We recently launched Instant Clusters at RunPod, enabling on-demand deployment of multi-node GPU clusters with up to 64 H100 GPUs. This helps scale massive AI workloads like LLaMA 405B and DeepSeek R1 in minutes, leveraging high-speed interconnects and frameworks like PyTorch's torchrun.
Would love to hear how others are tackling multi-node orchestration challenges or scaling distributed AI workloads!"
Hey Hacker News, this is my new personal website that I built over the Labor Day weekend completely redesigning it and implementing .mdx files so I can write blog posts even faster. Please let me know your thoughts about the site and also about the writing!
Must be pretty good since nothing seems to be an obstacle and it's easy for your content to be the main thing.
Very inspiring I would say :)
Now focusing on the second curve where the rapid growth is asymptotic to a ceiling, which is about 90 degrees off from what most people want.
The rapid gains, tapering to level off through time, apply not just to personal abilities, but also related opportunities. That's a very steep slope to start with. Lots of times things like that can seem unsurmountable or unsustainable. Usually they are.
I recognized this well early in my lifetime of cargo work, since any suitable port for my growth ambitions has always been full of ships since before I was born, and they have "virtually" all been spoken for before I came along. Even before most of today's ships had been built.
In "pure" financial services as a youngster this had not been a limitation, and I got good experience at compound growth until the market crashed and everybody downsized, never to recover but a fraction of the lost potential.
Applying carefully considered growth principles with physical commodities limited by a single port, it's an accomplishment to be able to get any growth at all in a non-growth industry, where the market is already saturated with participants who have already achieved steady gains. Gains in both ability and financial opportunity. The curve, if it can get off the ground, is like your second graph. Even more of an accomplishment when the "rapid gains" period is not leveling but increasing something like 1 percent per day, even though the same ceiling (for that one port) is still there and the curve will have to more drastically reduce slope when the time comes.
What does all this mean? In a long-term pursuit where opportunities are dependent on abilities, the graphs are more applicable to both than they would otherwise be.
It would be hard to put it in further perspectives without other somewhat arbitrary attributes to the graphs, but let's go with the X-axis as 0 - 10 in years. And stay away from things where the 100 on the Y-axis is right there where the established experts are leveling off. The further the 100 point is located off-scale on the Y-axis the better off you could end up. Especially when nobody actually knows where the exact 100 point is, regardless of conventional wisdom or painful limitations.
Well you would be fortunate to level off with the experts and as the curve suggests this could carry on for a couple more decades in an even more flatlined condition. However, the more decades, the more chances for an unprecedented milestone that can once again be built upon in a compound way. Where the slope is very slight from this point but will increase rapidly if allowed to. All of it above a previous ceiling. None of it predicted by the previous curve up until that point.
These are like the leaps you anticipate. Opportunities that arise like this may each be a niche, but it can also be something that others can not offer.
The resulting curve, where your second graph takes the symmetrical gradual upturn in the future, does resemble the electrochemical progress of a titration reaction, where the neutral point just happens to be 100 millvolts (on the Y-axis this time). As reactive material (a third dimension) is added at a steady rate starting from zero (then continuing beyond the point of obviously diminishing returns), you never get above 100 for as long as any buffer lasts. Before you eventually get the uptick from continuing to add reactive material at the same rate the whole time. Some buffers are just more extensive and effective at resisting change compared to others ;)
Looks like the 10-year snapshot on the X-axis is a good starting point for many things after all ;)
That can be where the buffer is just starting to take effect, and very few may be able to make it to the other end. Progress can still be very worthwhile, and when it comes to longevity, it does seem to be a good strategy if you can outlive any remaining alternatives.
Easier said than done but I think I'll keep working on it anyway ;)
I recently started reading it and I love it. Her style of storytelling is very unique. It feels very meandering in a way, but rather than that making reading feel like a drag, I feel like I end up getting more color on the world she built.
I’ve been trying out Zed for the past two days. It’s for sure way more zen and focused compared to VS Code and I really enjoy that. At the same time, I felt the syntax highlighting wasn’t to the degree I’m used to and so I had to look a little longer to find certain words I had typed. All in all, I’m trying to find ways to use it because of how smooth it feels
I really want to try Linux and escape the Apple ecosystem. The main things holding me back are two things.
1. I use Raycast(an alternative to spotlight) that allows me to search files, open applications, open browser bookmarks, chat with GPT 3.5/4 instantly all with my keyboard. They only work on MacOS and I don’t know how to attain those levels of personal productivity on Linux
2. Good design. I do not like the design of WindowsOS at all and really enjoy Apple’s design in general. My design tastes lean to towards minimal, dark mode, colorful, “rounded corners”. I am looking for an OS that would be designed like apps like Linear, Notion, Obsidian, etc.
If there’s a disro out there that can solve these two things. I would love to switch to Linux
I’ve seen Raycast adjacent apps for Linux, but I don’t know what the current go-to all the kids are using these days. I used Quicksilver on OS X back in the day, which kind of defined the category I think. But these days I try to keep it simple.
Elementary OS seems to be trying to solve for the design issue, but it’s not as polished as macOS and there are still all the 3rd party apps to contend with. I’ve tried it, and it didn’t have me considering a full time switch.
For me, one of the big things keeping me on macOS is the sync between devices. 20 years ago it seemed much easier to move to Linux, because I didn’t have to worry about my contacts, calendar, etc being in sync everywhere. Having all that stuff, and the handoff between devices, just work is a huge benefit. There are probably ways to sync various things between iOS and Linux, but then researching all the options, setting it up, and keeping up on it all, becomes a hobby. That’s not a hobby I want at this stage of my life.
I think you would like the modern default Gnome desktop. Your needs in regard to search are built-in, except for the GPT stuff. In Gnome you can almost always just type away to navigate by search, without memorizing shortcuts. It's really great! There may be extensions for your "AI" needs, but I don't know for sure. However, Gnome is still sometimes limited by the development status of basic apps and services like mail, calendar and online accounts. Ironically, mostly affecting free standards like CalDAV, while Google Mail and Exchange integration is there. Also some common UI tweaks like having a dock, are hidden away as extensions.
Apple made it quite painful to share with iOS devices and access iCloud services outside the Apple ecosystem. So if you got an iPhone, that's gonna be mildly annoying.
I'd say try Fedora in a virtual machine and see for yourself!
