s its verbose239regression outputs contain large execution traces with sparse hypothesis-relevant content.240MetaScience resists condensation entirely, its compact, already-aggregated outputs leave241little noise to remove.
do some qualitative validation!
Performance varies substantially: crisis communication and264expense anomaly score highest (85, 82
lmao
10,000 unique person mentions
given that the vast majority of these are outside of your sample, how accurate is the task construction?
Stage 3: Interesting findings extraction. To build the culture layer, a separate week-147batched pass uses elevated temperature to identify 3–10 qualitatively interesting moments148per week—humor, prescience, fun facts, cultural artifacts—each with a catchy title, de-149scription, category, excitement rating (1–10), and source email references. Findings below150excitement level 6 are discarded, yielding 382 curated findings across 119 weeks.1514
.............
emoving emails381with bodies <50 characters
This is also insane! What about literally yes or no questions!
3)Temporal density—email provides daily or even hourly granularity over121months or years;
Yes, but do you establish how frequently? I want to see this data. Taking the corpus size of 345k for 150 employees over 4 years works out as ~1.6 emails/employee-day. Is the claim here genuinely that people's work is summarized by two emails a day?
2)Comprehensiveness—email captures strategic119discussions, operational coordination, social exchanges, and administrative processes within120a single medium;
This cannot possibly have evidence
five-layer model
what's the fifth layer!!
simplicity, mobil-ity, and immutability
Are these the properties that are just kinda solved now with nonsymbolic approaches?
odels to obsolesce
passive voice is interesting here. I'm also not entirely sure that all of this activity is nomadic.
esource e"ciency.
Is this true??
the ability to directlymanipulate model weights allows users to bypass safety training entirely, modifying models so that they no longerrefuse any request
open models may not even have safety training, in that they may just be base model releases, rather than RLHFd ones.
Some participants argued that moredata is necessary to capture the full range of cultural expressions, while others contended thatthe focus should be on developing thicker development pipelines that incorporate expertiseand context. They discussed the limitations of current models, which often operate on crudemetrics and may not adequately represent the richness of cultural data.
This more vs thicker data debate is a good one, but it is also begging the question - do we want to model all cultural variation? e.g. https://aclanthology.org/2025.naacl-long.273.pdf
They highlighted the importance of interdisciplinaryteaming at specific points in the development pipeline, imagining pairs of experts fromtechnical and qualitative fields working together step-by-step to negotiate approaches thatmeet shared goals. This collaborative approach would ensure that both qualitative insightsand quantitative rigor are incorporated into AI development.
hell yeah, this is what we address in our interdisciplinarity piece
common basic language for evaluating iterations of AI that do not assumethere is a linear or universal path of improving AI for all users, regardless of context.This shared language would help clarify where disciplinary specificity is needed and whereinterdisciplinary collaboration can be most effective.
Again, the thing about culture is that it cannot be done with a monopolar view from the west, which is the circumstance that allows for the scalarization of "bad" and "good"
how small language models can contribute to responsible innovation, and howto design decentralized infrastructure architectures that enable users to choose how theyshare and distribute their data and models
These feel like really key questions that do strike at the centralization/hegemonic nature of LLM development.
the need for alternativeapproaches to AI development that prioritize sustainability, justice, and inclusivity
How alternative are we talking? I'd love some more context here.
The role of metadata is complex
who assigns metadata, and how is it reconciled?
ocused on the gaps between community knowledge and com-putational knowledge
What does this mean? Can community knowledge be encoded computationally?
Institutional aspects and interdisciplinarity play a significant role in the cultures of AIproduction. There is a need for alternative imaginaries of technology that go beyond thecorporate inclusion of data.
This feels like a real throughline regardless - how do we do technology that is not adapted noblesse oblige, but by the stakeholders who use it?
first exercise of the day was a series of “speed dating” rounds.
Speed dating!! I'd love to know more about how well this worked