At the Salishan HPC conference last week, I gave a talk called "AI doesn't need massive supercomputers after all!"
A couple people asked me for the slides (as crappy as they were), so here they are in hastily written blog format.
https://blog.glennklockwood.com/2026/05/ai-doesnt-need-giant-supercomputers.html
Speculative decode is an inferencing optimization that was mentioned a few times at #GTC26. I'd heard of it but didn't know how it worked, so I spent some time figuring it out. Some notes (and toy code that illustrates its benefits!) are here: https://glennklockwood.com/garden/speculative-decode
Sooo are #HPC centers now having to deal with students trying to run models for OpenClaw on their systems? Genuine question.
In today's edition of "smart people writing dumb things," Mustafa Suleyman wrote an op ed in the MIT Tech Review on how there are no scaling limits in sight. But his argument hinges on a 2022 study. Also goes on to imply that solar will make #AI affordable?
https://www.technologyreview.com/2026/04/08/1135398/mustafa-suleyman-ai-future/
Some notes I wrote up about #GTC26 last week: https://blog.glennklockwood.com/2026/03/gtc-2026-recap.html
I first noticed Claude and ChatGPT talk about "pressure testing" ideas ~six months ago, and only now am I hearing humans use the phrase in conversation. Is this a case of AI successfully training humans? Or is this just a coincidence?
These new NVIDIA STX storage node implementations from Quanta and Supermicro appear identical even down to the LED and USB-C placement. I wonder if the drive carriers are differentiated.
Interesting that the #SC26 panel submission process
1. Only provides a template in LaTeX, and
2. The template implies that panelists should all have University affiliations
I don't think this reflects how most of the world works these days.
I did an interview yesterday after a journalist found my website via Gemini. This wasn't the first time this happened. He said to keep doing whatever SEO I've been doing for AI-based search, but truth is, I do zero SEO. I just don't use LLMs to write.
Given how much AI-generated thinkpieces are popping up online (LinkedIn, Twitter, etc), I worried that the quality of search would go down. But it seems like search/AI is keeping up and not surfacing slop. Keeps me optimistic for the future.
These speculative decode results match up with my experimentation with different draft and target models. Getting the circumstances right is necessary to make speculative decode worthwhile.
Went down a rabbithole of looking at what MLPerf Storage Training benchmarks do; turns out they're not very meaningful IMO. Alarming how many practitioners (mostly storage vendors) run them and don't understand what the results do/don't represent.
Somehow it is less mind-numbing to tell agents how to do some tedious task (like a content audit) than do it myself. I doubt I'm saving any time, but it's satisfying to know I'm not the one copying, pasting, and ctrl+f'ing over and over.
Maine is about to lead the way in banning new datacenter buildout (https://www.wsj.com/us-news/maine-data-center-ban-e768fb18). Except it doesn't have any datacenter buildout going on (https://www.visualcapitalist.com/ranked-us-states-data-center-hotspots/).
What a loss.
