(WIP)
I’d like to restate the core suggestion of y’day’s post:
- value is qualitative
- LLMs embed qualities in quantitative vector spaces
- quantitative representations are communicable across computer networks
- lateral federation supports decentralization
- ‘embedding alignment’ can support decentralized value-coherence
and now, expanding on the above:
1: value is qualitative
this is largely based on / borrowing from Massumi’s 99 theses.
captitalism conditions us to prioritize the quantitative scalarized money-commodity over all else. it is our “yardstick” by which we measure and commensurate all things.
Massumi encourages us to explore value’s qualitative aspects, and (imo) suggests that the most important aspects of value are qualitative.
2: LLMs embed qualities in quantitative vector spaces
this seems self-evident. LLMs are strange and we perhaps do not understand exactly what they are doing, but they seem to have impressive ability to grasp “qualities”.
under the hood, they use a highly quantitative representation - high dimensional vector spaces with no intrinsic “qualitative meanings”. yet they are able to “embed qualities” into these spaces.
3: quantitative representations are communicable across computer networks
obviously, quantized digital data can be easily transmitted.
text, one quantized representation, is often used to communicate qualities between humans and machines. yet, some of the meaning is often lost in the “game of telephone”.
quantized vector-space representations could communicate complex qualities across computer networks without loss via encoding+decoding through text.
4: lateral federation supports decentralization
scale exerts a homogenizing & fungible-izing effect on what it touches. for example, labor markets commoditize and fungible-ize humans. imo, the dehumanizing effects of working a corporate job are connected to this.
small-scale collectives are better able to support the needs of their uniquely-individual constituents.
yet, large-scale coordination is necessary. lateral federation between small-scale collectives seems a desirable alternative to vertical large-scale, and better supports decentralization.
decentralization is desirable because it supports the preservation of each collective’s culture.
5: ‘embedding alignment’ can support decentralized value-coherence
imo, coordination of many small-scale entities has proved challenging and is perhaps the main reason why decentralization efforts are failing. vertical large-scale is more effective at coordination, perhaps largely due to its use of a “single central value-representation”.
if there are numerous LLM-centered collectives, each with their distinct flavor of “qualitative valuation”, how can they coordinate and interoperate with each other? how can they hope to challenge the efficiencies of vertical large-scale?
perhaps ‘embedding alignment’ offers hope. if qualitative value-representations can be aligned between different decentralized quality-collectives, it seems possible that lateral federation could achieve sufficient coordination-efficiency to be competitive.