AI Agents: Better Than Humans? Cognitive Architectures for Language Agents (CoALA). μAgents. Against LLM Maximalism. XTTS Voice Cloning. AI Forecasting. NeXt-GPT Any2any Multimodal LLM. phi-1 1.3B
Cognitive models for AI can only be problem-specific. Problems, however, are often culturally defined, and the AI needs to be able recognize and relate to a human speaker's stance. Beyond that, implications and inferences must be "translated" into a singly symbolic domain. Only a huge number of situational corpus and AI analysis can come close to relatively acceptable results, but variables change and even the ancient method of flagging by human operators is not foolproof, because conceptual segmentation is also problem-specific, while the scope and nature of the same problems can also vary by the context.
Hey Matt - thanks for sending the link, very interesting. It does seem to support my thesis :-)
There's research from Harvard/ MIT/ Wharton unis working with business indicating the same results. It’s early days and there’s a bit of debate on what evaluation methods & benchmarks we use, and how.
Most orgs try to automate the human agents with scripts, templates, etc for interacting with humans and resolving tasks derived from the chat. But being an automated human is not a human thing.
Most probably, the AI agents -with the frameworks I mentioned- and properly finetuned with instruction-task oriented datasets, and a bit of RLHF (soon RLAIF) will soon beat most human agents at resolving most customer queries.
Cognitive models for AI can only be problem-specific. Problems, however, are often culturally defined, and the AI needs to be able recognize and relate to a human speaker's stance. Beyond that, implications and inferences must be "translated" into a singly symbolic domain. Only a huge number of situational corpus and AI analysis can come close to relatively acceptable results, but variables change and even the ancient method of flagging by human operators is not foolproof, because conceptual segmentation is also problem-specific, while the scope and nature of the same problems can also vary by the context.
I hope, I am not repeating the evident. :)
Loved this post and many useful links as always. Thanks
AI agents are starting to beat human agents in many areas already. This link is a short article about a trial we did a few weeks ago for Financial Guidance on UK Pensions queries. Our LLM AI Chatbot makes 5-10x FEWER ERRORS than typical human agents of financial services companies. So very much supporting your thesis. https://medium.com/@mattgosden/trial-results-show-that-ai-answers-financial-queries-better-than-humans-6df77bbcab31
Hey Matt - thanks for sending the link, very interesting. It does seem to support my thesis :-)
There's research from Harvard/ MIT/ Wharton unis working with business indicating the same results. It’s early days and there’s a bit of debate on what evaluation methods & benchmarks we use, and how.
Most orgs try to automate the human agents with scripts, templates, etc for interacting with humans and resolving tasks derived from the chat. But being an automated human is not a human thing.
Most probably, the AI agents -with the frameworks I mentioned- and properly finetuned with instruction-task oriented datasets, and a bit of RLHF (soon RLAIF) will soon beat most human agents at resolving most customer queries.