The problem with this idea is that someone can, and likely will, come up with the next best architecture that leapfrogs the current frontier models at least once a year, likely faster, for the foreseeable future. This means by the time you've manufactured your LLM on an ASIC, it's 4-5 generations behind, and probably much less efficient than current SOTA model at scale.
It won't make sense for ASIC LLMs to manifest until things start to plateau, otherwise it'll be cheaper to get smarter tokens on the cloud for almost all use cases.
That said, a 10 trillion parameter model on a bespoke compute platform overcomes a lot of efficiency and FOOM aspects of the market fit, so the angle is "when will models that can be run on an asic be good enough that people will still want them for various things even if the frontier models are 10x smarter and more efficient"
I think we're probably a decade of iteration on LLMs out, at least, and the entire market could pivot if the right breakthrough happens - some GPT-2 moment demonstrating some novel architecture that convinces the industry to make the move could happen any time now.
We're teetering on the brink of highly capable software agents that can run on a phone using a local model, that can manage things like basic digital hygiene, operating a self-hosted cloud, with tailscale and other private vpns that can leverage your own home internet service with a well maintained set of firewall rules and level of locked-down access that it's actually practical.
An inspired nerd can do it right now, but grandma will be able to do a curated, accessible set of things by the end of the year, and by the end of 2027, the internet and self hosted things are going to be incredibly different. When people can self host plex and anonymously pirate anything, and their local model can do the ethically gray area stuff like ensure everything is done so they don't get caught - cloud services can't compete with that. Cable and netflix and spotify and the rest are going to have to up their game, and not do the stupid lashing out, price gouging, hunting the pirates type of thing or they're just going to burn down faster.
We're headed for some really cool, interesting times.
If it catches engagement, the main firehose feed will show it. They've begun using Grok and AI processes, which is hit and miss, but definitely improving.
Having Japanese, French, other countries' tweets automatically translated back and forth has been fun, too. It'll be interesting to see where it gets to in the next few years.
The other day, I looked at the trending topics. Top one was "Lesbians". I was wondering if there was some kind of development in politics. Nope.
It was all porn. I was on a call with a friend and he checked from his account too and it was there as well, so this wasn't some kinda A/B test thing. It disappeared after a bit. My point is the algo is a bit wonky.
Twitter had always been the modern day Playboy mag from Sci-Fi era. So there's Bradbury, Lenna, geopolitics, all bound in one.
The catch is it's a UGC based algorithmic system with instant feedback, which means the fastest adapting contents with most bandwidth absolutely wins, which tends to be, like that.
Does anyone have the solution to this problem anyway? I thought this was always inevitable on WWW.
Not with 600m organic global active users. The platform moral compass must align to the performance weighted sum total of its user, rather than the other way around.
Trying to bend the platform morality to suit your idealisms seriously ruin yours over time.
I don't see how seeing the current feed of words and going "not this one" before they go online is difficult. You could literally filter 10 per minute and clean up the misfires
It's awesome if you mod your own gear, and 3d printing / one off part services are ubiquitous, so if you see something you like online, it's cheap and easy to do little upgrades.
More companies should do what they do - the less ethical players are already cloning knockoffs anyways, stuff like this builds brand loyalty and probably makes it more likely that people stick with Keychron over going for the knockoffs.
It's worth it. The increased participation and discussion have given a little momentum in usability, and AI on hand makes the learning curve very manageable. If you're already familiar with vmware, virtualization in general, it's a pretty easy transition.
I switched from VMWare to Proxmox a few years ago because Proxmox supported a wider range of network cards that were more common in the cheap desktop computers I use in my homelab, whilst VMWare almost required an Intel network card (which was usually fine for server hardware).
It was a surprisingly easy transition that I have not regretted one bit. I'm not sure whether there that was an actual migration path, without reinstalling servers from scratch. Homelab meant it didn't quite have the requirements of a production system...
I'm kinda hoping AI agents pass the threshold of being able to reliably do a complete production migration sometime this year. We've got a couple years left on the vmware contract, and it's just obscene what they've done, with the price hikes, degraded support, etc.
At this point they're more an enterprise scheme to maximize license profiteering for compatible software and OS "per core" licensing in conjunction with hardware platform providers. It'd be cool if the support were worth the price, but in most enterprise cases, you're going to save a lot of money if you pay enough full time staff to purchase, build, maintain, and operate a virtualization environment compared to what the enterprise platforms provide. In most cases, you'll save more than enough to keep a better specced system completely redundant with spare hardware on hand.
What's the difference between the idle imaginings of a god's mind and a universe scale simulation?
I always got advanced AI vibes from Devs, that it was a mind interfacing with reality in some sort of weird inception / simulation / manifestation way.
That can be a type of mind, though? It can also be a type of interface - a tap into a system not fully understood, controlling the perspective or view but not the process. The whole "mind of god" Deus/Devs, etc - I think it's left ambiguous on purpose for the hook but I always took it to be an AI flavored story at the core.
Maybe. The plot itself is based on a short story that Garland read, both the 2007 original [0] and its 2022 rewrite [1]. Qntm is great and their latest book, There Is No Antimemetics Division, recently was on HN as well [2].
It won't make sense for ASIC LLMs to manifest until things start to plateau, otherwise it'll be cheaper to get smarter tokens on the cloud for almost all use cases.
That said, a 10 trillion parameter model on a bespoke compute platform overcomes a lot of efficiency and FOOM aspects of the market fit, so the angle is "when will models that can be run on an asic be good enough that people will still want them for various things even if the frontier models are 10x smarter and more efficient"
I think we're probably a decade of iteration on LLMs out, at least, and the entire market could pivot if the right breakthrough happens - some GPT-2 moment demonstrating some novel architecture that convinces the industry to make the move could happen any time now.
reply