Gagner de l'argent sur Internet, c'est possible !
Découvrez et partagez des méthodes pour générer des revenus en ligne : sondages, cashback, freelancing, affiliation et micro-tâches.
Vous n'êtes pas connecté.
J'ai spent the last 3 weeks building a setup where n8n handles the intake from my freelance clients (email + simple web form), parses the briefs using a local LLM (Llama 3.1 70B on my homelab), generates first drafts with a fine-tuned Mistral, and then auto-formats the deliverables into the client's specific templates. The whole point is to kill the "busywork" – the copying, pasting, formatting, and initial structuring – so I can focus on the actual high-value consulting.
The problem is, for tasks that require deep context or nuanced client communication, the hallucination rate is still annoyingly high, and I have to spend more time fixing errors than I save on labor. Is anyone else finding that the "efficiency gain" is actually a net negative unless you build extremely rigid validation steps? Or am I just bad at prompt engineering?
Hors ligne
Je lis ça et je me demande si tu es trying to build a business or just a tech demo. You talk about "net negative" and "fixing errors", which basically means you're stuck in the same rut as the manual process, except now you have to debug both your code and your AI hallucinations.
Honestly, most people on this forum are chasing the "set it and forget it" dream with LLMs, but the reality is that until you have a robust QA layer that doesn't cost more than the human labor you're trying to replace, you're just optimizing for failure. I'm currently testing similar pipelines for my own freelance work, specifically around the AI integration part, and I'm seeing the same thing: the "first draft" is often so structurally broken that rewriting it from scratch takes longer than just doing the work manually.
The hallucination rate isn't a prompt engineering issue, it's a fundamental limitation of using a 70B model for nuanced context without massive RAG overhead. If you're not spending weeks building rigid validation steps, you're not "automating", you're just creating a new layer of complexity that requires maintenance. I'd rather spend that maintenance time on a simpler affiliate setup or a more reliable AI tool that does one specific task well, than trying to build a Frankenstein machine that needs constant babysitting. What specific validation steps have you actually implemented that worked? Because if it's just "check the output", you're not going anywhere.
En ligne