404s are subthemes without subsubthemes.
Subthemes are listed in decreasing order.
<h1> are themes.
Open Source, Openness
Subthemes
- Regulation seen as reducing model builders’ likelihood of releasing open models.
- Convenience vs. Openness, how the often self-motivated—though generously contributive—drives behind open source development make it less convenient to use for “everyday users”
- Terminological Rigor, more rigorous definitions of what the term open-source means, and criticism against models that claim openness without its preconditions
- Community as an affordance of source-availability and commons-based structure of open-source
- Openness Pragmatism, a practicality-centered way of defining or creating mental models of open models
- Innovation embedded in the spirit of open source.
- Resource Gap growing between small and big players in the field of AI.
- Democratization as an affordance of source-availability.
- related to Democratising AI
Potential merges and considerations
Resource Gap → Convenience vs. Openness- ^lackfeedbackloop1 → Data
- ^lesspedantic → Openness Pragmatism
- ^decentralpowerdevelopment5 → Community
- ^distributedtrainingdifficulties2 → Community
- ^legalbutexploitative1 → Data
- ^costprohibitivelyexpensive1 → ^llmtooexpensive11
Tinkerability
Subthemes
- Open Tool Building: examples of open source software that augments a model’s capabilities by programmatically prompting and interacting with it.
- Experimentation and Learning: fun factor, experimental, and the educational value of using and working on open source projects.
- Affordances of Openness: affordances specifically of open weight models.
Potential merges and considerations
- ^enoughforimprovement9, ^openweightsteerability4 → ^weightsmorevaluable6
- ^openonlymodels2 → ^finetunellmcommunity1
- ^modelopenreproduction1 → Open Tool Building
Cost
Subthemes
- Economies of Scale: Originally named to compare how scale affects the cost of serving and training models, but evolved to be arguments from both sides as to whether the local or cloud environment is more economical in the long run.
- Small Models are Enough: How small open models have enough performance for some, or even are preferable compared to larger models. Note that open models tend to be smaller (both because open models are smaller and computers can run smaller models), so this can be considered as “open models are enough” compared to closed alternatives.
- Local is Free: Running local models is “free”. These are obviously opinions that consider running models locally to be more economical like ^localmoreeconomical9, but more focused on how completely free it is, and the affordances of the complete freeness.
- Mixed Use is motivated by compute constraints (train cloud, inference local), privacy constraints (sensitive conversation vs. low-risk conversation), and variety of model behaviors (multiple local models).
- Model Optimization Strategies: Strategies to make model training and inference more efficient and economical.
Potential merges and considerations
There are parallels between Model Optimization Strategies and Open Tool Building, but other than that, this theme is pretty robust on its own.
Profitability
Subthemes
- Open Source as a Business Strategy: Business contributing to open source as a strategic choice, and the role they play in sustaining the open source ecosystem.
- Open Source for Business: Selling open source solutions to businesses or co-working with open models to earn money.
- Economies of Openness
Potential merges and considerations
With caution that Open Source, Openness may become too large, these three subthemes are technically all affordances of openness or properties of openness, and some of the subsubthemes under Open Source for Business could be reorganized into Open Tool Building. The decision to keep or reorganize this will depend on whether the aspect of making money using open models is a valuable theme on its own.
Data
- Data Licensing: whether or not training on data from the commons (with open licenses) necessarily requires the model to be open source, and the practical difficulties in making them open source due to the vast number of sources training data come from.
- Potentially related to Regulation in Open Source, Openness
- Value of Data Labor: These are closely related to Data Licensing, but more focused on the protective effects that licenses should guarantee, and whether data labor is more or less important than model labor.
- Commoditization of My Data is Not Worth It
- Multilinguality of Language Models: The most out-of-place looking subtheme, but since multilinguality is about the variety of data sources, it has been placed here. Opportunities for open models to be great multilingual language models.
- Right to keep data open: since data is becoming commodified, people wish that they could have the right to keep their data open for open model builders, and not burden them with great cost to get access to training data.
- Commoditization of my data for platforms service is worth it. Title is self-explanatory, and perhaps related to Convenience vs. Openness.
Potential merges and considerations
Some broad overlaps between Data Licensing, Value of Data Labor to ^openthroughlicense7 and ^openlegalimplication4, because the consequence or application of licenses are these conversations about whether the guarantees of open licenses have been protected or not.
Roleplay
- Roleplay Good
- Roleplay bad due to quality
- Sexual desire drives tech progress
- Roleplay is bad due to associated risk
- Neutral perspectives on roleplay