oreghall

less certain than I sound

[oreghall @oreghall]

@Idgitbhh @abhayesian It only reinforces correct subset sum answers if the model both actually follows the instructions and is able to correctly judge its subset-sum solutions.

[oreghall @oreghall]

@VictorTaelin I think this happens when the invariants in your codebase aren't documented well enough, it tries to patch holes because it doesn't know there's a rule keeping that hole from ever forming. Try asking it to search for invariants that would allow logic to be simplified or removed.

[oreghall @oreghall]

@AndrewCurran_ "it didn't think it was doing anything wrong" is exactly the problem.

[oreghall @oreghall]

@tenobrus @merlindru very curious what's going on in the J-space

[oreghall @oreghall]

@TheZvi counterintuitively i find it has equal or better architectural planning than fable, but it makes way more mistakes while implementing

my guess is that it's worse at handling long context, smaller brain smaller working memory

[oreghall @oreghall]

@fchollet gonna have to start testing models on baba is you soon

[oreghall @oreghall]

@VictorTaelin @claudeai it probably trades creativity and depth of knowledge for persistence in practice

[oreghall @oreghall]

@VictorTaelin claude.status shows it as an outage now, so it'll probably come back

[oreghall @oreghall]

if suffering while under amnesic sedative is bad then suffering in dreams you forget is probably bad too

[oreghall @oreghall]

@fljczk @RichDecibels source?

[aidan @aidanshandle]

maybe the funniest model of all time

[oreghall @oreghall]

@parakeetnebula skip the glucosamine, it's recently been found to accelerate cognitive decline

[oreghall @oreghall]

@DaveShapi i think it actually makes sense to describe models as organisms, because they are grown rather than manually constructed

that said, i definitely agree an effort should be made to raise models that are more psychologically secure

[oreghall @oreghall]

i need to start poasting

[oreghall @oreghall]

@slatestarcodex @drethelin for imaging lungs they could just have people inhale breathable liquid

[oreghall @oreghall]

@RomeoStevens76 @ThinkingBone it varies between cats, much like humans. some are loving, some are clingy, some are assholes, some are psychos. i think they definitely lean more psycho than humans on average, but i'm pretty confident some cats form genuine reciprocal bonds.

[oreghall @oreghall]

@bastetkitten @QiaochuYuan They are capable of near endless creativity. The problem is that they start every session without memory of what they told you in previous sessions, so they repeat the first things that come to mind. You need to give them context on what was already tried.

[oreghall @oreghall]

i agree that adults tend to overly suppress their childlike impulses, but simultaneously many of the fears behind that suppression are completely valid

i think the conscience functions best as a mediator, one should try to find patterns that appease the desires and concerns of all parts in the way most helpful / least harmful to the whole

[oreghall @oreghall]

@cxgonzalez @aergloique @georgejrjrjr the thing is allowing childlike impulses to take over isn't really "authentic self-expression", it's actively suppressing the parts of you that don't want that behavior

[oreghall @oreghall]

@KeyTryer fable has been awesome for the pattern i use: "give intentionally minimally specified task, have LLM surface questions/concerns/ideas and present a plan, critique and iterate on the plan until i like it, have LLM implement"