Fatih Cetinkaya

Who came up with this FireTV PIN screen?

Full circle, text.

The weekend build: low cost and usable (sort of) local AI coding agent for people on a budget. Nvidia Tesla P40 (24GB), eGPU (Oculink) and a cheap mini PC (Intel N100). Running in VSCode via Ollama.

Ordered and looking forward

The AI model finally learned how to hit the ball, but it still has a long way to go. 🤖🕹️

I'm doing some AI training for the Duel of the Paddles game so it won't require two players to play. So far, after 250K+ steps (RL/PPO), it's still performing terribly. 😅🕹️🤖 #ai #artificalintelligence

There are some really points and thoughts by @aaronwhite in this "Who Will Win the Incumbent AI Battle?" video. youtube.com/watch?v=X9SgohV6… #ai #artificialintelligence

I finally had a chance to rearrange my Pi5 cluster today. For the compute-only Pi5s (fan only, no USB/PCIe/NVMe), I got rid of the official Pi5 power supplies and replaced them with Anker PowerPort 6, which is common for Pi clusters. One of the reasons was that no matter what I did, the Pi5s didn't go beyond 9W of power consumption (3W idle). So, the official Pi5 power supply was overkill for my use case. Although the Anker PowerPort can handle 60W, I decided to be a good citizen and only connected 4 Pi5s to each unit, leaving plenty of extra overhead. By the way, they are connected to two different switches. 😅 While looking for some cables, I found my old AC Infinity fans, which I used to cool an Xbox. I thought, why not, and set them up to cool the power supplies. The fans have a thermal trigger, so they only run once the temperature goes beyond ~29°C/84F. By the way, I asked ChatGPT how to position the fans, and it suggested the push-pull configuration. I was skeptical at first, but it seems to work well. See the temperature in the image below. Let me know if there is a better way. Finally, I got these fancy USB-C cables that show the current power consumption. I like them and hope they last long enough since they are not cheap. 💸 P.S.: I'm still waiting for the official Raspberry Pi5 PoE HAT for my power-hungry Pi5s, which have NVMe disks. Let me know if you have any suggestions. #raspberrypi #pi5 #anker #usbc

I saw this variation of the pong game (link below) where two balls continuously move, aiming to capture the opposing player's area. My first question was how this game could end, if at all. I didn't see a possibility (correct me if I'm wrong) without modifying the game. So, I'm recreating the game with some changes: 1) Adding paddles so that players can directly control the balls, and 2) Introducing power-ups to give players an advantage. I have some ideas for power-ups, such as increasing the ball's speed every time it passes to the opposite player's paddle and hits the wall or giving an extra ball to the players when that occurs. I'm open to any suggestions and ideas. I'm counting on folks out there who lived through the '80s and '90s and played these kinds of games. You can find the original game at pong-wars.koenvangilst.nl And below is a screenshot from the game that I'm developing. #gaming #pong

Sharing the pictures of my old Macs for the Macintosh 40th Anniversary. Happy birthday Mac! #Apple #Mac #Macintosh

Single Pi5 performs ~65 embeddings/sec (including network overhead)

I don't have those scores. But each Pi5 serves around ~65 embeddings/sec (including network overhead)

Everything is handled by @nats_io Check out @thecodegangsta ‘s video at youtu.be/byHGNUqIONw

For edge, I started with Terraform, which is my powerhouse, but it quickly became overkill. So, I switched to regular shell scripts. That being said, I recently watched a video by @synadia (youtube.com/watch?v=EfxtiKMn…) and I'm thinking of switching to their approach as soon as they release it. For cloud, I'm using Fly.io, which hosts the multi-region NATS server and the public API server.

I haven't run any LLM yet. Although you can find some numbers from other people here on X. I used it for @embedefy See xitter.cf/devfacet/status/174912… It can perform ~65 AI inference/sec with a single Pi 5 (including network overhead). But I'm also working on batch API to make it faster.

It can perform ~65 AI inference/sec with a single Pi 5 (including network overhead).

I’ve already tried some of them, the software support (e.g. Linux kernel) and system stability (Google it yourself) is no where near Raspberry Pi. They were keep crashing.