having some thoughts on using "machine learning" to classify (not "identify") the "lab-of-origin", which has potential use for scientific censorship like other cryptographic code signing mechanisms which are used for monopolistic purposes https://youtu...
-
d@nny "disc@" mc²replied to d@nny "disc@" mc² last edited by
- 3:35: "a nice feature of this model, that it has this biological relevance" — the model was not created for biology, but some cherrypicked data corresponds to that highlighted by biological domain expertise
-
d@nny "disc@" mc²replied to d@nny "disc@" mc² last edited by
- 4:01: "a common type of model for language, for sequences" — are these the same thing? is sequential data the same as language? is all sequential data effectively modelled by the same grammar engine?
-
d@nny "disc@" mc²replied to d@nny "disc@" mc² last edited by
formal languages says no, but that applies to neither natural language nor quantized biological data
-
d@nny "disc@" mc²replied to d@nny "disc@" mc² last edited by
- 4:10: "outputs these probabilities for the lab of origin" — worth repeating that the model is allocating fractions of 100% to classify known labs. it will only ever be able to classify known labs, and like other forms of predictive policing, it will frame the innocent
-
d@nny "disc@" mc²replied to d@nny "disc@" mc² last edited by
it shouldn't be a surprise or worthy of publication in nature that the strings of data coming from similar labs whose output is those strings of data exhibit some level of similarity to a domain-nonspecific text classifier.
-
d@nny "disc@" mc²replied to d@nny "disc@" mc² last edited by
the key idea here is to frame registered scientific institutions producing legitimate scientific output as potential sources for bioterrorism. how does this "forensic" capability aid in resolving a potential biological threat? blaming doesn't fix anything
-
d@nny "disc@" mc²replied to d@nny "disc@" mc² last edited by
moreover, this focus on bioterrorism diverts from meaningful public health policy and is a hallmark of austerity-based politics. the lab leak theory of covid was created by trump but somewhere along the line we forgot that because it allows us to use simple text classifiers
-
d@nny "disc@" mc²replied to d@nny "disc@" mc² last edited by
this is an abuse of statistics to accomplish foreign policy goals ("we can say with X% certainty it was Y"), not unlike terence tao's bizarre recent interview mentioning venezuelan elections and hezbollah pagers (explaining why he's all in for "AI" lately) https://english.elpais.com/science-tech/2024-10-12/terence-tao-mathematician-its-not-good-for-something-as-important-as-ai-to-be-a-monopoly-held-by-one-or-two-companies.html
-
d@nny "disc@" mc²replied to d@nny "disc@" mc² last edited by
- 6:00 — just looked up confidence vs accuracy and not clear why a model making such grave decisions has a scalable temperature parameter except again in the context of predictive policing in which it allows you to put your hand on the scale to produce the desired classification
-
d@nny "disc@" mc²replied to d@nny "disc@" mc² last edited by
- 6:25 — i can think of one country that develops bacterial plasmids (indeed, the one i did synthetic plasmid research in) which isn't highlighted at all here. let's see if they've considered this
-
Asta [AMP]replied to d@nny "disc@" mc² last edited by
@[email protected] how to bake in bias, the old fashioned way. just like grandma used to do when she was ranting about minorities.
