lordstorm

[lordstorm @lordsturm]

@MaximChistov @Timeroot This. Benchmaxing and reward hacking is most dangerous rot for frontier model currently.

[lordstorm @lordsturm]

@blagz22 @Mr_BenPrime Problem is that modern auto-solvers that are used to "confirm" this slop to be able to do it in finite time (less than few billion years on mainframes) are using dirty logical approximations and logical tricks that are considered ok, but can backfire epically.

[lordstorm @lordsturm]

@berd72018 @Headshok1962 They didn't really have any basis for advanced industry and there one-trick-pony of worlds best producer of beef.

[lordstorm @lordsturm]

@bridgebench Bad output happens when you force Kimi into an OpenAI-standard wrapper. Universal harnesses just fail. The future is modular, model-specific modular harnesses like DeepSeek Harness functioning as custom "Docker" environments to let every model hit peak performance.

[lordstorm @lordsturm]

@archohon @gen_enhazed About infrastructure toughness - true. Another thing is that he had a lot of long-distance things... but they are expensive and made for war with NATO (same cost problem as USA had with Iran).

Main point now is that allure of capture for profit was overdone by just winning.