Is it the case that UnitedHealth and Cigna each own (or control) one of the "big three" PBMs? If so, that is a just crazy - the control insurance premium pricing, benefit decisions, AND the pricing of covered medications?
yadaebo wrote below "Medical Loss Ratio (MLR) is capped at 85% in the US which means 85% of revenue must go to patients". Does controlling a big PBM allow an insurance company a loophole?
I think you both make valid points, but I also get the sense that the article is articulating insights gained from pure math explorations into the theoretical limitations of learning, which in the article can sound "turboencabulator-speak" when compressed into words.
Maybe I should have just linked to the research paper:
Basically, as LLMs scale up, the author (Soatto, VP at AWS) suggests they're beginning to resemble Solomonoff inference: hypothetically optimal but computationally infinite approach that executes all possible programs to match observed data. Repeating this approach for any given question by definition gives the best answer, yet requires no learning, since the entire process can be repeated for any query (thanks to infinite computation).
The article develops a theoretical framework contrasting traditional inductive learning (which emphasizes generalization over memorization) with transductive inference (which embraces memorization and reasoning). Here's a quote:
"What matters is that LLMs are inductively trained transductive-inference engines and can therefore support both forms of inference.[2] They are capable of performing inference by inductive learning, like any trained classifier, akin to Daniel Kahneman’s “system 1” behavior — the fast thinking of his book title Thinking Fast and Slow. But LLMs are also capable of rudimentary forms of transduction, such as in-context-learning and chain of thought, which we may call system 2 — slow-thinking — behavior. The more sophisticated among us have even taught LLMs to do deduction — the ultimate test for their emergent abilities."
Yes, I should have made that clear in my first comment. Thanks for doing so. I used the quote in my title because I found it a fascinating way to start a technical blog post, and it made me want to read the article to understand what the author was planning to write from such a beginning.
I found the opening quote of this article to be intriguing, especially since it was from a 1992 research lab:
“One year of research in neural networks is sufficient to believe in God.” The writing on the wall of John Hopfield’s lab at Caltech made no sense to me in 1992. Three decades later, and after years of building large language models, I see its sense if one replaces sufficiency with necessity: understanding neural networks as we teach them today requires believing in an immanent entity.
reminds me of the anthropic's recent work on identifying the neuron sets that correlate to various semantic concepts in Claude: https://news.ycombinator.com/item?id=40429540 "Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet"
Ditto. The larger question is how does the linear sequence of DNA (primary sequence) ultimately drive the spatio-temporal development program that leads to mature differentiated cells working together at the organoid / organ / organism scale. How do cells know what to do in time and space as the organism grows? How is that logic encoded in the genome?
Eric Davidson did a bunch of pioneering work meticulously "debugging" this spatiotemporal genomic logic in the sea urchin. Pretty amazing. Eukaryotes like us have control elements directly upstream of our genes (trans-acting aka close acting) and also 100,000's of base pairs distant (cis-acting). The region of DNA directly preceding the beginning of an open reading frame at the start of a gene usually has a sequence of DNA motifs that bind proteins that can increase or decrease expression of the gene. Davidson and others showed that the transcription factor proteins that bind to these control motifs actually have additional other proteins that bind to them, in literally a layer on top, and that the sequences of proteins in this second layer recruit a tertiary layer of proteins that conditionally cause more or less gene expression, depending on their identities. You could say the secondary and tertiary layers are a form of "abstraction" in a literal sense, since they encode a hierarchy of logical operations.
Here's an open-access overview of Davidson's work which incidentally illuminates a lot of these concepts in more detail for a lay audience: "ERIC DAVIDSON: STEPS TO A GENE REGULATORY NETWORK FOR DEVELOPMENT" by Ellen Rothenberg, 2016; doi: 10.1016/j.ydbio.2016.01.020 https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4828313/
to see the decoded logic in pseudocode and with a diagram, see "cis-Regulatory control circuits in development", Howard and Davidson, 2004, Developmental Biology, vol 271, https://doi.org/10.1016/j.ydbio.2004.03.031 (open access)
one of my favorite books since high school. I remember mentioning it to my HS English teacher before class one day (20 years ago) and she amazed and surprised me by replying "oh that's one of my favorites!" It seemed like an esoteric book to me at the time, but her generation also knew of it. maybe the OP can share what they find interesting about it...
This comment is tone-deaf to the unique (and effective? TBD) arrangement of the board OpenAI 501(c)3 without compensation and the company they regulate. Your comment strikes me as not appreciating the unusually civic-minded arrangement, at least superficially, that is enabling the current power play. Maybe read the boards letter more carefully and provide your reaction. You castigate them as “non-techies” - meaning… what?
VSI thermodynamics is my favorite so far, it covers statistical mechanics of gasses, Boltzmann distribution defn of temperature… if you don’t know what these concepts are, I am confident you can get a good intro from the VSI thermodynamics book.
The author referenced a paper from 2009 about the microRNAs and BK subunits: "Post-transcriptional regulation of BK channel splice variant stability by miR-9 underlies neuroadaptation to alcohol" https://www.ncbi.nlm.nih.gov/pmc/articles/PMC2714263/
"Tolerance represents a critical component of addiction. The large conductance calcium-and voltage-activated potassium channel (BK) is a well-established alcohol target, and an important element in behavioral and molecular alcohol tolerance. We tested whether microRNA, a newly-discovered class of gene expression regulators, plays a role in the development of tolerance. We show that in adult mammalian brain alcohol upregulates microRNA (miR-9) and mediates post-transcriptional reorganization in BK mRNA splice variants by miR-9-dependent destabilization of BK mRNAs containing 3’UTRs with a miR-9 Recognition Element (MRE). Different splice variants encode BK isoforms with different alcohol sensitivities. Computational modeling indicates that this miR-9 dependent mechanism contributes to alcohol tolerance. Moreover, this mechanism can be extended to regulation of additional miR-9 targets relevant to alcohol abuse. Our results describe a novel mechanism of multiplex regulation of stability of alternatively spliced mRNA by miRNA in drug adaptation and neuronal plasticity."