KI -Agenten falsch ~ 70% der Zeit: Carnegie Mellon Studie
https://www.theregister.com/2025/06/29/ai_agents_fail_a_lot/?td=rt-3a
KI -Agenten falsch ~ 70% der Zeit: Carnegie Mellon Studie
https://www.theregister.com/2025/06/29/ai_agents_fail_a_lot/?td=rt-3a
7 Kommentare
Unsurprising. This is the most pressing problem in LLM based AI right now and seems to be getting the least attention.
But hey, cool videos.
For all the hype around AI agents they simply don’t work or in some cases really even exist yet. Despite that companies are still selling these to unsuspected consumers and desperate middle managers trying to show their worth. One under discussed aspect of regulating this stuff is a need for a certification process to weed out the pure scams.
This is the main issue with LLMs. They’re good at specific tasks like coding small and specific tasks. But if there’s any nuance you will need to use your critical thinking skills and existing knowledge to correct the mistakes. When searching for stuff these models behave like someone taking google search results literally. If a wrong thing is high up on the results the LLM will present it as truth
So in case anyone misses the subtlety of the title, this is about agentic AI. Which is assigning AI tasks that it has to complete independently, these tasks often involve multiple steps, reasoning, and keeping track of what has already been done and what needs to be done. This is a new and developing set of abilities, which per this report and many others is an area LLMs are currently struggling in.
I wouldn’t feel too happy about the results, in 30% of cases an AI with a half baked version of agentic behavior, could replace a human. This technology is less than a year old, and will assuredly get better. Where will we be this time next year, 60% or more?
Humanity is on the clock, people will rightfully argue that getting AI to 100% might not happen soon, but no one is going to wait that long. As soon as agentic behavior hits accuracy rates above coin flip odds there are tasks they can start taking massive chunks out of our labor force.
What makes an Agent an Agent is the ability to reason, ask itself questions, the propose solutions to build out chains of systems and applications without human interaction.
Buttttttt there’s one giant flaw. We know LLMs produce garbage phantom data and results. So when an Agent uses an LLM as part of its process it still can’t distinguish a garbage or incorrect result before it moves onto the next step.
Basically, an Agent is only as good as the LLMs it relies on and those are still garbage themselves. Who cares if it can “reason” if it is never able to distinguish the bad data it is reading.
“Hmm… Yes. The floor is made of floor.” Moment. We’ve been seeing AI be wrong for a long time now.
That sounds about right for all but the easiest of questions I’ve asked it. Admittedly, I’ve only asked a small amount of coding questions and some vanilla PC questions probably considered difficult for most.
Literally, if you ask it a question that is specific, and uniquely different or opposite of from a very common question, it will answer the common question and pretend like it answered yours.
People who think AI is wizardry must be asking it stuff like “when does the WalMart across the street close?” Or “who was the 32nd president?”