Comment by camillomiller
6 months ago
>> Developers haven't even started extracting the value of LLMs with agent architectures yet.
What does this EVEN mean? Do words have any value still, or are we all just starting to treat them as the byproduct of probabilistic tokens?
"Agent architectures". Last time I checked an architecture needs predictability and constraints. Even in software engineering, a field for which the word "engineering" is already quite a stretch in comparison to construction, electronics, mechanics.
Yet we just spew the non-speak "Agentic architectures" as if the innate inability of LLMs in managing predictable quantitative operations is not an unsolved issue. As if putting more and more of these things together automagically will solves their fundamental and existential issue (hallucinations) and suddenly makes them viable for unchecked and automated integration.
This means I believe we currently underuse LLM capabilities and their empirical nature makes it difficult to assess their limitations without trying. I've been studying LLMs from various angles during a few months before coming to this conclusion, as an experienced software engineer and consultant. I must admit it is however biased towards my experience as an SME and in my local ecosystem.
Hallucinations might get solved by faster, cheaper and more accurate, vision and commonsense-physics models. Hypothesis: Hallucinations are a problem only because physical reality isn't text. Once people switch to models that predict physical states instead of missing text, then we'll have domestic robots and lower hallucination rates.
Where is the training data for that? LLMs work because we already had tons of text that could be obtained cheaply. Where is the training data for physical reality?