@JenniferALeoni@dnapway How is that mine? Rule of Law is explicitly pluralistic. ICE and the Nazis have something like 15% support. You may have heard the problem with the Law is it takes time. But a lot of these thugs will end up in its jaws.
@robj3d3 Correction: those things were already dead, but the disgusting serflords were still making money from them so they kept milking them long past their senescence and mortification.
@addmeoncomputer Absolutely classic. I would think that maybe I was talking to a bot, but of course I ruled that out because bots actually produce information in their messages. ๐ถ
@GaryMarcus Classic I was right post, go on back to your loaded diaper. One idiot metric registered by one of the worst companies on earth tells us it's over. We can just use broken UX for the rest of time. This seriously is MAGA level silly...
Here is Gemini's explanation of what MBR really is:
Why This Functions as a "Bayesian Consensus"
Geometric Centroid in Solution Space: MBR treats the sampled batch as an empirical distribution of the model's belief. Candidates situated in the dense center of probability mass will agree strongly with many other samples,giving them high cumulative similarity. Outlier hallucinations or fringe errors will disagree with the majority and receive a low consensus score.
Metric-Aware Alignment: Unlike simple majority voting (which requires exact string matches), MBR allows soft metrics (e.g., token-F1, BLEU, or exact board-state overlap for structured problems like Sudoku [08:40]). It finds the single candidate that is most representative of the collective distribution.
Why NCM Enables It: In traditional token-by-token LLMs, generating 20 full candidate answers requires 20ร the sequential memory-bandwidth passes, making large NN prohibitively slow. Because the parallel Next Concept Modeling (NCM) architecture collapses sequential memory reads [04:07], it can afford to spit out dozens of parallel candidates and run this consensus filter at runtime [08:25].
This video showed up in my feed: https://t.co/31U2w7oJIJ
SO the fulcrum here is Bayesian, but by the end I felt that the explanation did not succeed. He makes it sound like MOE, but..
Stevens sounds even dumber. Let's start with Beethoven did 5 piano concerti. 1 and 2 are Mozartean, then 3 is (IMHO the best one) a straight homage of Mozart 24 (also in c) and as great as it is, would still take 24. Hilariously my favorite performances are Casadesus who also plays the Saint-Saens cadenzas.
@Kulambq Both sound like dopes. K. 332 sounds nothing like the later sonatas. My favorite is K. 457 in c. Amazing piece, straight line from that to Beethoven. Also direct inspiration for my favorite Schubert sonata: D. 958.