Liu Grey
@liu-eroteme.bsky.social
non-binary audhd person, background in compsci and phil, interested in AI and evolution of consciousness, doing MLops for a public transit company
Twitter: @Liu_eroteme
(she/they)
Twitter: @Liu_eroteme
(she/they)
Reposted by Liu Grey
imo o1 and the like do best if you clearly lay out how things are at the moment and how things should be when done, and let them find the delta
e.g. codebase counts as is-state-representation, description of new feature as end-state-representation, models job is to find the delta between entropies.
e.g. codebase counts as is-state-representation, description of new feature as end-state-representation, models job is to find the delta between entropies.
January 19, 2025 at 7:54 PM
imo o1 and the like do best if you clearly lay out how things are at the moment and how things should be when done, and let them find the delta
e.g. codebase counts as is-state-representation, description of new feature as end-state-representation, models job is to find the delta between entropies.
e.g. codebase counts as is-state-representation, description of new feature as end-state-representation, models job is to find the delta between entropies.
obviously not a rigorous study, but..
describing the reasoning & generation as a markov process is pretty straightforward, the application of active inference makes sense, and this shouldn't be that hard to formalize and verify through some basic tests, no?
anyway, I'm off to bed now ^~^
describing the reasoning & generation as a markov process is pretty straightforward, the application of active inference makes sense, and this shouldn't be that hard to formalize and verify through some basic tests, no?
anyway, I'm off to bed now ^~^
January 19, 2025 at 8:14 PM
obviously not a rigorous study, but..
describing the reasoning & generation as a markov process is pretty straightforward, the application of active inference makes sense, and this shouldn't be that hard to formalize and verify through some basic tests, no?
anyway, I'm off to bed now ^~^
describing the reasoning & generation as a markov process is pretty straightforward, the application of active inference makes sense, and this shouldn't be that hard to formalize and verify through some basic tests, no?
anyway, I'm off to bed now ^~^
we know our action (coding), we know our current state (model with context, prompt) and current observations (codebase), we know our approximate posterior (description of finished code after execution in prompt) the models job is to bridge the gap, so to speak
(alpha nearly reflects logprob here)
(alpha nearly reflects logprob here)
January 19, 2025 at 8:12 PM
we know our action (coding), we know our current state (model with context, prompt) and current observations (codebase), we know our approximate posterior (description of finished code after execution in prompt) the models job is to bridge the gap, so to speak
(alpha nearly reflects logprob here)
(alpha nearly reflects logprob here)
it's like how, in a Markov process formulation of actinf, range of possible behavior depends on a ratio of two entropies: approximate posterior entropy of current action at current state and current sensory observations at current state
January 19, 2025 at 8:12 PM
it's like how, in a Markov process formulation of actinf, range of possible behavior depends on a ratio of two entropies: approximate posterior entropy of current action at current state and current sensory observations at current state
imo o1 and the like do best if you clearly lay out how things are at the moment and how things should be when done, and let them find the delta
e.g. codebase counts as is-state-representation, description of new feature as end-state-representation, models job is to find the delta between entropies.
e.g. codebase counts as is-state-representation, description of new feature as end-state-representation, models job is to find the delta between entropies.
January 19, 2025 at 7:54 PM
imo o1 and the like do best if you clearly lay out how things are at the moment and how things should be when done, and let them find the delta
e.g. codebase counts as is-state-representation, description of new feature as end-state-representation, models job is to find the delta between entropies.
e.g. codebase counts as is-state-representation, description of new feature as end-state-representation, models job is to find the delta between entropies.
pretty sure that's it questions wise .. I'd be very grateful if y'all could help me figure this out soon ^~^
.. I need to get away from twitter
.. I need to get away from twitter
December 12, 2024 at 10:03 AM
pretty sure that's it questions wise .. I'd be very grateful if y'all could help me figure this out soon ^~^
.. I need to get away from twitter
.. I need to get away from twitter
If 2\
where should I post my shit? substack?
or rather.: it's a subquestion:
do you want..
2.1\ an established site, and if so, which one?
2.2\ a self-hosted blog written from scratch, and if so, what are you thinking features-wise?
interactive visualizations?
where should I post my shit? substack?
or rather.: it's a subquestion:
do you want..
2.1\ an established site, and if so, which one?
2.2\ a self-hosted blog written from scratch, and if so, what are you thinking features-wise?
interactive visualizations?
December 12, 2024 at 10:03 AM
If 2\
where should I post my shit? substack?
or rather.: it's a subquestion:
do you want..
2.1\ an established site, and if so, which one?
2.2\ a self-hosted blog written from scratch, and if so, what are you thinking features-wise?
interactive visualizations?
where should I post my shit? substack?
or rather.: it's a subquestion:
do you want..
2.1\ an established site, and if so, which one?
2.2\ a self-hosted blog written from scratch, and if so, what are you thinking features-wise?
interactive visualizations?
If 1\
what would 'highly readable formatting' look like for you?
should I.. try to hit max characters on every post & minimize total blocks?
or should I try to chunk things into sensible paragraphs, even if it means more total blocks?
and how do I split? characters, words, sentences, paragraphs?
what would 'highly readable formatting' look like for you?
should I.. try to hit max characters on every post & minimize total blocks?
or should I try to chunk things into sensible paragraphs, even if it means more total blocks?
and how do I split? characters, words, sentences, paragraphs?
December 12, 2024 at 10:03 AM
If 1\
what would 'highly readable formatting' look like for you?
should I.. try to hit max characters on every post & minimize total blocks?
or should I try to chunk things into sensible paragraphs, even if it means more total blocks?
and how do I split? characters, words, sentences, paragraphs?
what would 'highly readable formatting' look like for you?
should I.. try to hit max characters on every post & minimize total blocks?
or should I try to chunk things into sensible paragraphs, even if it means more total blocks?
and how do I split? characters, words, sentences, paragraphs?