Max Kreminski
@maxkreminski.bsky.social
also @sh1m.bsky.social and @jjyc.bsky.social!
December 13, 2025 at 1:59 AM
also @sh1m.bsky.social and @jjyc.bsky.social!
congratulations, super well deserved!
December 6, 2025 at 1:38 AM
congratulations, super well deserved!
poems are generated via the Savelost algorithm: given an input text, we repeatedly "compress" it by erasing a single letter at a time
a text embedding model is used to evaluate deletion candidates & choose the one that least disrupts the "meaning" of the original input
bsky.app/profile/maxk...
a text embedding model is used to evaluate deletion candidates & choose the one that least disrupts the "meaning" of the original input
bsky.app/profile/maxk...
🫠 new poetry machine: Savelost
given a sentence, Savelost repeatedly picks one letter to delete…
while trying to keep the new sentence as semantically similar as possible to the orig sentence
concept + algo by me, js port + frontend by @barrettrees.bsky.social
barrettrees.com/savelost
given a sentence, Savelost repeatedly picks one letter to delete…
while trying to keep the new sentence as semantically similar as possible to the orig sentence
concept + algo by me, js port + frontend by @barrettrees.bsky.social
barrettrees.com/savelost
November 30, 2025 at 10:52 PM
poems are generated via the Savelost algorithm: given an input text, we repeatedly "compress" it by erasing a single letter at a time
a text embedding model is used to evaluate deletion candidates & choose the one that least disrupts the "meaning" of the original input
bsky.app/profile/maxk...
a text embedding model is used to evaluate deletion candidates & choose the one that least disrupts the "meaning" of the original input
bsky.app/profile/maxk...
great paper! reminds me of some recent work in creativity support tools research on designing for "vertical movement" and "horizontal movement" between tools dl.acm.org/doi/abs/10.1...
Beyond the Artifact: Power as a Lens for Creativity Support Tools | Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology
dl.acm.org
November 25, 2025 at 11:45 PM
great paper! reminds me of some recent work in creativity support tools research on designing for "vertical movement" and "horizontal movement" between tools dl.acm.org/doi/abs/10.1...
so if you want a relatively straightforward prompting-based method i think unfortunately some kind of combinatory approach that incorporates an outside source of randomness/control is still as good as it gets
November 25, 2025 at 1:12 AM
so if you want a relatively straightforward prompting-based method i think unfortunately some kind of combinatory approach that incorporates an outside source of randomness/control is still as good as it gets
yeah, everything better seems to involve either:
- complex sampling techniques – e.g., beam-searching for diverse strings in token sequence space dl.acm.org/doi/full/10....
- finetuning – e.g., "forcing diffuse distributions" arxiv.org/abs/2404.10859; our recent COLM work arxiv.org/abs/2503.17126
- complex sampling techniques – e.g., beam-searching for diverse strings in token sequence space dl.acm.org/doi/full/10....
- finetuning – e.g., "forcing diffuse distributions" arxiv.org/abs/2404.10859; our recent COLM work arxiv.org/abs/2503.17126
November 25, 2025 at 1:11 AM
yeah, everything better seems to involve either:
- complex sampling techniques – e.g., beam-searching for diverse strings in token sequence space dl.acm.org/doi/full/10....
- finetuning – e.g., "forcing diffuse distributions" arxiv.org/abs/2404.10859; our recent COLM work arxiv.org/abs/2503.17126
- complex sampling techniques – e.g., beam-searching for diverse strings in token sequence space dl.acm.org/doi/full/10....
- finetuning – e.g., "forcing diffuse distributions" arxiv.org/abs/2404.10859; our recent COLM work arxiv.org/abs/2503.17126
also check out Luminate! rough process:
1. ask LLM for list of "dimensions" that stories might vary along (eg tone, genre, protag)
2. ask LLM for list of "values" each dim could take (eg protag -> young boy, old woman…)
3. generate stories with diff value combinations
dl.acm.org/doi/full/10....
1. ask LLM for list of "dimensions" that stories might vary along (eg tone, genre, protag)
2. ask LLM for list of "values" each dim could take (eg protag -> young boy, old woman…)
3. generate stories with diff value combinations
dl.acm.org/doi/full/10....
Luminate: Structured Generation and Exploration of Design Space with Large Language Models for Human-AI Co-Creation | Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems
dl.acm.org
November 25, 2025 at 1:00 AM
also check out Luminate! rough process:
1. ask LLM for list of "dimensions" that stories might vary along (eg tone, genre, protag)
2. ask LLM for list of "values" each dim could take (eg protag -> young boy, old woman…)
3. generate stories with diff value combinations
dl.acm.org/doi/full/10....
1. ask LLM for list of "dimensions" that stories might vary along (eg tone, genre, protag)
2. ask LLM for list of "values" each dim could take (eg protag -> young boy, old woman…)
3. generate stories with diff value combinations
dl.acm.org/doi/full/10....
i think a big driver of this (and reliance on AI in “evaluative” situations broadly) is that the evaluator’s goal is often less “make the right call” than “avoid being responsible for making the *wrong* call”
October 29, 2025 at 7:43 PM
i think a big driver of this (and reliance on AI in “evaluative” situations broadly) is that the evaluator’s goal is often less “make the right call” than “avoid being responsible for making the *wrong* call”
i'm kind of a doomer about this too but the x-axis *does* seem to mysteriously end right around the year 2020
October 17, 2025 at 1:02 AM
i'm kind of a doomer about this too but the x-axis *does* seem to mysteriously end right around the year 2020
i think of these as “target callers” and have been contemplating ways to autodetect them (see thread) bsky.app/profile/maxk...
it's been ~10yrs since the *second* time Kathy Sierra was chased off the web
as always, her harassment was driven by a handful of "target callers": big accounts that pick out targets for a swarm of part-time volunteer harassers
if we want public social media, we have to solve target calling
as always, her harassment was driven by a handful of "target callers": big accounts that pick out targets for a swarm of part-time volunteer harassers
if we want public social media, we have to solve target calling
“why isn’t <person I really liked> posting anymore?”
the answer is pretty much always that people tore them to shreds for something they did not remotely deserve being torn to shreds for, and did it over and over until they decided they were happier offline
the answer is pretty much always that people tore them to shreds for something they did not remotely deserve being torn to shreds for, and did it over and over until they decided they were happier offline
October 6, 2025 at 10:05 AM
i think of these as “target callers” and have been contemplating ways to autodetect them (see thread) bsky.app/profile/maxk...
ya i’ve only gone a couple times but it’s so huge that i mostly just find it draining
September 5, 2025 at 8:19 AM
ya i’ve only gone a couple times but it’s so huge that i mostly just find it draining
tempted but not sure, soooo much travel coming up this fall. UIST + Portland + Northwestern for sure, plus maybe COLM EMNLP AIIDE NeurIPS. hbu
September 3, 2025 at 10:43 PM
tempted but not sure, soooo much travel coming up this fall. UIST + Portland + Northwestern for sure, plus maybe COLM EMNLP AIIDE NeurIPS. hbu
wait this is kinda profound tho. a trace of the true self remains in the false self
September 2, 2025 at 6:21 PM
wait this is kinda profound tho. a trace of the true self remains in the false self
yeah, the currently online version is correct as of the last time i checked! didn't see that correction language tho, points for internal consistency i guess
August 13, 2025 at 7:20 PM
yeah, the currently online version is correct as of the last time i checked! didn't see that correction language tho, points for internal consistency i guess
i’m biased bc Noah was one of my phd advisors, but i still think his take on “alternative”, “expansive”, and “inventive” approaches to relating mechanics and meaning (from How Pac-Man Eats) provides a good theoretical framework for addressing these questions manchestergamecentre.org/blog-db/2021...
How Pac-Man Eats (book review) — Manchester Game Centre
In this blog post, Man Met Game Centre member Rob Gallagher reviews How Pac-Man Eats by Noah Wardrip-Fruin.
manchestergamecentre.org
August 12, 2025 at 9:56 PM
i’m biased bc Noah was one of my phd advisors, but i still think his take on “alternative”, “expansive”, and “inventive” approaches to relating mechanics and meaning (from How Pac-Man Eats) provides a good theoretical framework for addressing these questions manchestergamecentre.org/blog-db/2021...