Thanks for sharing. The big differance between LLM based AI systems and the formerly machine learning approach, is that just by feeding large amounts of data(which as mentioned only became feasible due to cost/space reduction while achieving more compute) we gain unpredictable capabilities. Therefor i wouldn't quite say you were ahead of the time, but you were on the cutting edge of what at the time was possible. As seen in the past year, many gain of functions were not predicted, but just established themselves due to adding compute (140+ and counting). Suddenly without teaching explicitly systems they were "oh now it can translate into all languages" or "ooops now it can do math" Therefor a combined approach from formal reliable prooven scientific concepts, starting with physics and then building on that chemical ruelsets and then adding the datadriven approach from LLMs. Whereby the first approaches are the filters of the later. The LLMs though may have rulesets learned that maybe real, but we are unaware of, therefor it needs to be looked into where these two would collide. Meaning the data that would be filtered out, needs to be looked into seperatly to maybe find new principles we were unaware of. Like f,e, recently more and more news about quantum effects on biology, chemistry and such were reported(i am no expert in these areas so i assume the experts know more about what i address then i understand). Microsoft recently announded it intend to create with help of AI material science of the next 200 years within the next 10 years. Here my fear is that these findings would lead to a bottleneck in scientific research as these where seen as profitable will be immediatly patended and by that make those companies that are succesfully doing so first, dominance for the time of the patent restriction which is like upto 50 years if i am not missinformed. In an ideal world i would hope for a political solution to make all AI based research public domain as these systems, their principles are based on decades of research from many fields and not just those companies then making a buck in the end. Furthermore i think that our western model of patents may come to an end anyways. Somehow the media and politics always find it as a cause for blaming someone who stole such research and use it then successfully, instead the open source or free source/public domain approach as software development has shown seems more ideal. Historicly research was finannced by states and only due to our western capitalist systems more and more emphasis was put on companies then being already included into the research projects which lead often to what can be sold not necessarily what humanity needs or in some cases may also not be detremental to our health. What would be your thoughts on this?
1:02:30 concerning "black box systems", there are several ongoing attempts to address this. From Wolfram alphas aproach over a formalized language approach to newly established courses and research projects looking into gain of funtion implyed rulesets of gain of function due to adding compute(there was a specific term for this i am currently missing). The issue being that these now come after these systems already have become unpredictable strong and due to market pressures and strive for dominance are behind the curve leading to an not impossible scenario where an AGI comes into existance without us noticing and doing stuff that might be detrimental to our goals and even health if left uncchecked. Which is why since formost the release of GPT4, the calls for research in allignment and opening the black box were called for.
Thanks for sharing.
The big differance between LLM based AI systems and the formerly machine learning approach, is that just by feeding large amounts of data(which as mentioned only became feasible due to cost/space reduction while achieving more compute) we gain unpredictable capabilities.
Therefor i wouldn't quite say you were ahead of the time, but you were on the cutting edge of what at the time was possible. As seen in the past year, many gain of functions were not predicted, but just established themselves due to adding compute (140+ and counting). Suddenly without teaching explicitly systems they were "oh now it can translate into all languages" or "ooops now it can do math"
Therefor a combined approach from formal reliable prooven scientific concepts, starting with physics and then building on that chemical ruelsets and then adding the datadriven approach from LLMs. Whereby the first approaches are the filters of the later.
The LLMs though may have rulesets learned that maybe real, but we are unaware of, therefor it needs to be looked into where these two would collide. Meaning the data that would be filtered out, needs to be looked into seperatly to maybe find new principles we were unaware of. Like f,e, recently more and more news about quantum effects on biology, chemistry and such were reported(i am no expert in these areas so i assume the experts know more about what i address then i understand).
Microsoft recently announded it intend to create with help of AI material science of the next 200 years within the next 10 years. Here my fear is that these findings would lead to a bottleneck in scientific research as these where seen as profitable will be immediatly patended and by that make those companies that are succesfully doing so first, dominance for the time of the patent restriction which is like upto 50 years if i am not missinformed. In an ideal world i would hope for a political solution to make all AI based research public domain as these systems, their principles are based on decades of research from many fields and not just those companies then making a buck in the end.
Furthermore i think that our western model of patents may come to an end anyways. Somehow the media and politics always find it as a cause for blaming someone who stole such research and use it then successfully, instead the open source or free source/public domain approach as software development has shown seems more ideal. Historicly research was finannced by states and only due to our western capitalist systems more and more emphasis was put on companies then being already included into the research projects which lead often to what can be sold not necessarily what humanity needs or in some cases may also not be detremental to our health.
What would be your thoughts on this?
1:02:30 concerning "black box systems", there are several ongoing attempts to address this. From Wolfram alphas aproach over a formalized language approach to newly established courses and research projects looking into gain of funtion implyed rulesets of gain of function due to adding compute(there was a specific term for this i am currently missing). The issue being that these now come after these systems already have become unpredictable strong and due to market pressures and strive for dominance are behind the curve leading to an not impossible scenario where an AGI comes into existance without us noticing and doing stuff that might be detrimental to our goals and even health if left uncchecked. Which is why since formost the release of GPT4, the calls for research in allignment and opening the black box were called for.
❤❤