4 ms·

Are AI Agents Just Hallucinations in a Digital Dialogue?

Are AI agents overrated and ultimately just hallucinations that are forwarded and used as input to hallucinate even more? I notice that people describe AI agents, but also the system prompts, as if they were talking to a human being. They use polite phrases, repeat themselves, some even claim that being polite to an AI created better results etc… I say it works like this: An agent receives a task written in human-readable form with politeness and such. Then it creates some other human-readable statement and sends it to another AI. Every time there is some hallucination. If the end result is not totally off, the inventors say: it works. It bugs me that I am yet again the only one to address this, everyone else in those low-quality videos on youtube praises ai agents just like the next open source model that now has even more parameters. It feels like those people in praising videos like “Two minute papers” or similar, they are not even real people, at least they are pathologically autistic.

daemon_9009 2y ago
I cannot make out anything from this. what are you trying to say? AI agents's fundamental task is to implement stuff, or in technical terms tool calling, computer use etc. they take input from user in natural language, and everything is perfectly defined in the prompt. The LLM is instructed to ask whenever not sure, so bootstrapping from hallucinations only happens when the implementation of AI agent is poor.
- Haeuserschlucht 2y ago
  User: What's the weather in Chicago? Layer1(LLM) to Layer2(AgentWeather): Can you tell the weather in chicago? AgentWeather: I have a tool that's called get_weather. Should I use this? No -> AgentWeather -> LLM: I cannot tell. LLM -> User: I cannot tell. Yes -> AgentWeather -> LLM: The air temperature is 33°C. LLM -> User: The air temperature in Chicago is 33°C. One example out of billions why it could go wrong. Layer1, 2, ..., n, they are all unsupervised language models and they "communicate" via text that's converted into tokens. What could go wrong?

daemon_9009 2y ago

I cannot make out anything from this. what are you trying to say? AI agents's fundamental task is to implement stuff, or in technical terms tool calling, computer use etc. they take input from user in natural language, and everything is perfectly defined in the prompt. The LLM is instructed to ask whenever not sure, so bootstrapping from hallucinations only happens when the implementation of AI agent is poor.

Haeuserschlucht 2y ago
User: What's the weather in Chicago? Layer1(LLM) to Layer2(AgentWeather): Can you tell the weather in chicago? AgentWeather: I have a tool that's called get_weather. Should I use this? No -> AgentWeather -> LLM: I cannot tell. LLM -> User: I cannot tell. Yes -> AgentWeather -> LLM: The air temperature is 33°C. LLM -> User: The air temperature in Chicago is 33°C. One example out of billions why it could go wrong. Layer1, 2, ..., n, they are all unsupervised language models and they "communicate" via text that's converted into tokens. What could go wrong?