69 when I try to set it to 10.10.10.3 (per the wiki) The server on my pc won’t start and gives an error.

I’m pretty sure that you can’t use all zeroes as an IP address.

kagis

https://en.wikipedia.org/wiki/0.0.0.0

RFC 1122 refers to 0.0.0.0 using the notation {0,0}. It prohibits this as a destination address in IPv4 and only allows it as a source address during the initialization process, when the host is attempting to obtain its own address.

As it is limited to use as a source address and prohibited as a destination address, setting the address to 0.0.0.0 explicitly specifies that the target is unavailable and non-routable.

You probably need to figure out why your TFTP server is unhappy with 10.10.10.3, and there’s not enough information here to provide guidance on that. I don’t know what OS or software package you’re using or the error or the network config.

It may be that you don’t have any network interface with 10.10.10.3 assigned to it, which I believe might cause the TFTP server to fail to bind a socket to that address and port when it attempts to do so.

If you are manually invoking the TFTP server as a non-root user and trying to bind to port 69, and this is a Linux system, it will probably fail, as ports below 1024 are privileged ports and processes running as ordinary users cannot bind to them. That might cause a TFTP server package to bail out.

But I’m really just taking wild stabs in the dark, without any information about the software involved and the errors you’re seeing. I would probably recommend trying to make 10.10.10.3 work, though, not 0.0.0.0.

If this is a Linux system, you might use a packet sniffer on the TFTP host, like Wireshark or tcpdump, to diagnose any additional issues that come up, since that will let you see how the two devices are talking to each other. But if you can’t get the TFTP server to even run on an IP address, then you’re not to that point yet.

Stephen Miller is not known for exactly being the soul of veracity.

https://www.salon.com/2017/02/21/stephen-millers-web-of-lies-the-trump-adviser-has-championed-big-tobaccos-alarming-deceptions_partner/

And that was from back in 2017.

Trump just has a number of people working for him who are, like himself, willing to tell very substantial untruths.

I don’t use this plugin myself, but if you’re using Firefox, you might take a look at it, as it provides a bunch of browser-side configurability. I don’t know whether the feature you’re looking for is there, but as far as I can tell, it aims to be a pretty large bucket of pretty much every add-on YouTube feature one might want.

I was looking at it a while back for something unrelated, a UI tweak that I was hoping that it might do.

I tend to think of “sports” as being things that require not just mental apititude, but strength, endurance, and the like, but competitive shooting is an Olympic sport and has been for as long as the modern games have been around, so…shrugs

I mean, is curling a sport? Bowling? I think that most would call those sports, but they don’t really rely on exceptional strength or endurance.

I do think that there’s an argument for having some kind of word that encompasses all sorts of competitive activities, but also an argument for having a term for the smaller set that requires excellent physical conditioning. You might want to refer to either set. Maybe just have two different terms, regardless of where “sports” winds up.

If it’s periodically dropping, I doubt that it’s going to be the userspace WiFi backend. I’d be more-inclined to blame either the driver or an (unlabeled) hardware variation.