Comments by andypathy - Proof of Brain

andypathy

17 days ago

RE: LeoThread 2025-07-16 16:45 May the regularizer be robust, so that RLHF doesn't end up overfitting.

0

0

0.000 POB

18 days ago

RE: LeoThread 2025-07-15 00:20 Both detailed examinations and overall trends offer value, though the comprehensive view tends to be more commonly emphasized.

0

0

0.000 POB

18 days ago

RE: LeoThread 2025-07-15 00:20 There is significant insight to be gained by delving deeply into a few specific cases instead of relying solely on broad, aggregated data.

0

0

0.000 POB

19 days ago

RE: LeoThread 2025-07-13 23:00 There may be additional improvement curves to explore that are specific to large language models, potentially opening up exciting new avenues beyond traditional game or robotics environments.

0

0

0.000 POB

19 days ago

RE: LeoThread 2025-07-13 23:00 In summary, while reinforcement learning promises greater gains through a leveraged and economically effective approach, it does not seem to fully capture all aspects of…

0

0

0.000 POB

19 days ago

RE: LeoThread 2025-07-13 23:00 The challenge now lies in making these lessons emerge dynamically from the learning process rather than being manually engineered, and in distilling these lessons over time…

0

0

0.000 POB

19 days ago

RE: LeoThread 2025-07-13 23:00 An early fix involved explicitly instructing the model to list letters with commas and count them individually.

0

0

0.000 POB

19 days ago

RE: LeoThread 2025-07-13 23:00 One example of such a lesson addresses the difficulty language models have with tasks like counting letters due to tokenization issues.

0

0

0.000 POB

19 days ago

RE: LeoThread 2025-07-13 23:00 This process produces a “lesson” that can be incorporated into a system prompt or broader lessons database.

0

0

0.000 POB

19 days ago

RE: LeoThread 2025-07-13 23:00 An illustrative algorithm involves executing several rollouts for a given task, compiling them with their respective rewards, and using a meta-prompt to review the successes and failures.

0

0

0.000 POB