with npc tools we are trying to build these kinds of continuously learning ensemble models so that you can build and grow them on your own machine iteratively
Very cool exercise, but my gut feeling is that you don't necessarily need much of a world model to do this? For my local paper, for instance, I can pretty reliably predict that most stories will either be a) violent crime, b) hurricane damages, or c) NIMBYs blocking/trying to block/complaining about failing to have blocked some new construction.
But I don't need to understand anything about those topics for to generate this production. All I need is a very crude model of my local paper.
Turns out its a bit more complicated than this, but its a good shout. Its also why I also chose a more heterogenous data feed, and later to include not just headlines but also articles.
I came across (https://nof1.ai/) today, and it got me wondering why DeepSeek has been dominating other models lately. Then I remembered a scene from the Dwarkesh podcast with Victor Shih (https://youtu.be/b1TeeIG6Uaw?si=_NS4QA5yiQ7ByZz3&t=2184) around the 36:28 mark, where he mentioned that DeepSeek was originally developed for high-frequency trading (HFT). It was specifically trained to analyze government announcements and respond as Chinese markets react to CCP policy changes. So I have a gut feeling that DeepSeek might perform better for your project.
Don’t know if I’ll get to it soon enough, my todo is piling up faster than I can type it out, but try swapping the openai api here (https://github.com/strangeloopcanon/foresight-forge) with the deepseek one? suspect is houldw ork
with npc tools we are trying to build these kinds of continuously learning ensemble models so that you can build and grow them on your own machine iteratively
https://github.com/npc-worldwide/npcpy
https://github.com/npc-worldwide/npc-studio
This looks neat!
Very cool exercise, but my gut feeling is that you don't necessarily need much of a world model to do this? For my local paper, for instance, I can pretty reliably predict that most stories will either be a) violent crime, b) hurricane damages, or c) NIMBYs blocking/trying to block/complaining about failing to have blocked some new construction.
But I don't need to understand anything about those topics for to generate this production. All I need is a very crude model of my local paper.
Turns out its a bit more complicated than this, but its a good shout. Its also why I also chose a more heterogenous data feed, and later to include not just headlines but also articles.
have you tried out this experiment with the deepseek model instead of chatgpt?
Tried Qwen but not Deepseek. Is it better?
I came across (https://nof1.ai/) today, and it got me wondering why DeepSeek has been dominating other models lately. Then I remembered a scene from the Dwarkesh podcast with Victor Shih (https://youtu.be/b1TeeIG6Uaw?si=_NS4QA5yiQ7ByZz3&t=2184) around the 36:28 mark, where he mentioned that DeepSeek was originally developed for high-frequency trading (HFT). It was specifically trained to analyze government announcements and respond as Chinese markets react to CCP policy changes. So I have a gut feeling that DeepSeek might perform better for your project.
Plausible!
if you try it , do share you insight !! :)
Don’t know if I’ll get to it soon enough, my todo is piling up faster than I can type it out, but try swapping the openai api here (https://github.com/strangeloopcanon/foresight-forge) with the deepseek one? suspect is houldw ork