You can’t predict how the next twenty years of research improves on the current techniques because we haven’t done the research.
Is it going to be specialized agents? Because you don’t need a lot of data to do one task well. Or maybe it’s a lot of data but you keep getting more of it (robot movement? stock market data?)
We do already know about model collapse though, genai is essentially eating its own training data. And we do know that you need a TON of data to do even one thing well. Even then it only does well on things strongly matching training data.
Most people throwing around the word agents have no idea what they mean vs what the people building and promoting them mean. Agents have been around for decades, but what most are building is just using genai for natural language processing to call scripted python flows. The only way to make them look coherent reliably is to remove as much responsibility from the llm as possible. Multi agent systems are just compounding the errors. The current best practice for building agents is “don’t use a llm, if you do don’t build multiple”. We will never get beyond the current techniques essentially being seeded random generators, because that’s what they are intended to be.
You can’t predict how the next twenty years of research improves on the current techniques because we haven’t done the research.
Is it going to be specialized agents? Because you don’t need a lot of data to do one task well. Or maybe it’s a lot of data but you keep getting more of it (robot movement? stock market data?)
We do already know about model collapse though, genai is essentially eating its own training data. And we do know that you need a TON of data to do even one thing well. Even then it only does well on things strongly matching training data.
Most people throwing around the word agents have no idea what they mean vs what the people building and promoting them mean. Agents have been around for decades, but what most are building is just using genai for natural language processing to call scripted python flows. The only way to make them look coherent reliably is to remove as much responsibility from the llm as possible. Multi agent systems are just compounding the errors. The current best practice for building agents is “don’t use a llm, if you do don’t build multiple”. We will never get beyond the current techniques essentially being seeded random generators, because that’s what they are intended to be.