Lwar
Region: US
Thursday 18 June 2026 13:58:06 GMT
1305
93
33
5
Music
Download
Comments
Frank Erling :
The data centers will be obsolete before they are turned on
2026-06-18 18:33:12
3
fortunateloss :
Clustering consumer laptops over a network would have horrible performance since inference is almost exclusively memory bound. The model needs to read the entire model weight matrix and the KV Cache memory for every single token generated so trying to split via pipeline parallelism (Laptop A runs layers 1-10, Laptop B runs 11-20, Laptop C runs 21-30), Laptop A has to send the intermediate activation tensors to Laptop B over the network for every single forward pass and unless you’ve got 100G+ Ethernet or a miniature IB network at home that’s going to perform like trash.
2026-06-18 17:11:42
2
Neal :
The aftermarket prices on those machines are gonna be nuts
2026-06-21 02:57:12
0
XYZ :
Bro NVIDIA them selves recently unveiled a whole new line of laptops capable of running large LLMs, coming out soon 🤔
2026-06-18 16:20:15
1
Chris Baker :
They’re making RAM too expensive for most people to afford. It’s more than doubled in price in the last year
2026-06-19 13:41:46
0
Travisty :
The problem isn’t that it can’t run locally, it’s that we are still early stage. Running local right now would be like buying a VR headset in 2004. It’s fine for the time but it won’t scale.
2026-06-21 14:56:20
0
The CCSI :
My stack is routing for token efficiency. prompts that require cloud inference get cloud access, everything else takes the local slow lane. Local can run on almost anything from the last 10 years. Speed varies by hardware stack and model, but it will chug through the work and respond. If you have a decent gaming machine from the last 4 years, you may seldom go to the cloud.
2026-06-20 14:45:41
0
seansie :
real question; why can AI do things that humans aren't allowed to? they can steal and pirate data to train, but all of that data and info is behind a paywall for us. they can produce AI slop video and music but if we did it it would be copyright infringement or a violation of trademark. if we had a private data center in our living room hooked up to our hose spigot sucking up water, not only would we be charged a full price for that water as opposed to some friggin discount or tax break, they would actively shit off our water supply if we continued.
2026-06-19 13:47:50
0
miked :
200B parameter is small in the world of frontier models.
2026-06-19 08:00:42
2
Genxmahm :
So on-Prem again. I’m tired.
2026-06-19 11:05:12
1
Jordan Rey :
Local AI would be super useful
2026-06-18 14:03:29
3
MetalDrgn :
It's both really that we need
2026-06-18 22:58:20
0
scottlevine112 :
I called this a year ago.
2026-06-19 14:51:03
0
Space Christ :
I want 1T params !! 😂
2026-06-19 01:17:19
0
Anonymoss :
I barely comprehend what you’re saying but this feels related to the rise of Model Context Protocols? I remember engineering explaining this to us. MCPs almost feel necessary if companies want to remain sustainable in their technology plus more control over governance.
2026-06-18 15:44:56
0
prof_info :
I thought the data centers were required for the training of the model. the execution of the model requires much less compute and can be run locally. Running larger models is an exponential problem.
2026-06-19 14:03:42
0
Charmeleon07 :
For end users and data security that would be awesome for lots of corporations
2026-06-18 14:14:47
0
so.over.vic :
How big is Codex 5.5 Max? This could pay for itself in 18 months
2026-06-20 14:59:02
0
XYZ :
Btw stock market at all time high 🤔😳😉
2026-06-18 16:13:34
0
CBrawlerSr :
I’ll just use DeepSeek which is ALREADY robbing the U.S. of profits because SO MANY U.S. COMPANIES USE IT FOR PENNIES 😁😁😁
2026-06-20 13:35:56
0
Russell Freeland :
spot on. the cloud models will still be superior but most of us don't need Superior we just need a certain set of functions. like most of us don't need military grade encryption we settle for pgp pretty good privacy. will be settling for pretty good AI, because for the things that we do everyday we don't need ridiculous amounts of thinking power from the AI or all kinds of new developments
2026-06-19 07:45:52
0
To see more videos from user @lwar510, please go to the Tikwm
homepage.
