so what is the overlap between "my usecase is not ML, math, physics, biology, or cybersecurity" and "I would pay exorbitant per-token rates for a better model" on an enterprise level, who wants to use a model with a silent active-sabotage feature and which is trained to never talk about security?

dramatization of Mythos finally meeting Dario Amodei

wwdc stands for "what would desus chew"

i am either cooking or completely out of my mind

if gpt-image fucked up the whiteboard don't blame me

not about anything in particular but some people on here are a bit too gullible w/r/t new AI papers no, this novel training method didn't make a 3B model better at all tasks than Claude, that paper didn't find a 10x efficiency gain, that new VRAM-saver is slow or degrades performance, etc.

I guess they're trying not to push their luck with already-tepid non-SF municipalities but I wonder how long it'll be until cities start making infrastructure explicitly more legible to AVs via short-range comms, including IR legibility in standards for signage, etc.

Noticing when I get routed to the Colossus Claude instances because it suddenly calls me a tranny

theory: the unifying principle between cranks, naïve people, and grifters is "using literary analysis in lieu of actually knowing what you're talking about"

complain about rationalists all you want but they were really cooking with "ideally you should only have beliefs that you could conceivably be convinced are false" thing

unironically a beautiful and effective way to keep grifters out. if you get filtered by an anime catgirl popping up for 2 seconds because it's too cringe and gay for you was your heart really in the systems programming?

I think I might be a bad scientist I am in the nasty habit of running my experiment before my baseline, so whenever I start the baseline run I spend the whole day rooting for the gap between the grey line and the brown one to get wider

fwiw, besides the obvious SWE-maxxing RL, it seems the actual baseline intelligence capacity of Claude Mythos is not that much higher than GPT-5.4 Pro (probably <3T), it's just that they don't really evaluate the Pro models on agentic stuff because that's unhinged

many events competing for the headline today but we all know the real groundbreaking news is Minecraft switching to Vulkan

I do like running my own AI at home, but it's worth acknowledging that my box pulls >100W idle, fights my air conditioner, and has a dismal PUE of probably <~30%, compared to >90% for a datacenter. you should use datacenters more and self-hosting less if you care about energy usage.

everyone quoting this with approval is either named something like "fizzy the tiefling prince 🌈 acab blm 🖌️ COMMS OPEN" or "anticiv anarcho-revanchist 🔻🪂" and they're all the most unbearable kind of person you will ever meet

bskypbc might be vibecoding, but the haters are viberootcauseanalysising

trying for impeachment right now is probably a very bad idea if you think impeachment is a real possibility in the future if his approval drops into the 20s then getting a bunch of GOP senators on the record as a "no" makes them less likely to switch in some face-saving cut and run down the line

overall living in Seattle has been like making the internet real in many pleasant and unpleasant ways

need

project hail mary: a good movie

but i don't want to use SPSA. i want to use some bullshit i half-remembered seeing on arxiv

Had an interesting discussion with a hiring manager !

I know it's a silly jokey joke, but people REALLY need to stop anthropomorphizing their boyfriends. "My boyfriend told me" your boyfriend didn't "tell" you anything. "My boyfriend thinks..." your boyfriend can't actually think or feel; just output a convincing simulation of it based on probability 🙄

getting intense gemhunger. need to cultivate wisemind

while cool, tbh it is not clear what the utility of differentiable code execution is you're still pretraining on text so you're only using this in SFT, because you'd almost certainly want to mask off the execution tokens during RL where it would learn the vast majority of its code execution skills

Labor saving technology is good, actually. I’m glad I own a washing machine and a dryer so I don’t have to spend a full day each week doing laundry.

Person with an advanced degree in a field whose value to humanity is only realized if people are broadly educated in it: its not my job to educate you

moved here ~8mo ago and really cannot imagine it without the 1 line it was a much better than driving for the vast majority of places I've needed to go in that time

watching The Sopranos it has this bizarre quality that it shares with Breaking Bad of being contemporary shows that feel like period pieces they both capture the Vibe of the time so precisely that they feel like exactly the kind of show you'd make today about that time period

what do you wanna bet that the "truth manifold" is, in fact, "recalling midtraining data designed to resemble the relevant benchmarks near-exactly without actually being copies of the test set"

zohran mamdani is personable in the same way that Claude models are send post

i really didn't expect the lasered hairs to take so long to be pushed out i have to shave DEAD hair. this is bullshit

my guess: american frontier labs spend all that money because they can. at every stage of the model training process you can decide to expend twice the money for an additional 5% improvement and if you're OpenAI/Anthropic/Google you can just do that

laser hair removal. yeowch

seeing a lot of excitement around this but worth noting: the site is extremely vague but everything about their language suggests that the hardware is not reprogrammable as far as I can tell this is a dedicated chip entirely for llama 3.1 8B and transistor count will scale linearly with params

Columbus Day is important to remember the lives lost 😔

pybubble v0.3.0 is out! it adds a bundled default rootfs, PTY support, interactive execution, and writable overlay filesystems github.com/arcee-ai/pyb... normal disclaimer applies, don't use this for untrusted user-supplied code, it's just for protecting against overzealous LLMs

ik the community around LLMs here is pretty sparse and just starting but, a critique: I would suggest that your enthusiasm for this technology go beyond its consumer/API applications and into the increasingly-approachable field of open LLM research there's a whole world out there!

pet peeve of mine is when people use "wire mother" as a shorthand for simulacra of real life friends/lifestyle/whatever. no! the wire mother is the one that had the milk, the cloth mother is the one that was comforting but fake!

yeah no i'll get right on that. let me just task my clawdbot. i mean my openclaw. they'll send beads to each other. and then the polecat in the gastown will pick it up. and then CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS will spin up on the gastown. and then the moltbook will communicate it

Drunken Seahawks Fans Climbing Space Needle

girlfriend just put me on a facetime with her family while they sang happy birthday to their dog and gave her a cake made entirely of apples and peanut butter

i think we all need to step back and realize that peak art was made when neolithic pot in the shape of a pig was fired

my experience has basically been that i no longer get "i cannot make myself read or write code" burnout, i get "i cannot make myself think" burnout, since the bottleneck is basically how fast I can process what needs to be done

recently I started getting ads for this smart budgeting app called "Rocket Money". I liked the concept but found the huge ad budget and free business model suspicious so I just went with the best service that charges upfront

as cool as this paper is, it seems to prove more that Qwen3 is super overindexed on gsm8k-style problems rather than a general statement about LLM capabilities llama3.2 apparently took about 500 params for a similar jump

why am I seeing a bunch of posts being like "children being unable to access the internet is absurd! clearly we should be banning it for everyone"