@GhostOnTheHalfShell
@olibrendel
In my previous reply I said that the “oops I’m sorry” part is probably the response to being caught hallucinating. It’s a fact of how they work, and clearly worse with with smaller, less capable models.
That’s why grounding (enriching the context with your own documents) is critical, if you really want to use them for any research work. #DeepSeek discusses #Tiananmen massacre if you give it relevant sources (though maybe not DeepSeeks’s own hosted model).
The “I was programmed” part though seems to be product of the algorithm controlling the workflow, or logic. True to the American mindset of their developers, they’re programmed for both-side-ism. But since the model doesn’t have the data, it hallucinates it.
These tools are a disaster for research when used trivially by people who are not experts on the topic, don’t know how to guide the model using system prompts or ground it in facts (reliable diverse sources).
I’ve seen worse produced by #Perplexity for example, which is grounded in web searches. There the problem is the sources it prefers. If I ask it if #Israel is committing genocide in #Gaza, or occupying Palestinian lands, it’s very likely to use some Israeli sources to answer these questions, often the non peer reviewed Zionist Jewish virtual library which is very difficult to get rid of. To get it to quote any of the Khalidis or more contemporary research is hopeless.
When I asked llama3.2 if it’s conceivable that 2nd and 3rd generation Israeli holocaust survivors be involved in a genocide, it completely denied such a possibility due to Israel’s high moral standards, rule of law, adherence to international law etc … the data about Israel not respecting any UN resolution ever, apparently never been fed to any of these models.