You can set up an MCP server in front of the LLM so that it can reach out to external APIs, like news sites/feeds/etc…
But, the majority of its behavior is going to be from its original training set, which is likely 12+ months old at this point. The pipeline for this is so long to generate and refine a good new AI model, that your data sets are constantly out of date.
Duckduckgo’s AI search seems to slightly agree with you, meaning Futurism and Engadget:
Grok, the AI chatbot, initially spread misinformation about Charlie Kirk’s death by claiming he survived the shooting and that videos of the incident were fake. This confusion stemmed from the chatbot’s inability to accurately process breaking news and conflicting information, leading to a series of incorrect statements before eventually acknowledging Kirk’s death.
That means AI really, really sucks and can be manipulated easily.
I think it’s well known at this point that grok in particular has been designed to be easy to manipulate by deliberately keeping it in the dark and feeding it only select information so that Musk can make it say what he wants.
Yes, LLMs, or what people call AI, are absolutely manipulated easily. Just the way you phrase your question can steer it to answer in a particular way. I haven’t been on Twitter in a long time, but I hopped on yesterday and today to check out all of the Kirk memes.
I saw so many comments from people both happy and upset with Kirk dying, that were asking questions in manipulative ways (to the llm) to try and get the response from grok they want.
LLMs are indeed horrible for live or recent events, and more importantly horrible for anything that is super important to not get wrong.
Don’t get me wrong I personally find llms useful and I use open source models occasionally for tasks they are better at, for me typically that means reformatting or compiling shorter notes from documents. Nothing super critical.
Good point, but wouldn’t it be time sensitive? Meaning, it would give more weight to recent events? I’m going with the “kirk is always right theory.”
You can set up an MCP server in front of the LLM so that it can reach out to external APIs, like news sites/feeds/etc…
But, the majority of its behavior is going to be from its original training set, which is likely 12+ months old at this point. The pipeline for this is so long to generate and refine a good new AI model, that your data sets are constantly out of date.
An MCP server will only get you so far.
Duckduckgo’s AI search seems to slightly agree with you, meaning Futurism and Engadget:
That means AI really, really sucks and can be manipulated easily.
I think it’s well known at this point that grok in particular has been designed to be easy to manipulate by deliberately keeping it in the dark and feeding it only select information so that Musk can make it say what he wants.
Yes, LLMs, or what people call AI, are absolutely manipulated easily. Just the way you phrase your question can steer it to answer in a particular way. I haven’t been on Twitter in a long time, but I hopped on yesterday and today to check out all of the Kirk memes.
I saw so many comments from people both happy and upset with Kirk dying, that were asking questions in manipulative ways (to the llm) to try and get the response from grok they want.
LLMs are indeed horrible for live or recent events, and more importantly horrible for anything that is super important to not get wrong.
Don’t get me wrong I personally find llms useful and I use open source models occasionally for tasks they are better at, for me typically that means reformatting or compiling shorter notes from documents. Nothing super critical.
ddg doesn’t run it’s own llm, they’re just a frontend to chatgpt that (allegedly) strips out all the tracking.
So… just like people?