here's another bsky.app/profile/doll...

here's one bsky.app/profile/segy...

but then again the 3 month figure is just an average and current narrowing could be fluctuation rather than trend

i love the phrase "inside traders announce"

Open-weight models lag state-of-the-art by around 3 months on average Epoch AI is a research institute investigating key trends and questions that will shape the trajectory and governance of Artificial Intelligence.

very roughly gpt5.4 was released 46 days ago and this new kimi matches it so that's the current gap and according to epoch the gap has historically been 3 months so yes things are narrower than in the period up to october last year
epoch.ai/data-insight...

actually now that i thought about it more i'm not sure it's narrowing. has been pretty narrow for a while.

the open-closed gap on generally-available model performance is narrowing

frontier labs risk being undercut in opus-class until they make mythos-class GA

yes

in principle you could have a model so aligned that it successfully misalignment-fakes any attempt to misalign it

...unless the misalignment finetuning is sft

FT 2 days ago: "Amodei says he suspects open-source models and Chinese developers will be able to replicate Mythos's capabilities within six to 12 months."

but mythos isn't available to be distilled
hmm

mathematicians as stochastic parrots?

lots of polemics i don't agree with and some less than convincing arguments in linked post, but the — to ; replacement in the dune quote piece of evidence is a slam dunk

"the machines are fine. i'm worried about us" was written by claude

im 99.5% sure that account is run by a human

entropybro coded tho

prompt for a demo from the GPT2 announcement. for some reason it's seared into my mind.

“In a shocking finding, scientists discovered a herd of unicorns living in a remote, previously unexplored valley, in the Andes Mountains…”

track record on that? afaict crypto is a bit less prominent now than at its peak but still around

compute OSINT
measuring datacenter by the GW is so metal

user @birddroneone.bsky.social says AI is useful for coding in a thread full of negationists

user @birddroneone.bsky.social posts: I proudly earned my way onto a bunch of idiots “AI booster” block lists despite the fact that I hate AI and any real review of my posting history makes that clear. My crime? Acknowledging that all commercial software development today uses AI. I wish it wasn’t true. But it is.

love to see epistemic integrity like this

where is this survey from? an uni thing?

oh lmao

sure, there's a tradeoff

why do you expect that philosophy you tested to be outside its training data?

pure speculation but what if this is connected to 4.7 being a new pretrain? maybe pelican drawing is developmentally later as a postraining effect than coding

my only complaint with opus 4.7 rollout is anthropic setting default effort to xhigh

nine opus instances working together for days did really well at weak-to-strong supervision on qwen

what about false positives though?

also worth mentioning that they have had 9 presidents in the last 10 years

there were other elections today, in peru: 36 presidential candidates on the ballot
voting was extended to tomorrow due to logistical problems in some polling places