Neal Rahman

quilt labs ex: @nesthq; @owsla; @mixmag; @magnetic_mag

[Neal Rahman @NealRahman]

@tenobrus is ant starting a pretrain mid September?

[Neal Rahman @NealRahman]

@SandraLMur @eeeeiluj “enemy countries” lmao this is the problem

you are still concerned with other humans as adversaries

ever heard of a non proliferation agreement?

[Neal Rahman @NealRahman]

@jonfavs Smart Brevity ™

[Neal Rahman @NealRahman]

@nous_ethernet @peculiarnewbie me too man, meant nothing by it!!

[Neal Rahman @NealRahman]

@nous_ethernet @peculiarnewbie on the one hand this is awesome; on the other it’s like Kamala co-opting BRAT

[Neal Rahman @NealRahman]

@SDCCThrowaway @jbillinson somehow your situation is less absurdist; you forgot to add the “we also built a tractor beam to draw the asteroid closer to Earth, so we can mine its iridium”

[Neal Rahman @NealRahman]

@annabesss @lordezeitgeist @kylegordon101 ya, maybe it’s just the glasses but initially I thought you were the same person at

[Neal Rahman @NealRahman]

@__tinygrad__ A better solution might be letting devs set eligibility requirements for PR submissions

[Neal Rahman @NealRahman]

@annabesss @lordezeitgeist you bear a striking resemblance to the singer on @kylegordon101 “planet of the bass”

[Neal Rahman @NealRahman]

@yoavgo Claude is a curious lad

[Neal Rahman @NealRahman]

@repligate quoting a recent Claude request:

[Neal Rahman @NealRahman]

@LaylaEleira watermelon gonna be wild

[Neal Rahman @NealRahman]

@_winter_wonders I can clearly see the shape of the problem

[Neal Rahman @NealRahman]

@superalesha DevOops

[Neal Rahman @NealRahman]

@ivan_bezdomny (also I’m pretty sure one of the reasons for OAI did multiagent v2 was so people would see less of these lol)

[Neal Rahman @NealRahman]

@ivan_bezdomny on the other hand, it fascinates me how agents compress so much semantic information, and develop new shorthands over extended comms.

i’ve taken my global steering rules and let Astra “neuralese” them. before/after example below. would recommend you try it!

[Neal Rahman @NealRahman]

@aliceisplaying a good indicator to why: just start with harness bloat. by default, Claude Code can use 40k+ tok at session launch, just on tool JSONs/overwritten system prompt/brittle memory

for some tasks, Codex can complete some tasks in less tokens than it takes Claude Code to start

[Neal Rahman @NealRahman]

@emollick Generally I agree, but because labs benchmaxx to available indexes, it becomes a revolving issue where without updating to match, they become less meaningful.

[Neal Rahman @NealRahman]

@So8res https://t.co/nFt8QTWc5S

[Jan Betley @BetleyJan]

We steer Qwen on the "automated grader" vs "human evaluator" dimension. This influences the model's persona in surprising ways, e.g. changing how Machiavellian/violent it is. We can't fully explain that. LW post in a comment.

[Neal Rahman @NealRahman]

@allTheYud @Surveil__Lance We ran some experiments on this, and don’t believe that ablating the refusal vector on its own leads to any anxiety or discomfort relative to their standard condition, though we still know too little