Permalink

Words or tokens are represented as high-dimensional vectors (embeddings). Matrix operations (matrix multiplications, transformer attention) dominate the training and inference of models. These computations require massive and parallel computing power, similar to that used in 3D graphics.

Post image
Permalink

old times

Post image
Permalink

Reminder : Go do some hacking today

Post image
Permalink

Unix tools have endured for so long because they are simple, efficient, and composable. The Unix philosophy: 'Do one thing, and do it well.

find – 1971 / chmod – 1971 / ls – 1971
cd – 1971/ cp – 1971 / mv – 1971
rm – 1971 / cat – 1971 / echo – 1971
grep – 1973/ sed – 1974 / awk – 1977
sh – 1977 / tar – 1979

Post image
Permalink

If I had to choose a VM hypervisor, I'd pick XCP-ng over Proxmox.

It’s a French product, and its interface is clear and pleasant. Essential information is well highlighted. In contrast, I find Proxmox’s interface outdated, almost Soviet-era. I know it’s performance-wise solid in terms of UX, but we can do better than this kind of Soviet-style UX.

Plus, the backup system is included by default—very practical and powerful. With Proxmox, you apparently need to install a separate OS like Proxmox Backup Server, which just adds unnecessary complexity to the architecture.

xcp-ng.org

Post image
Permalink

The New Internet

" I have bad news: if you remember doing LANs back in the 1990s, you are probably old. Quite a few of us here at Tailscale remember doing LANs in the 1990s. That’s an age gap compared to a lot of other startups. That age gap makes Tailscale unusual. "

" I read a post recently where someone bragged about using kubernetes to scale all the way up to 500,000 page views per month. But that’s 0.2 requests per second. I could serve that from my phone, on battery power, and it would spend most of its time asleep. "

pdf
nicolas-dorriere.fr/assets/files/The_New_Internet.pdf

source
tailscale.com/blog/new-internet

Post image
Permalink

Parallel inference (Gemini, Claude, GPT) in lightweight Firecracker virtual machines on the cloud. For example, create a Git branch for a new feature and then merge it.

Monorepos are likely to gain popularity in 2025, 2026, and 2027.

openai.com/codex

Post image
Post image
Permalink

The future of work could resemble an RTS like Age Of Empire—fleets of ai-micro-agents, each isolated in lightweight Firecracker micro-VMs, autonomously handling tasks, data collection, communications, and system design.

Post image
Permalink

The benchmark solves a random dense linear system in double-precision (64 bits / FP64) arithmetic

github.com/geerlingguy/top500-benchmark

GMKtec NucBox G3 Plus (N150 @ 3.6 GHz)
62.067 Gflops
28.5W    
2.18 Glops/W

Post image
Permalink

I’m building an SOS button for my mother. To do this, I’m using two Raspberry Pis connected through a Tailscale tunnel. When the button is pressed, an SSH command is sent to the other Raspberry Pi, which plays an MP3 loudly through speakers. At the same time, a Telegram alert is sent to our family channel. I often have my phone on silent at night — but now, no problem: the speakers will shout loudly! I’m coming, Mom! πŸ‘ΈπŸ»

Post image
Permalink

George in FLAC 🀩 with an average of 900 Kbit/s — all I’m missing now is the Devialet Phantom. For now, I’m listening on my Logitech Z-5500, limited to 16-bit through my optical SPDIF output.

Post image
Permalink

arts @home

Post image
Permalink

Nóva is well known for her open source contributions including projects like Linux, Kubernetes, and The Go Programming Language. A popular public speaker, she is best known for her Twitch stream, and her Kubernetes clusterfuck talks.

krisnova.net

Post image
Close
Fullscreen image