• froztbyte@awful.systems
    link
    fedilink
    English
    arrow-up
    5
    ·
    20 hours ago

    a’ight, sure bub, let’s play

    tell me what hw spec I need to deploy some kind of interactive user-facing prompt system backed by whatever favourite LLM/transformer-model you want to pick. idgaf if it’s llama or qwen or some shit you’ve got brewing in your back shed - if it’s on huggingface, fair game. here’s the baselines:

    • expected response latencies: human, or better
    • expected topical coherence: mid-support capability or above
    • expected correctness: at worst “I misunderstood $x” in the sense of “whoops, sorry, I thought you were asking about ${foo} but I answered about ${bar}”; i.e. actual, contextual, concrete contextual understanding

    (so, basically, anything a competent L2 support engineer at some random ISP or whatever could do)

    hit it, I’m waiting.