3 ms·

I think part of the problem is due to scientists and philosophers speculating as they do and regular folks taking them at their word. This causes an over-estima

I think part of the problem is due to scientists and philosophers speculating as they do and regular folks taking them at their word. This causes an over-estimation of the abilities of current AI. Nick Bostrom Says AI Chatbots May Have Some Degree of Sentience: More and more people are saying this. https://futurism.com/the-byte/nick-bostrom-ai-chatbot-sentience https://futurism.com/the-byte/nick-bostrom-ai-chatbot-sentie...

fnordpiglet 4y ago
I think folks on our side of the equation are underestimating the value of generative AI by focusing on its obvious flaws. But the flaws (reasoning, goal formation, agency, optimization, information retrieval) are all well established sciences with excellent models and algorithms behind them. What they’ve lacked is the ability to do semantic synthesis in abstract spaces, which LLM do quite well. Use a multimodal AI and present it with a picture of a bar with cowboys playing poker. Pretend to be a goal based agent and assert you need $5 to the LLM. It will analyze the image and discern the is a poker table and you can make money from the poker game. It can describe in detail and annotate the image where to go. This can inform a navigation model where to go, and the navigation system can optimize and execute a route. Once there it can take images of the poker game and the LLM can describe the hand and the state of the game to a poker playing agent. The agent can tell the LLM moves to make, and the LLM can convert that into instructions for the navigation system. The LLM itself can’t create agency, goals, reliably win at poker, navigate, manage control systems, etc. But it can very well “glue” these traditional AI systems in a novel way that’s never been possible to date. Btw, I tried this as an experiment (with me being the traditional agents). By not delegating the parts that require agency, reasoning, calculation, or optimization to the LLM, I was more than able to prove to myself this approach is entirely feasible. With finetuning I was able to get it to output popular encodings as well so the input and output was properly structured.

fnordpiglet 4y ago

I think folks on our side of the equation are underestimating the value of generative AI by focusing on its obvious flaws. But the flaws (reasoning, goal formation, agency, optimization, information retrieval) are all well established sciences with excellent models and algorithms behind them. What they’ve lacked is the ability to do semantic synthesis in abstract spaces, which LLM do quite well. Use a multimodal AI and present it with a picture of a bar with cowboys playing poker. Pretend to be a goal based agent and assert you need $5 to the LLM. It will analyze the image and discern the is a poker table and you can make money from the poker game. It can describe in detail and annotate the image where to go. This can inform a navigation model where to go, and the navigation system can optimize and execute a route. Once there it can take images of the poker game and the LLM can describe the hand and the state of the game to a poker playing agent. The agent can tell the LLM moves to make, and the LLM can convert that into instructions for the navigation system. The LLM itself can’t create agency, goals, reliably win at poker, navigate, manage control systems, etc. But it can very well “glue” these traditional AI systems in a novel way that’s never been possible to date. Btw, I tried this as an experiment (with me being the traditional agents). By not delegating the parts that require agency, reasoning, calculation, or optimization to the LLM, I was more than able to prove to myself this approach is entirely feasible. With finetuning I was able to get it to output popular encodings as well so the input and output was properly structured.