layer4down

subagents = stack
files = heap

Interesting way to think about. 👍

Side‑by‑side pair of line charts comparing how GPT‑5 (left, in blue) and an RLM system built on GPT‑5 (right, in orange) handle extremely long input prompts on three evaluation tasks: S‑NIAH, OOLONG, and OOLONG‑Pairs. Both charts share the same vertical axis “Score (%)” from 0 to 100 and the same horizontal axis “Input Context Length (log scale)” with labels 8k, 16k, 33k, 66k, 131k, 262k and then very long ranges 524k and 1M tokens.[arxiv] In the left GPT‑5 panel, three blue lines show performance dropping as context grows. The S‑NIAH line with circles starts near 100% and stays flat until about 131k tokens, then enters a pink shaded region beyond the model’s native window where it is no longer plotted. The OOLONG line with triangles begins around the 80–90% range but steadily declines toward roughly the mid‑40% range by 131k. The OOLONG‑Pairs line with squares starts lower and falls close to 0% well before the longest lengths, illustrating that pairwise long‑context reasoning collapses even when the text still fits in the context window.[arxiv] In the right RLM(GPT‑5) panel, the layout is the same but curves are orange and extend into a green shaded region covering 262k to 1M tokens, emphasizing ultra‑long contexts. The S‑NIAH line with circles remains essentially flat at or near 100% across the entire range out to 1M tokens, showing robust “needle‑in‑a‑haystack” retrieval at any length. The OOLONG and OOLONG‑Pairs lines, drawn with different orange markers, still decline somewhat but stay in a moderate accuracy band instead of crashing, indicating that the recursive setup substantially mitigates the degradation seen in the plain GPT‑5 chart.

I used to think that until I read MIT’s paper on RLM in which they showed the impact a complex prompt can have even in a frontier model at just 16-32k. System prompts alone can eat that up. We can prune and subtask and engineer around it but it’s nice to have options.

arxiv.org/html/2512.24...

@vastatepolice.bsky.social the URL is broken for the #amberalert you just issued. Please fix and also post here.

I Trained With Russian Diplomats. I Can Tell You How They Work - The Moscow Times Opinion | Last week, while Russian missiles rained down on Ukrainian cities, the Kremlin’s special envoy Kirill Dmitriev landed in Washington with an offer of economic cooperation.

#OpEd Inna Bondarenko takes you inside Russia's famous school for diplomats, where Moscow's negotiators of the future are taught to break the rules of international relations while appearing to play by the book.

Here's what she learned and how she got out.

”This is a warning.” The West should listen.

"If the West wants to deal effectively with 🇷🇺, it needs to understand how its diplomacy works," says researcher Inna Bondarenko, who studied at 🇷🇺 (MGIMO), an elite Russian Diplomatic Academy often referred to in the West as the "Russian Harvard."

1/9🧵

a woman wearing a white turban and earrings looks serious Alt: a woman wearing a white turban and earrings looks serious, tells off screen person “go ask your father”.

At least that person is not delusional. Maybe even reachable. Tho, I thought I was getting thru to my own cousin last year and he couldn’t wait to brag that “Uncle tRump is back baby!” the morning of 11/6 (and had the nerve to ask to borrow money at the same time!). Y’all already know my response 😏

And luckily no one who lost everything already has enough struggles. 👍

He should wrap and rebadge it as a a Chinese brand. Restore some honor.

Defend Castro? Wow that's an excellent point! 😂

spongebob is standing in front of a sign that says krabby patty Alt: spongebob is on his knees bowing and worshipping in front of a sign that says krabby patty

JFC 😖 🤦‍♂️ still cradling Putin’s b@llsack after what he did to Melania on R1 Russian TV? What a cuck

a close up of a snow leopard with a national geographic logo behind it Alt: a close up of a snow leopard licking its chops with a National Geographic logo behind it

*ahem* “…mi-mi-mi-mi-mihhhh”

And take on $7M in ONE week for our troubles 😊

It’s like he’s cussing himself out 😭

A photo of Secretary of Defense Pete Hegseth with caption “D.U.I. Hire”.

Hegseth and Chill.

FBI Uncovers Al-Qaeda Plot To Just Sit Back And Enjoy Collapse Of United States WASHINGTON—Putting the nation on alert against what it has described as a “highly credible terrorist threat,” the FBI announced today that it has uncovered a plot by members of al-Qaeda to sit back an...

FBI Uncovers Al-Qaeda Plot To Just Sit Back And Enjoy Collapse Of United States

Have a written record of society’s collapse.

Join America's Finest Membership Here:
membership.theonion....

A photo of a black man spanking a Klansman in broad day light ca. 1980’s. Caption reads: “Let us never forget Willie ‘Whoop Ass’ Johnson, the first Black man to whoop a Klansman on camera and not give a fuck. Happy Black History Month!”

Happy Black History Month!
#blacksky

What apologies? 😂 Anyway, we don’t know the cousin’s situation. I know people a few years away from retirement who might be pushed out that’s a pretty scary thing, considering they’ve devoted to past two decades to public service! Yeah, no apologies. 😘🖕✌️

a man wearing a striped sweater is sitting on a couch Alt: a man wearing a striped sweater is sitting on a couch eating popcorn, watching the show.

I love this tough guy routine 😂. Can’t WAIT to see how they respond! Trying to imagine MBS bending the knee 😆

Any other election cycle I’d say “oh well, we’ll get ‘em next time.” This one just feels wayyy different and unfortunately it requires removing the kid gloves. We thought swing voters got it after last midterms but those lessons did not hold two more years. We were dead wrong. So a change is needed.

I know your approach because it was my approach for years and what I learned is that it doesn’t work because these people aren’t connecting their choices to consequences. We can not convince them with logic or empathy. They only understand self-interest. And theirs are coming to a head with reality.

I’m on the opposite end of the spectrum. I know constant positivity and it’s exhausting. I’m not angry. I’m quite resolved. It’s only been a week and that’s not nearly enough time for the consequences of their actions to have truly sunken in yet (they’re not THAT smart). And it’s just the beginning.

a man in a suit is standing in front of a crowd and says oh no . Alt: a man in a suit is standing in front of a crowd and says “oh no! Anyway…”

Now they’re getting it worst and first.

No. You don’t get to do that. We get to be assholes. We’re on vacation for the next four years because we told those ppl and they refused to listen. “Trump derangement syndrome”? No. Trump engagement syndrome. They wanted it and now we’re just giving it to then 24/7/365 x 4 + 1 (leap year). Fk ‘em✌️

I think most of us tried so hard for so long not to lean into tribalism that now that we’re here, it’s the only solace remaining. Because it’s that or capitulation. I’m not convinced that these people don’t want this because it’s only been a week. We need to give it time to really sink in for them.

Who is Liang Wenfeng? DeepSeek founder comes from AI investing DeepSeek, a Chinese AI lab, has Silicon Valley reeling with its R1 reasoning model, which it claims uses far less computing power than those of American AI leaders — and, it’s open source. The DeepSeek app has shot to the top of the App Store charts this…

Who is Liang Wenfeng? DeepSeek founder comes from AI investing

Hey let’s go a step further. We should end the state governments and local governments and let the broligarchs duke it out neo-feudalism style, amirite? Make the Middle Ages Great Again! #neoserfdom

@deepseek.bsky.social what’s was up with that API #whalefail yesterday?

On the 671B model? Or lower precision models?