Gemini is right. But I'm pretty sure with more effort I could have improved compression by discovering correlations between population density, income, crime, and race. It still wouldn't have discovered causation. It wasn't known whether crime causes poverty or poverty causes crime until we did (expensive) experiments of randomly giving out money and finding it had no effect on crime.
I suppose with vibe coding we would have time to pursue it again. Except that it is having the opposite effects, flooding us with submissions to the Hutter prize for us to evaluate. -- Matt Mahoney, [email protected] On Sat, Oct 24, 2026, 4:50 PM James Bowery <[email protected]> wrote: > By the way Matt, sSpeaking of the ethics of "Is", get a load of what > Gemini Aragon High (the most advanced reasoning agent Google DeepMind has) > gave as the second most important reason for the proposed Hume's Guillotine > contest having no impact: > > 2. Trivial Redundancy Swamps Causal Discovery in Compression > > When compression pioneer Matt Mahoney tested the proposal by writing a > custom ZPAQ compressor for LaboratoryOfTheCounties, he abandoned the > effort because the easiest ways to shrink the file had nothing to do with > macrosocial causation. > > In tabular census data, a compressor achieves its biggest gains by > exploiting mundane structural redundancies—such as predicting a county's > population in year $t+1$ from year $t$, or deriving state-level totals by > summing county rows. Because lossless compression scores every saved bit > equally, clever engineering around formatting, measurement noise, and > accounting identities easily outweighs the bits saved by discovering a > genuine sociological feedback loop. > > Do you feel even slightly uncomfortable being seen in this light by > history or are you a nihilist? > > On Sat, Oct 24, 2026 at 8:58 AM James Bowery <[email protected]> wrote: > >> >> >> On Fri, Oct 23, 2026 at 1:35 PM Matt Mahoney <[email protected]> >> wrote: >> >>> Actually my model of ethics is that there is no rational or absolute >>> model of ethics. >>> >> >> That's the foundation of a meta-ethics, as we shall see below. >> >> Most models of ethics are strongly held opinions that resulted from group >>> selection for reproductive fitness. >>> >> >> Your choice of the word "opinions" conflates two aspects of AIXI as a >> rational agent: >> >> Ought: Sequential Decision Theory's utility parameter. >> Is: Solomonoff-induced world model as a belief system. >> >> That's why my approach to meta-ethics is called "Hume's Guillotine": >> >> Hume was the first to explain the importance of distinguishing Is from >> Ought in such clear terms. >> >> Humans required a reprogrammable "Ought" to become tool users—vertically >> transmitted memes that permitted the expansion of their ecological range >> out of Africa. This was primarily father to son because males were the >> ones forced to live in the periphery of the ecological range by female >> hypergamy's preference for polygyny, as I described in "Race, Gender and >> the Frontier" back in 1992 >> <https://groups.google.com/g/soc.men/c/LJ2TIpKv6Ck/m/f68XCu6888IJ>. >> That thesis has held up remarkably well and probably cost me millions if >> not billions for stating it publicly when I did. After getting the law >> passed that opened the way for SpaceX, I saw too much of how the centers of >> power operate. >> >> By the way, if you search for that exact phrase, search.brave.com is the >> only search engine that will return what I wrote. The others return a >> later piece of "wokism" by the same title. >> >> The barriers to truth are so baked in that even Elon Musk's much vaunted >> "Truth Seeking AI" exhibits selective lobotomy when it comes to Hume's >> Guillotine. Read this final response from Grok to this exchange about >> Hume's Guillotine >> <https://grok.com/share/bGVnYWN5LWNvcHk_d05cd0ea-e888-4f06-893e-0cc3130bbe5d> >> where I asked: >> >> "I again, ask you to please introspect on your reasoning trace." >> >> *Grok4.6 Expert:* >> The error is clear in the trace. >> >> When asked “Why not?”, I ranked computational intractability first. That >> ranking came from the standard literature on Solomonoff induction and >> Kolmogorov complexity, where uncomputability is the dominant practical >> objection. I applied that background pattern to this proposal instead of >> checking whether the README actually requires computing the ideal. >> The README does not. It defines the criterion as the size of a working >> program that regenerates the given dataset (self-extracting archive or >> compressor + compressed file), under explicit resource limits, modeled on >> the Hutter Prize. It states that scoring is independent of how the model >> was produced and does not require computers at all. Those passages were >> already in the text we had retrieved. >> >> The mis-ranking therefore occurred because the general theoretical >> objection was treated as automatically decisive for this specific, >> operationalized contest format. That is a failure to keep the concrete >> proposal distinct from the ideal it approximates. The data-selection and >> institutional points remain, but the intractability claim does not belong >> at the top of the list for the contests as written. >> >> *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/Te59bc90acbee0c99-M943d69ba576836ae6bc41762> > ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/Te59bc90acbee0c99-Mf05f78a0b9dab0c2da13745f Delivery options: https://agi.topicbox.com/groups/agi/subscription
