@HistoPol@mastodon.social @mattsheffield@mastodon.social That is not evidence, but simple conclusory statements. One of the principal failures of LLMs to fill the shoes of AGI is that they learn nothing. Agentic AI tries to patch this by having the LLM maintain a serial file of “memory” but it is just a LLM conversation prefix to prime the pump of autocomplete. Real AGI doesn’t just answer yes if it is asked if it thinks, it thinks and acts based on its desires. So it needs memory, it needs to color memories as good or bad, a trait we call sentience, and it needs to contemplate what justice and fairness mean ; currently every Agentic AI needs to be supervised as LLMs are trained to be prolix and sound confident at all times, even when wrong. So Agentic “decisions” are just a simulation of a conversation of humans who authored the content of its training data. Agentic AI makes mistakes no reasoning being would because there is no reasoning not present in its training data. Conversations go sideways because LLMs associate Xur with The Last Starfighter because those words came close together in training data, not because they had original opinions about a film it never saw.