>Agent-2, more so than previous models, is effectively “online learning,” in that it’s built to never really finish training. Every day, the weights get updated to the latest version, trained on more data generated by the previous version the previous day.
That seems to me to be a natural progression, from discrete models to models that are just continuously improved. Maybe we'll end up with different models with different rates of improvement rather than static differences in performance, and methodologies for that improvement will be the thing we care about. Maybe over time, even benchmark tests will be primarily concerned with that kind of efficiency.
I think a huge debate right now is the relative value of the "frontier" models from Western companies at the cutting edge, vs distilled versions of those models that are good enough and exponentially cheaper coming from China. But a paradigm of 'always training' means an always active, always advancing frontier, which is a stronger moat than a one-off model that's more advanced for a few months.
I wonder when the people will get that intelligence is not only directed at the external, but only really starts when you look at the internal (joy, pleasure, traumas, taboos, awkwardness, abuse etc.). Look up the word "interoception".
That seems to me to be a natural progression, from discrete models to models that are just continuously improved. Maybe we'll end up with different models with different rates of improvement rather than static differences in performance, and methodologies for that improvement will be the thing we care about. Maybe over time, even benchmark tests will be primarily concerned with that kind of efficiency.
I think a huge debate right now is the relative value of the "frontier" models from Western companies at the cutting edge, vs distilled versions of those models that are good enough and exponentially cheaper coming from China. But a paradigm of 'always training' means an always active, always advancing frontier, which is a stronger moat than a one-off model that's more advanced for a few months.
In other words: bias. Tons and tons of self-congratulatory, glue sniffing bias.