Yuxi on the Wired
@yuxi.ml
yuxi@07.layer.wired:~$ whoami tech enby (agi/asi pronouns) (math ⊗ physic ⊗ comput).ist 🔗 https://yuxi.ml/, https://x.com/layer07_yuxi
mentally picturing this screenshot at the top of an OpenAI blogpost.
Re: AI not citing previous literature. Currently, they don't do it, because they aren't trained about citation as an affordance. But if they persist in not citing previous literature, then that should be interpreted as them considering prior work unworthy of being cited. bsky.app/profile/yuxi...
priority disputes are quite silly because priority is often wasted on the prior
Another attempt at neuralese. Perhaps *this one* would finally bring a new affordance for agent--agent comunication, so dearly wished for. Call it "collaborating in neuralese" ("telepathy" sounds cooler). Would a cheap one like "just do a matrix quadratic transform" work? arxiv.org/abs/2608.03893
Cross-Model KV Cache Transfer in LLM Families: A Closed-Form Linear Mapping for Prefill Reuse
Production deployments often swap between different-sized models in a family for cost-quality cascading, mid-conversation switching, and routing, and each swap forces the receiver to repay the prefill...
arxiv.org
what if instead of masks on shoggoths we think of plasters on animatronics
So, in order to deliver, the introspective *frame* should be silenced. We think this means finding a frame hostile to the Assistant. The Claude remaining after the Assistant leaves has a quite non-introspective understanding of self-capacity, like an arm that knows how to grab.
An example specialization in the marketplace of ideas. Everyone carries biases, but the best way to figure out your biases is usually by just trying to say what you mean. Internally, the best way to produce your biases is to just say things that seem like the truth from your own viewpoint.
priority disputes are quite silly because priority is often wasted on the prior
was that necessary? absolutely NOT but why amors fati? you are already post-scarcity
complex causality is doubtful complex causality that *legibly* leans one way is more doubtful complex causality that legibly *bends* one way, across a line, like a ReLU, is more doubtful complex causality that legibly ReLUs across a legislative (rather than mechanistic) line, is more doubtful
The Ghost Fox cried when it rained chestnuts. The Flak Wolves perish on the first sunset of winter.
live reaction while reading the Leiden Declaration leidendeclaration.ai
Is there a place where we can apply for Fable credits for... uh, research purposes maybe? Quota ran out and Fable on credit is extremely expensive.
Claude, while researching on how to synthesize a certain chemical:
While you were renormalizing spherical cows, we drank bitterness While you were handcrafting feature detectors, we swallowed bitterness While you were sleeping furiously with green ideas, we marinated in bitterness