Bonus Panel

Transcript

[A man in a suit stands on stage, arms outstretched before a crowd]

Suit: And with AI we won’t need apps anymore. We’ll have a single terminal from which you can find anything you want!

[A man in the crowd shouts back]

Heckler: That used to be called the internet!

[The man on stage shouts, his arms raised in anger. His robot minions rush after the heckler, drones fly after the heckler overhead as the heckler flees]

Suit: Seize him!

[Bonus Panel: The author shouts at the stage]

Author: Bring back Memepool!

Source

  • fonix232@fedia.io
    link
    fedilink
    arrow-up
    31
    arrow-down
    1
    ·
    3 days ago

    On the topic of AI… I’ve found that smaller, self hosted models are actually quite adept at helping you search better. I’m not talking about replacing a search engine with an AI interface (like e.g. Perplexity did), but when you struggle to find relevant hits, throwing the query into a mini LLM (3-4B params), one you can even run on your phone, can be super helpful in refining the query itself, including further keywords you might not have thought of. Again, not talking about using a cloud hosted big model that burns through ten times more kwh to give a worse result.

    • Doublenut@lemmy.zip
      link
      fedilink
      arrow-up
      7
      ·
      3 days ago

      Do you have any examples of such models? Models like this are the only interest I have in LLMs.

      • fonix232@fedia.io
        link
        fedilink
        arrow-up
        6
        arrow-down
        1
        ·
        3 days ago

        I’ve been mainly using gemma4:e4b - interestingly, even though it’s supposed be a subset/lesser variant of Gemini for self-hosting, its overall approach in reasoning and iterations over search terminology and query optimisation is much, much better than what you get from Gemini.

        And to be fair I also quite like Perplexity - its ability to find relevant information from multiple sources, find both supporting and counter arguments where relevant, and build quite precise summaries. I do wish it was DIY-able, but with all the malicious actors I don’t feel comfortable allowing any LLM to just browse the net, who knows, even with properly secured tool calling, prompt injection can happen.