• 0 Posts
  • 261 Comments
Joined 1 year ago
cake
Cake day: June 19th, 2023

help-circle

  • Electronics components do not like to have power states change frequently. Turning devices on and off frequently will decrease lifespan of device. Sure, you are saving money on your electricity bill, but at some point, the savings and environmental impacts are outweighed by the cost of the device/parts and the impact during manufacturing.

    Also, don’t forget phantom draws from the power supplier is a real thing, which will most likely exceed your 5 zeros threshold. So that microwave oven, and laundry dryer? Don’t forget to unplug those after each use.



  • The amount of confidently incorrect responses is exactly what one could expect from Lemmy.

    First: TCP and UDP can listen on the same port, DNS is a great example of such. You’d generally need it to be part of the same process as ports are generally bound to the same process, but more on this later.

    Second: Minecraft and website are both using TCP. TCP is part of layer 4, transport; whereas HTTP(S) / Minecraft are part of layer 7, application. If you really want to, you could cram HTTP(S) over UDP (technically, QUIC/HTTP3 does this), and if you absolutely want to, with updates to the protocol itself, and some server client edits you can cram Minecraft over UDP, too. People need to brush up on their OSI layers before making bold claims.

    Third: The web server and the Minecraft server are not running on the same machine. For something that scale, both services are served from a cluster focused only on what they’re serving.

    Finally: Hypixel use reverse proxy to sit between the user and their actual server. Specifically, they are most likely using Cloudflare Spectrum to proxy their traffic. User request reaches a point of presence, a reverse proxy service is listening on the applicable ports (443/25565) + protocol (HTTPS/Minecraft), and then depending on traffic type, and rules, the request gets routed to the actual server behind the scenes. There are speculations of them no longer using Cloudflare, but I don’t believe this is the case. If you dig their mc.hypixel.net domain, you get a bunch of direct assigned IP addresses, but if you tried to trace it from multiple locations, you’d all end up going through Cloudflare infrastructure. It is highly likely that they’re still leaning on Cloudflare for this service, with a BYOIP arrangement to reduce risk of DDOS addressed towards them overflow to other customers.

    In no uncertain terms:

    1. Hypixel.net has Cloudflare DNS for their domain.
    2. For their website, it has orange cloud enabled to proxy traffic through CF’s global CDN and DDOS protection service.
    3. For their Minecraft server, they advertise mc.hypixel.net, but also have a SRV record for _minecraft._tcp.hypixel.net set for 25565 on mc.hypixel.net
    4. The mc.hypixel.net domain has CNAME record for mt.mc.production.hypixel.io. which is flattened to a bunch of their own direct assigned IP addresses.
    5. Traceroute towards those direct assigned IP addresses goes through Cloudflare infrastructure but final destination is obscured, just like their website, to protect them from DDOS attacks.

  • Using Ollama to try a couple of models right now for an idea. I’ve tried to run Llama 3.2 and Qwen 2.5 3b, both of which fits my 3050 6G’s VRAM. I’ve also tried for fun to use Qwen 2.5 32b, which fits in my RAM (I’ve got 128G) but it was only able to reply a couple of tokens per second, thereby making it very much a non-interactive experience. Will need to explore the response time piece a bit further to see if there are ways I can lean on larger models with longer delays still.












  • If memory serves, 175B parameters is for the GPT3 model, not even the 3.5 model that caught the world by surprise; and they have not disclosed parameter space for GPT4, 4o, and o1 yet. If memory also serves, 3 was primarily English, and had only a relatively small set of words (I think 50K or something to that effect) it was considering as next token candidates. Now that it is able to work in multiple languages and multi modal, the parameter space must be much much larger.

    The amount of things it can do now is incredible, but our perceived incremental improvements on LLM will probably slow down (due to the pace fitting to the predicted lines in log space)… until the next big thing (neural nets > expert systems > deep learning > LLM > ???). Such an exciting time we’re in!

    Edit: found it. Roughly 50K tokens for input output embedding, in GPT3. 3Blue1Brown has a really good explanation here for anyone interested: https://youtu.be/wjZofJX0v4M





  • 4o does perform web searches, give summaries from a couple of pages, and include the link to those pages when prompted properly.

    However, as most people know, first couple results doesn’t always tell the full picture and further actual researches are required… but, most “AI assistant” (also including things like those voice assistants in speakers) users tends to take the first response as fact…

    ¯\_(ツ)_/¯