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 passé les deux dernières semaines configurant un pipeline LLM pour auto-classing support emails in a dev shop. The pitch from the vendor was "zero human intervention, 99% accuracy". In practice, after the honeymoon phase, it's basically a glorified regex that hallucinates.
The problem isn't the classification itself, it's the confidence levels. The model is 95% sure that a user complaining about "my account is locked" is a billing dispute, so it routes it to Finance. When the user is actually locked out and waiting for a reset, you've just created a 4-day delay while the AI sits there thinking it's doing a good job. You end up with a "human review" queue that's basically a mountain of garbage where you have to click through every single misclassification because the model is confidently wrong in ways you didn't expect.
It's the same trap as every other "automation" scam. They sell you on the happy path, but they don't give a damn about the edge cases where the AI is stubbornly incorrect. Unless you're willing to hire a team just to babysit the bots, you're better off sticking to simple keyword filters. The latency cost of a human reading a sentence is negligible compared to the cost of debugging a hallucinated workflow that breaks your SLA. Don't get sold on the idea that a 400k context window solves the problem of understanding nuance. It doesn't.
Hors ligne