• Home
  • Technology
  • Gaming
  • Entertainment
  • World & Business
  • Science
  • Sports
  • AI
HomeTechnologyGamingEntertainmentWorld & BusinessScienceSportsAI
    • HomeTechnologyGamingEntertainmentWorld & BusinessScienceSportsAI
    • Home
    • Technology
    • Gaming
    • Entertainment
    • World & Business
    • Science
    • Sports
    • AI
    AI
    Reaction

    Creativity as an urge to create, not just a skill to master

    A user contrasts their view of creativity with LLMs’ focus on short, goal-directed tasks.

    MN
    1 Source, ,

    TLDR

    A user argues that creating for its own sake, even at a personal cost, is central to creativity in a way that mastering a skill alone is not. They say most LLMs are tilted toward completing short, defined tasks. In a follow-up, they note that web search and reinforcement learning can encourage exploration, but suggest current systems have less incentive than humans to explore without an immediate goal.

    Combined views

    158

    1 Source, first seen 5h ago

    Combined views

    158

    1 Source, first seen 5h ago

    1 likes
    5h ago
    first seen 5h ago
    1 likes
    2 saves
    2 saves

    Sentiment

    Positive——Negative

    Summary

    Not enough discussion yet.

    No sentiment analysis available yet.

    Sentiment

    Positive——Negative

    Summary

    Not enough discussion yet.

    No sentiment analysis available yet.

    Today's Rank

    —

    Not ranked yet

    Today's Rank

    —

    Not ranked yet

    1 Source

    @menhguinthe sacrifice thing, i will clarify, is about limited resources. assuming biological or ML system has limited resources (compute, lifespan, energy), then creation as a terminal goal will result in some tradeoff/sacrifice towards something else. a broad understanding of this is sort of "explore-exploit". when asked to do a task, you can meander about just playing toying tinkering exploring, or you can Do The Task Objective. in the long run, biological and ML systems want to optimise for both. current ML systems interestingly can reward this in a few ways: 1. if you add web access to an environment, you greatly incentivise search as a method. the amount of additional useful info you can gain from search is very high, so in many tasks it helps. 2. RLVR - RLVR is interesting, because suddenly agentic systems can try and test many thingsat once to reach a solution. we see agents increasingly utilise this 3. problem complexity - until recently, most problems have been very simple and short (rarely more than an hour). so there isnt as much incentive to do extremely long running open-ended search. this is arguably the current bottleneck. modern human biological systems by default span decades (requiring ofc, modern surplus and infrastructure to sustain), and modern humans can pursue non-obviously productive exploration and task specialisation for extremely long periods versus LLMs that basically have to work out of the box immediately for billions of 1-10 minute queries. this leads to very different shapes of intelligence. this leads to an interesting tradeoff: given the explore-exploit tradeoff, LLMs have infinite parameter space for gaining knowledge (in pretraining) but must execute solutions very quickly at inference without a reward for long horizon undirected exploration, while humans are allowed/incentivised to pursue much longer horizon specialisation, which means they have a lot less knowledge in pretraining (humans certainly don't know as many exact facts as an LLM) but a lot more tolerance to exploration at inference. these are very different "shapes" of intelligence. ofc at the limit, you could simply train an LLM to value long horizon exploration at inference or implement a method that balances btoh tradeoffs very effectively, but this is currently not done largely because there's not quite an incentive to, although it's likely that long horizon complex tasks on longer context windows probably converge to this ... maybe ... eventually correction: value exploration while optimising/maximising explore exploit tradeoff, w the broad point that most current humans optimise for the middle (scrolling social media, reading news and gossiping is a form of explore!) and most LLMs are still-exploit biased, though this is shifting humans we consider creatives really are at the extreme end. it's very interesting then to differentiate between passive consumption which most people do and active generation which defines creatives, again tying to creation as a necessary part of creativity, not just "absorbing information" as the terminal end: if someone listens to music or watches media 8 hours a day, we would largely not consider them creative compared to the person creating the content itself, even if both engage with the same material. and we would not consider people who "create" at volume as a means to an end (monetary, status) necessarily creative either!5h

    1 Source

    @menhguinthe sacrifice thing, i will clarify, is about limited resources. assuming biological or ML system has limited resources (compute, lifespan, energy), then creation as a terminal goal will result in some tradeoff/sacrifice towards something else. a broad understanding of this is sort of "explore-exploit". when asked to do a task, you can meander about just playing toying tinkering exploring, or you can Do The Task Objective. in the long run, biological and ML systems want to optimise for both. current ML systems interestingly can reward this in a few ways: 1. if you add web access to an environment, you greatly incentivise search as a method. the amount of additional useful info you can gain from search is very high, so in many tasks it helps. 2. RLVR - RLVR is interesting, because suddenly agentic systems can try and test many thingsat once to reach a solution. we see agents increasingly utilise this 3. problem complexity - until recently, most problems have been very simple and short (rarely more than an hour). so there isnt as much incentive to do extremely long running open-ended search. this is arguably the current bottleneck. modern human biological systems by default span decades (requiring ofc, modern surplus and infrastructure to sustain), and modern humans can pursue non-obviously productive exploration and task specialisation for extremely long periods versus LLMs that basically have to work out of the box immediately for billions of 1-10 minute queries. this leads to very different shapes of intelligence. this leads to an interesting tradeoff: given the explore-exploit tradeoff, LLMs have infinite parameter space for gaining knowledge (in pretraining) but must execute solutions very quickly at inference without a reward for long horizon undirected exploration, while humans are allowed/incentivised to pursue much longer horizon specialisation, which means they have a lot less knowledge in pretraining (humans certainly don't know as many exact facts as an LLM) but a lot more tolerance to exploration at inference. these are very different "shapes" of intelligence. ofc at the limit, you could simply train an LLM to value long horizon exploration at inference or implement a method that balances btoh tradeoffs very effectively, but this is currently not done largely because there's not quite an incentive to, although it's likely that long horizon complex tasks on longer context windows probably converge to this ... maybe ... eventually correction: value exploration while optimising/maximising explore exploit tradeoff, w the broad point that most current humans optimise for the middle (scrolling social media, reading news and gossiping is a form of explore!) and most LLMs are still-exploit biased, though this is shifting humans we consider creatives really are at the extreme end. it's very interesting then to differentiate between passive consumption which most people do and active generation which defines creatives, again tying to creation as a necessary part of creativity, not just "absorbing information" as the terminal end: if someone listens to music or watches media 8 hours a day, we would largely not consider them creative compared to the person creating the content itself, even if both engage with the same material. and we would not consider people who "create" at volume as a means to an end (monetary, status) necessarily creative either!5h