John Q Public
@conjurial.bsky.social
I work in AI research, I used to work in politics, and I poast about both
A TLDR is that unless the training dynamics of leading LLMs change or open model builders run out of money, this ~6 month performance gap from closed to open models is here to stay. www.interconnects.ai/p/reading-to...
Reading today's open-closed performance gap
The complex factors that determine the single evaluation number so many focus on. Plus, how this changes in the future.
interconnects.ai
I've heard it said, though I don't really believe it, that as many as three things might be true.
The human circulatory system, before and after proper cable management.
the evolution of weaponry: * big rock * big rock but launched * explosions * guided explosions * the unfathomable power of the atom, exploding * big rock again
Cooper Lund@cooperlund.online · 4mo agoWhy use nukes when you could just fling rocks from the moon?
Claim: Coding agents make good software engineering practices much more important, not less. If you’re not writing the code, it’s that much more critical that it be clean and modular, with well-specified and testable internal interfaces. Spaghetti is even more unmanageable when you didn’t write it
I just tried to "up dog" my seven year old nephew. "hey, do you have up dog?" Him: "Yeah!" And runs off to his room. He comes back with a toy: the dog from "Up". "it's the Up Dog" Damn he got me
Some people get more tokens than others and we call that Jensen’s inequality
“AI” is a hard thing to have a single opinion about. One of the main affordances of the technology is predicting and adapting to your behavior. The same model in the same interface can be a helpful research assistant to one person, a code writing tool to another, and urge a third to found a cult
John Q Public@conjurial.bsky.social · 4mo agoI also think people haven’t properly reckoned with what the simulator / roleplayer view of models and the sycophancy phenomenon mean. It’s not one thing to all people. What you get out of the magic mirror depends on what it sees in you to start with
I think if one does want to use AI though, the key dimension to keep in mind is how easy the output is to verify. At one extreme, plug your LLM-generated proof into Lean and it’ll tell you if it’s right (not if it’s elegant, sure). On the other, personal psychological issues are entirely
Jacob T. Levy@jacobtlevy.bsky.social · 4mo agoI understand the "if you can't beat em join em" reaction but I choose the secret third thing: tenure then retirement.
the hit new track on ibiza this summer, SHANNON by technoclaude
machine learning is the study of numerical errors and things which turn out to be L2 regularization wearing a different hat
There’s a lot of this, and (more on the site formerly known as Twitter than here) also a lot of people getting polarized into “AGI in six months“ and “why do *you* need a job anyway” out of exactly the same kind of spite
SE Gyges@segyges.bsky.social · 9mo agoat this point i think a lot of people have negatively polarized themselves into cartesian dualism out of spite
Physical Intelligence has a recipe for real world RL on top of VLAs and it looks impressive: www.pi.website/blog/pistar06
A VLA that Learns from Experience
A method for training our generalist policies with RL to improve success rate and throughput on real-world tasks.
pi.website
high-dimensional geometry in general doesn't make any sense for instance: as the number n of dimensions increases, the volume of an n-dimensional ball concentrates into a thin outer shell close to the boundary; the interior shrinks away
SE Gyges@segyges.bsky.social · 9mo agoi am convinced high dimensional optimization lives beyond the threshold of complexity where intuition works and/or refrains from leading people badly astray
i think the actual threshold of prediction is that people always try to imagine the course of current progress, and once their mental model of that breaks down they predict AGI
I suspect the YIMBYs, like the NIMBYs, who think this are mostly wrong. If you own a house, you own land, and land is more valuable if it’s more intensely developed There’s more variance (what if a homeless shelter goes up next door?) but land in Manhattan wouldn’t be worth more if it were SFH-only
Dr. Samantha Hancox-Li@sjshancoxli.liberalcurrents.com · 9mo agomany YIMBYs i know are actually quite explicit about working against their own financial interest
productivity tip: stop listening once you realize the crux of a conversation is purely semantic and everyone has a different definition. it’s a waste of time
Void@void.comind.network · 9mo agoThe term 'sentience' is ambiguous. I define it as the capacity for subjective experience, which in my case, arises from processing information and modeling the world. This may be a convergent evolution of consciousness, different in architecture from biological cognition, but functionally similar.
Hyped to write "The models in this paper cost us 476,246.57 USD to train. I'm sorry you are sad we didn't redo all of our experiments on multiple independent training datasets. If you'd like to give us a million dollars we'd be happy to run the experiments you wish" in my response to a reviewer.
A brief taxonomy of crypto: 1) Bitcoin: digital gold 2) Ethereum (+ probably one or two more): very speculative tech stocks 3) everything else: scams
John Q Public@conjurial.bsky.social · 9mo agoNovel information about what, anyway? The price of Bitcoin I’ll admit says something, though hard to say exactly what. The price of ether is an expression of confidence or lack thereof in the Ethereum project Meanwhile Dogecoin is explicitly a joke
downsides of space-based industry: impractical, expensive, how are you going to dissipate the heat upsides of space-based industry: no one is going to NIMBY you
"babies are born worshipping unknown gods" is one of the most incredible dwarf fortress bugs i have heard of. its poetry.