It's a machine that tries to guess numbers on a page having looked at lots of pages of numbers before.

ELI5 ?

Futurism links to the SCMP (which acts under the Chinese Communist Party's censorship regime) which in turn is referencing a CNAS' 'online event' and a 'CNAS report published last week', though there are no links to the online event nor the report.

I couldn't find the online event, but here is the report for those who want to read it: Superpowers and AGI - How U.S.-China Competition Could Be Shaped by the Emergence of Artificial General Intelligence

The linked article is quoting Stokes, for example, saying,

"Data centers can be bombed with conventional munitions,"

and,

"Bombing would cross a major threshold that heretofore has not been crossed in US-China relations."

I can't find these and other quotations from the linked article in the original report (not even the SCMP is quoting these).

In the original report, there is no mention of a "US military strike against China" as the SCMP headline tries to make people believe.

In the original report, words like 'bombing' and 'espionage' - quoted by Futurism - don't even exist.

Though the report discusses the U.S./China rivalry regarding AGI, it does so in a much larger and more diverse frame, including the common U.S./China competition with the Global South and other regions, what supposedly happened if and when any other country develops AGI first, and more things related to this topic. It is by no means a report "U.S. versus China" as the linked article (and the SCMP) makes people believe.

AI critic Eliezer Yudkowsky - which is also not mentioned in the SCMP - indeed suggested in 2023 that countries needed to be ready to “destroy a rogue da center by airstrike," but this was not directed specifically to China or any other country. He meant that as a possible (and, for now, theoretical) future measure in case a data center becomes dangerous for humans and humanity at some point in the future.

But you can perfectly combine all these fake quotations and half-truths to create a hopelessly distorted reality. This is pure propaganda aiming to sow division.

sooooo.. this is one of those articles you have to read. the guys is in a think tank where they come up with different ideas. he even talks about the issues. i guess someone heard it and ran with it as a deceleration of sonething we must do

Anyone in position of power who suggests actions that may bring the world to brink of ww3 must be prosecuted for crimes against humanity.

On another perspective you have to be a complete dumbfuck of an idiot to suggest bombing a country who supplies you a great chunk of your daily and advanced tech components. All because of what? Because Chinese companies were able to make LLMs that prove Western counterparts were just big inflated bubbles? Is that your forecast of achieving AGI? JFC the people running the most powerful country on Earth..

wat

jesus fuck ww3 christ in a crockpot

Holy shit, how the fuck does someone say something like that with a straight face?

"Yeah I'm a little concerned about the idea someone might make something I saw about in a movie, lets trigger ww3 because fuck it"

Lol, do it. I want to see the US economy crash completely.

Thats the stupidest idea I’ve heard in a long time.

Direct link to the actual blog post this clickfarm article is about: https://eebench.org/blog/can-ai-design-circuit-boards-yet/

I am brand new to Pi as of this week, so far my evolution in the past years went from... Visual Studio (oh wow it'd been so long i've forgotten the name 🤣).... then Cursor... then Claude Code. I tried a few neovim integrated AI workflows over a year ago, but they were all too buggy.

I caught wind of Pi I think from the youtuber omerxx and decided to give it a spin cause of its claims to be highly customizable. I haven't had the same time I did in the past to really tinker with my setups, but I figure i'll do it in baby steps over time.

Thanks for the info, it's totally welcomed - I am trying my best to learn by sponge mode 🧽 😄

Yeah, I had not considered using one of those middle grounds and will check them out 🙏

Great, thanks for the info! How do you like hermes, have you tried out pi?

Did you give Qwen3.8-27B-FP8 a try?

I haven't, it seems like I can't do this with my current Vulkan setup, but I'll do some exploration.

I'm also waiting for the nixos support for glm 5.3 flash

Hey! thanks for the reply :) yeah the slowing down part is a big factor.. i'm certainly being a little impatient, so well received 🙏😁
I had also noticed I had a big mismatch in my pi models.json's contextWindow and maxTokens, which I just fixed (shown in another comment) so going to test it out some more.

Hey, thanks for the reply, the speed is actually pretty good!

So it's running on llama.cpp (llama-server) b10408, Vulkan/RADV, behind llama-swap.

Your questions made me realize i also had a mismatch in my pi model config, the contextWindow and maxTokens had drifted apart pretty significantly so I just fixed that now.

/etc/nixos/configuration.nix

  services.llama-swap = {
    enable = true;
    package = pkgs.llama-swap;
    listenAddress = "0.0.0.0";
    port = 8080;
    openFirewall = true;

    settings = {
      # 63 GB off NVMe cold is slow. The default 120s will time out.
      healthCheckTimeout = 600;
      logLevel = "info";

      models = {
        "glm-flash" = {
          # Keep "local" as an alias so existing clients keep working
          # without touching their base URL or model name.
          aliases = [ "local" "fast" ];
          ttl = 1800;
          cmd = ''
            ${lib.getExe' pkgs.llama-cpp-vulkan "llama-server"}
            --port ''${PORT}
            -m /srv/llm/models/GLM-4.7-Flash-Q8_0.gguf
            -ngl 999 --flash-attn on -c 131072 -np 1
            --jinja --no-webui
            --reasoning-format auto --reasoning-preserve
            --temp 1.0 --top-p 0.95 --repeat-penalty 1.0
            --cors-origins "http://optimus.local:8080,http://localhost:8080/"
          '';
        };

        "gpt-oss-120b" = {
          aliases = [ "smart" ];
          ttl = 1800;
          cmd = ''
            ${lib.getExe' pkgs.llama-cpp-vulkan "llama-server"}
            --port ''${PORT}
            -m /srv/llm/models/gpt-oss-120b/gpt-oss-120b-MXFP4.gguf
            -ngl 999 --flash-attn on -c 65536 -np 1
            --jinja --no-webui --reasoning-format auto
            --temp 1.0 --top-p 1.0 --top-k 0 --min-p 0.0
            --cors-origins "http://optimus.local:8080,http://localhost:8080/"
          '';
        };
      };
    };
  };

~/.pi/agent/models.json (fixed)

{
  "providers": {
    "optimus": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "ollama",
      "models": [
        {
          "id": "local",
          "name": "GLM-4.7-Flash Q8_0 (optimus)",
          "aliases": ["fast", "glm-flash"],
          "reasoning": true,
          "contextWindow": 131072,
          "maxTokens": 32768,
          "cost": { "input": 0, "output": 0, "cacheRead": 0, "cacheWrite": 0 }
        },
        {
          "id": "smart",
          "name": "GPT-OSS-120B (optimus)",
          "aliases": ["gpt-oss-120b", "oss"],
          "reasoning": true,
          "contextWindow": 65536,
          "maxTokens": 16384,
          "cost": { "input": 0, "output": 0, "cacheRead": 0, "cacheWrite": 0 }
        }
      ]
    }
  }
}

I know it's not exactly what you're asking for, but I've been using qwen3.8 flash-next, glm 5.3 flash, deepseek v4 flash, v4 pro and gemma-4 26b-a4b through Cortecs (openrouter alternative) and they've surprised me by how good they are for most tasks. I still have access to big models like Sonnet and so on but I rarely reach out for them.

If you're not completely ready to abandon the big models, perhaps try something like Cortecs or Openrouter as a middle ground? You can still self-host the smaller ones and reach out and use something like Sonnet through them when needed.

Some providers have strict data retention policies, are powered by green energy and so on... I use open-webui front-end to interact with Cortecs.

GLM 5.3 flash if i remember correctly had even some benchmarks that even beat Opus. I know there might be some benchmaxxing going on but still. And it's really cheap.

Is GLM too slow or not capable enough? It's quite a big model, if the runtime is setup correctly you should have good quality but getting a throughput similar to cloud models is going to be difficult. Did you give Qwen3.8-27B-FP8 a try?

Also you mentioned Pi as your harness, but what are you using as the inference backend?

If your spec is really what your computer is, you should be able to run the models you need to.

Just slow down, have patience, and be opinionated on what you want it to look like, from code to architecture to user experience.

I hit more failure modes than when I use Claude, but there are always failure modes, and the answer is always slowing down, learning what it's doing, and having an opinion.

I will use, sometimes, free web search tools to architect, plan, and generally get ideas. Their ability to search the web directly helps. And to code, I sometimes build a cheap RAG with the language documentation (nushell needs help most of the time) and that helps.

Again, in general, just slow down and take your time. Like a jr dev, just cut it into bite-sized pieces with clear goals, all of which a more capable model can help with.

GLM worked well and did some impressive one-shots. Gemma can be more sophisticated in what it writes, but Qwen is less acrobatic and stable. Ornith looks interesting but I was never able to run it.

Apparently you just need to give them a good harness to use tools like write code and then debug it, that makes it way more efficient.

And please run them inside a VM or at least a container sandbox, not just unsandboxed on your machine.

ah thanks :)

I've had great success, but I have a pretty beefy gpu: radeon pro w7900 (48GB vram).

Definitely not as fast or good as frontier models, but quite good! I'm currently using qwen3.8-27b, but have used qwe3.6-35b-a3b and other models (including using openrouter and opencode go, but you asked about self-hosting).

Also, FYI, I use opencode.