{
  "video": {
    "id": "AVMr9PMINyo",
    "title": "Agents at Scale: Inside MiniMax's Model and the Infrastructure Behind It — Dan Fu and Olive Song",
    "duration": 1214,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer",
    "event_format": "In person"
  },
  "analysis": {
    "video_id": "AVMr9PMINyo",
    "title": "Agents at Scale: Inside MiniMax's Model and the Infrastructure Behind It — Dan Fu and Olive Song",
    "one_liner": "A model creator (MiniMax's RL lead Olive Song) and an inference optimizer (Together AI's VP of Kernels Dan Fu) walk through what happens to an open-weight model like MiniMax M3 — trained multimodal from scratch with sparse attention and 1M context — between the day it drops and the day it's fast.",
    "summary": "The panel traces one open model across the two halves of its life: Olive Song on post-training M3 — multimodal from step zero, RL in designed environments for kernel-writing and 12-hour paper-reproduction tasks, and evaluation that has to detect reward hacking — and Dan Fu on what Together AI does the moment early architecture details land, writing and benchmarking kernels for MiniMax sparse attention before launch and then grinding on KV cache, attention kernels and quantization for weeks after. Both argue the open-source frontier is effectively caught up, pointing at M3, GLM and Kimi. A recurring theme is that agentic workloads — whole codebases uploaded, hundreds of multi-turn tool calls — are a different optimization, routing and KV-cache problem than the chat workloads the stack was built for.",
    "key_points": [
      "MiniMax M3 is multimodal where the M2 series was not, and was trained multimodal from scratch — text and image data from step zero. Olive says it's normal at other labs for the model to collapse partway into that kind of training; MiniMax solved it. The payoff shows in the attention map: text tokens attend to visual tokens, so the modalities are 'naturally combined'.",
      "Hidden-gem uses Olive calls out: computer use (the model navigating a computer and using its tools) and game development — 'the model can help you develop a real cool games'.",
      "Together AI released a benchmark called parallel kernel bench, deliberately stocked with unsolved problems found by surveying every way people serve model inference. They discovered many obvious model speedups have no good kernel. The anti-bench-maxing stance is inverted on purpose: if you overfit to it, Together takes your kernels and uses them.",
      "Day-zero inference work starts before launch, from early architecture details shared by MiniMax — for M3, the MiniMax sparse attention. Together writes/benchmarks kernels, decides what to reuse versus write from scratch, checks launch-day quality, then works a pre-written list (KV cache, attention kernels, quantization) so the model measurably speeds up between day 0, day 7 and day 14.",
      "The workload shift from chat to agentic changes what to optimize: a chat workload is a few-thousand-token system prompt plus logs, while a coding agent uploads your whole codebase — 'a very different optimization and routing and kernel challenge'.",
      "M3's blog highlighted SVG bench, kernel bench and OS World, plus a 12-hour run reproducing an ICLR paper. Olive says the key to training such long-horizon tasks is environments, data, reward formulation, and modifying the RL algorithm for efficiency — and that the tasks themselves need GPUs, so they carry hardware constraints.",
      "Evaluation over long-horizon runs uses the model's multiple submissions as intermediate signals, with validation/test to catch cases where the model hacks rather than genuinely improves. MiniMax also builds internal evals out of its own 'self-evolution' work — using the model to speed up its own development.",
      "Serving 500k–1M-token concurrent contexts means the KV cache becomes, in Dan's framing, a distributed file system or very big database: where to store it, whether you've seen it before, how to fetch it, how to move it between places.",
      "Sparse attention lessons transfer across models — MiniMax's differs from DeepSeek's and GLM's, but the kernel-writing and optimization process carries over."
    ],
    "takeaways": [
      "If you serve or fine-tune open models, treat day-zero performance as a floor, not a number: the same endpoint is materially faster weeks later as kernels, KV cache and quantization get worked.",
      "Design your benchmark so that overfitting to it produces artifacts you actually want — parallel kernel bench is built from unsolved problems whose solutions Together will ship.",
      "For long-horizon agentic RL, put the effort into the environment, the data and the reward formulation rather than the algorithm alone — and build validation/test splits specifically to detect the model hacking the reward.",
      "Stop optimizing your inference stack for chat shapes. Profile the agentic pattern — full codebase in context, hundreds of tool-call turns — because KV cache strategy, routing and prompting all change.",
      "Build internal evals out of your own team's real work (the 'self-evolution' loop) rather than relying only on public benchmarks."
    ],
    "topics": [
      "agents",
      "inference",
      "kernels",
      "reinforcement-learning",
      "open-source-models",
      "multimodal",
      "kv-cache",
      "evals"
    ],
    "tools": [
      "MiniMax M3",
      "MiniMax M2",
      "MiniMax 2.5",
      "MiniMax 2.7",
      "Together AI",
      "parallel kernel bench",
      "KernelBench",
      "SVG bench",
      "OS World",
      "DeepSeek",
      "GLM",
      "Kimi",
      "ICLR"
    ],
    "quotes": [
      {
        "text": "One of our intentions with this benchmark was that if you overfit to it, that's great cuz we'll go take those kernels and use them to accelerate the inference and the development.",
        "at": "06:53",
        "url": "https://www.youtube.com/watch?v=AVMr9PMINyo&t=413s"
      },
      {
        "text": "And I meant over the last month and she said, oh, did you mean from last night? And this is the pace at which these guys work.",
        "at": "09:01",
        "url": "https://www.youtube.com/watch?v=AVMr9PMINyo&t=541s"
      },
      {
        "text": "It's pretty simple in theory. It's like the type of thing that you should have done in your third year of undergrad or something like that, but most of us actually skipped that class, so now we're rediscovering it live in industry.",
        "at": "16:47",
        "url": "https://www.youtube.com/watch?v=AVMr9PMINyo&t=1007s"
      },
      {
        "text": "If someone tells me you can't do the thousand first thing, yeah, like I don't know, try harder.",
        "at": "12:28",
        "url": "https://www.youtube.com/watch?v=AVMr9PMINyo&t=748s"
      }
    ],
    "words": 4587
  },
  "summary_url": "/#AVMr9PMINyo",
  "transcript": {
    "html": "/transcripts/AVMr9PMINyo.html",
    "txt": "/transcripts/AVMr9PMINyo.txt",
    "vtt": "/transcripts/AVMr9PMINyo.vtt"
  }
}