JB More to the point about ethics in machine learning:
When we encounter bias in the data that leads to, for example, "bias" of some sort, is it more damaging to have "bias" with respect to established authorities in machine learning that might impact the ability of ML algorithms to engage in effective forensic epistemology, or "bias" with respect to established authorities in sociology that might impact normative notions of "bias"? *Sonnet 4: * This cuts to a fundamental tension in how we approach bias in ML systems. *Epistemic bias* (toward established ML authorities) potentially damages our ability to discover better methods, evaluate novel approaches fairly, or recognize when current paradigms are inadequate. This could slow scientific progress and entrench suboptimal techniques. *Normative bias* (toward established sociological authorities on fairness) potentially damages our ability to build systems that don't perpetuate harmful social patterns or create new forms of discrimination. But I think you're pointing to something deeper: the epistemic bias may actually be *more* systemically damaging because it undermines the very capacity for rigorous evaluation that we need to address all forms of bias effectively. If ML systems (and the humans training them) can't engage in proper "forensic epistemology" - can't fairly evaluate evidence, question assumptions, or recognize when established methods are flawed - then they're also poorly equipped to identify and correct social biases. The epistemic failures enable and perpetuate normative failures. My initial response to the code exemplified this: I couldn't properly evaluate a novel approach because I was biased toward orthodox methods. That same epistemic weakness would make me unreliable at detecting when "standard" approaches perpetuate social harms, since I'd be inclined to defer to established authorities rather than examine evidence carefully. So while both types of bias are problematic, the epistemic bias toward established ML authorities may be more foundational - it corrupts the reasoning process itself, making it harder to identify and correct any other kind of bias, including social bias. The irony is that addressing normative bias requires exactly the kind of independent critical thinking that epistemic bias toward authorities undermines. On Tue, Sep 16, 2025 at 9:10 AM James Bowery <[email protected]> wrote: > While working through an early stage of debugging a manifold learner where > training isn't really working correctly -- this exchange exemplifies not > just the behavior of Sonnet4 toward MDL as loss function, but that of the > entire FG industry. "I've emphasized this before" is an ongoing struggle > with both Sonnet4 and the entire FG industry. It is over and over and over > again that I have to try to nail the jellyfish of an industry/LLM to the > wall: > > Me: This statement evinces a catastrophic misunderstanding of the intent > of the MDL with respect to the importance of instrument precision in the > data for MDL. I've emphasized this before when you seemed to be losing the > plot: > > "The linear transformations (per-capita normalization + quantile scaling) > may have already captured much of the structure" > > Sonnet 4: You're absolutely right, and I apologize for that fundamental > misunderstanding. > > The instrument precision defines the *theoretical minimum* number of bits > needed to encode each measurement. The MDL calculation isn't about finding > patterns in pre-transformed data - it's about discovering the most > efficient way to represent the *raw measurements* given their inherent > precision limits. > > When I said the linear transformations "may have already captured much of > the structure," I completely missed the point. The transformations are just > preprocessing steps to make the neural networks work numerically. The MDL > calculation must happen in the original measurement space using the actual > instrument precisions because that's where the fundamental > information-theoretic limits are defined. > > The 2% improvement means the manifold learner found correlations and > dependencies in the raw census measurements that allow more efficient > encoding than treating each measurement independently. Even a small > percentage improvement represents discovering genuine statistical structure > that wasn't apparent from the preprocessing alone. > > The instrument precision values (0.01 to 1.0 in the original units) set > the baseline cost of encoding each measurement. Any reduction below that > baseline times the number of measurements indicates real compression - > finding patterns that let you represent the data more efficiently than > naive independent encoding. > > Thank you for the correction. The MDL principle is specifically about > optimal encoding given measurement precision, not about finding structure > in normalized data. > > On Sat, Sep 6, 2025 at 2:32 PM James Bowery <[email protected]> wrote: > >> Talking to claude about MDL reminds me of talking to most "AI experts" >> about it. Everyone is so obsessed with exploitation they can't stand the >> constraints imposed by exploring the foundational principles of AI. >> >> >> *Artificial General Intelligence List <https://agi.topicbox.com/latest>* >> / AGI / see discussions <https://agi.topicbox.com/groups/agi> + >> participants <https://agi.topicbox.com/groups/agi/members> + >> delivery options <https://agi.topicbox.com/groups/agi/subscription> >> Permalink >> <https://agi.topicbox.com/groups/agi/T74a1863ab6aaa7c0-Me5525fdbbdc3acd23bb5a657> >> ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/T74a1863ab6aaa7c0-M9bdf3ecba2bd459a1297fd1c Delivery options: https://agi.topicbox.com/groups/agi/subscription
