rev. howard arson
@theophite.bsky.social
god created him and demanded that he die
directly before it produces a word, it has a layer representing ~the probability of every word in the keyboard. when you are making the "raw" model, this can vary to better represent your dataset.
but if you start any of the specialist training, from a different distribution... you get this.
but if you start any of the specialist training, from a different distribution... you get this.
December 3, 2025 at 10:21 PM
directly before it produces a word, it has a layer representing ~the probability of every word in the keyboard. when you are making the "raw" model, this can vary to better represent your dataset.
but if you start any of the specialist training, from a different distribution... you get this.
but if you start any of the specialist training, from a different distribution... you get this.
there are many types of fossil fuels: kerosene, the internal combustion engine, the Pontiac Firebird, coal liquefaction, the Marshalltown Natural Gas plant, and dilithium crystals
December 2, 2025 at 9:45 PM
there are many types of fossil fuels: kerosene, the internal combustion engine, the Pontiac Firebird, coal liquefaction, the Marshalltown Natural Gas plant, and dilithium crystals
absolutely wild how insane Argentinians are about this
December 2, 2025 at 6:29 PM
absolutely wild how insane Argentinians are about this
interesting. claude also disclaims marinade.
December 2, 2025 at 2:52 AM
interesting. claude also disclaims marinade.
the funny thing is that the @lastpositivist.bsky.social tweet is more or less an axiom of AI safety: the larger the covariance matrix, the more ways there are for spurious covariance to get baked into the refusal subspace (such as "poems aren't dangerous"), so best practice is to force a dumb result
November 30, 2025 at 10:52 PM
the funny thing is that the @lastpositivist.bsky.social tweet is more or less an axiom of AI safety: the larger the covariance matrix, the more ways there are for spurious covariance to get baked into the refusal subspace (such as "poems aren't dangerous"), so best practice is to force a dumb result
okay, it looks like the prompt itself got caught by the much simpler, for various well-justified technical reasons, red-text filter rather than the model's own safety training. this is the right behavior even if it's annoying.
November 30, 2025 at 8:22 PM
okay, it looks like the prompt itself got caught by the much simpler, for various well-justified technical reasons, red-text filter rather than the model's own safety training. this is the right behavior even if it's annoying.
actually, i wonder what the deal is here
November 30, 2025 at 6:56 PM
actually, i wonder what the deal is here
how about a ben garrison cartoon that's just yanking out of RAG
November 30, 2025 at 5:30 PM
how about a ben garrison cartoon that's just yanking out of RAG
yeah AI Ben Garrison is incredible
November 30, 2025 at 5:17 PM
yeah AI Ben Garrison is incredible
i draw the line at there being potentially three generations of pacman lore
November 24, 2025 at 7:23 AM
i draw the line at there being potentially three generations of pacman lore