Backscroll ranked feed — 2026-06-26 11:34Z

A frozen ranked feed of 60 substantive tweets published on Backscroll (2026-06-26 11:34Z).

Conquest's Third Law says that you should model any bureaucratic organisation as if it was controlled by a cabal of its enemies. It's a fun exaggeration, but it gets at an important phenomenon. Organisations have continuous generational turnover. The founding members eventually
23 likes · 3 retweets · 3 replies
The US Government has requested a slow staggered rollout of GPT-5.6, and OpenAI has agreed. During this phase the government will approve each user individually. This will probably be the norm for all frontier models from all labs from now on.
2302 likes · 492 retweets · 224 replies
Look at how that green (teal? cyan? celes? edgar?) line at the bottom is flat as hell—so even in times of increasing total homelessness like post-pandemic, the people actually living rough on the streets is pretty consistent and has been for twenty years. There’s a bit of a drop
1 likes · 0 retweets · 1 replies
one of the craziest bets of prime intellect was swinging for everything from the start. we operate datacenters, we build our own trainers and envs and sandboxes, we have a platform for large enterprises and solo devs, we train huge models, we publish. ~40 people btw. hiring, DM
177 likes · 15 retweets · 13 replies
Custom Land Rover Defender known as "Baloo" designed and built by The Landrovers
206 likes · 24 retweets · 1 replies
hmm okay so now Ornith-1.0-35b-a3b is Qwen3.7-35b-a3b then? benchmark looks good, will give it a spin
16 likes · 1 retweets · 2 replies
If we look at the percent of unhoused homeless as a percent of the US total population, we see a shocking stability across time and our current homeless population is returning to the pre-Millennial dip of the 2010s. That makes me wonder if there is a kind of min-max
1 likes · 0 retweets · 1 replies
A woman harvesting roses - Oman
88 likes · 10 retweets · 1 replies
but there is a very narrow narrow path that passes through "most of us leading overall pretty normal and happy lives right up through aligned singularity" and this is a step that stays on it
23 likes · 0 retweets · 1 replies
Very cool work with connections to the schedule free work and iterative averaging work. Interestingly, a natural step is to derive it for vanilla shampoo because it keeps the pre-conditioner around (b2>0) around. Essentially replace diagonal with the kronecker factors Could be
23 likes · 2 retweets · 1 replies
OpenAI’s ARR stood at $40bn as of June.
112 likes · 11 retweets · 10 replies
this is an arrangement of 50 Cent's "In Da Club" by my pianist michael potter. somehow he turned this certified hood classic into an epic that's in 11 and has metric modulations all over the place. but yeah other than that it sounds exactly like the original
6 likes · 0 retweets · 0 replies
finally read the OSTP NSTM-4 memorandum so proudly cited by anthropic, and turns out the entire thing is just a self-own whine-fest about getting commoditized at a fraction of the cost by maliciously exploitative chynese actors who dare remove the performative "security
80 likes · 7 retweets · 9 replies
these people want ai models to be treated like nuclear weapons, researchers treated like they own nuclear state secrets during ww2, and geopolitical tensions escalated until we get ww3, where all production power can then be seized for the "ai industrial revolution" they strive
69 likes · 11 retweets · 4 replies
Looks like anti-cheat software (GameGuard) is causing bun to crash. It’s interfering with system calls. https:// github.com/oven-sh/bun/is sues/22747 … GameGuard also causes other software to crash (example https:// issues.chromium.org/issues/40762507)
37 likes · 0 retweets · 4 replies
Poore's law
332 likes · 5 retweets · 7 replies
we built HANDBOOK.md: a new benchmark for long-context agentic instruction following. we drop an agent into a live company environment with files (PDFs, Excel, Word Docs, ...), tools (email, Slack, Jira, calendar, ...), and a dense corporate handbook (up to 124 pages), across 5
10 likes · 2 retweets · 0 replies
imo every autoresearch is trying to figure out "the scaling law of the scaling law". - scaling law tells you what happens once you have fixed a config. It does not tell you how to train in the first place. - a scaling recipe is the missing map. It says how width, depth, batch
2 likes · 0 retweets · 0 replies
he’s joking but this is what the coasean boundaries lead to: a firm size of “everybody”; ultra space communism. if you want to find another way out you need to give it a hard think
101 likes · 3 retweets · 14 replies
it's not necessarily a victory for safety in the more absolutist unaligned ASI risk sense. but i do think these regulatory actions are probably safety victories wrt bio and cyber and general short-term mere megadeaths and societal disruption risk. if the USG has woken up enough
83 likes · 7 retweets · 12 replies
Overheard in Florida: “The West doesn’t exist. It died with the French Revolution. America supplanted it.”
330 likes · 21 retweets · 34 replies
They bought distressed debt at ~1 cent on the dollar and got headline credit for the sticker price. Obviously great that they spent $5.5M getting people out of debt but all the headlines are about $550M
36 likes · 2 retweets · 3 replies
A couple walks down a street in Sana'a, Yemen, 1997
56 likes · 5 retweets · 1 replies
the reason we excluded frontier models from our data page is they are artificially low in usage and didn't want people concluding this most people using them use them directly, they don't buy them through us we wouldn't be surprised if they're 90%+ of total token spend
7 likes · 0 retweets · 2 replies
Kraken in talks to buy 15% stake in DeFi lender Aave at $385 million valuation
45 likes · 17 retweets · 14 replies
believe an information theoretic view of vertical integration is helpful in this era (and probably could be further refined) fwiw
0 likes · 0 retweets · 0 replies
No joke: when i saw this tweet earlier today was soooo close of digging up your tweet from a few years ago where you basically introduced the term to the world by saying "we're hiring for midtraining team" Even funnier is Google now also recently has an Omni team!
19 likes · 0 retweets · 2 replies
71% of teams that switched on our self driving mode after we invited a bunch of customers yesterday have already merged a pr that they didn’t have to decide to build or ship themselves! Super excited to see the high conversion rate
17 likes · 1 retweets · 4 replies
There is an element of that sophisticated, urbane (yet decaying and evil to the core) Old World vampirism feeding on Lolita through Humbert. Similar to what happens to Daisy Miller in Henry James’ study; she gets courted by a man literally named “Winterbourne,” wilts, and dies.
3 likes · 0 retweets · 0 replies
While we eagerly await Fable 5's return, our agentic WebGPU kernel optimization framework kept running. Opus 4.8 picked up where Fable left off, pushing Liquid AI's new LFM2.5 230M to an unbelievable 1,400 tok/s... running locally in your browser. Don't blink or you'll miss it.
113 likes · 21 retweets · 14 replies
I hate “AI is bad because it becomes autonomous” genre. Anthropomorphizing does a great disservice relaying the actual risks Class power is about control over scarce resources. Everyone’s fighting for land because, as the saying goes, they’re not making more of it
0 likes · 0 retweets · 0 replies
ProgramBench is now accepting submissions to the official leaderboard! We also open sourced *all* of the 2000+ agent trajectories and codebases from our paper’s main results.
36 likes · 10 retweets · 2 replies
2026 has produced more $KOSPI circuit breakers in under 7 months (5) than the previous ~25 years combined (one in 2001, two in the 2020 COVID crash). 3 of this year's 5 came this week How levered are these Koreans?
35 likes · 4 retweets · 4 replies
This is the pinnacle of Chinese watchmaking The "Fam al Hut" Möbius This watch blows Richard Mille out of the water.. a 32000$ biaxial tourbillon, double retrograde, jumping hour, just how freaking insane is that... Makes me really happy that they won the GPHG award, because
199 likes · 8 retweets · 12 replies
Feynman is likely delayed to 2029 if they are adopting CoPoS and glass substrates (Innolux is hinting towards that)
39 likes · 3 retweets · 1 replies
Yes. The degree to which chat is really being relegated. And codex / Claude code have gigantic motion even outside engineering is very true
6 likes · 0 retweets · 0 replies
What you guys are referring to as MCP, is in fact a Spec/Protocol, or as I've recently taken to calling it, a Spec plus a Protocol. The spec of an MCP server is flexible, can be represented through lazy loading, code mode, as a CLI whereas the protocol is a fully functioning
64 likes · 3 retweets · 8 replies
3/ Amazon's AgentCore Runtime gives each rollout its own sandboxed MicroVM and scales hundreds of sessions per training step. That keeps arbitrary repo builds isolated, moves CPU-heavy Maven/JUnit work off the GPU node, and gives us per-session logs and traces.
1 likes · 0 retweets · 1 replies
New: Granular evidence citations in QA Beyond chunk-level citations, our QA now points to the exact evidence snippet supporting each answer. Easier to inspect, verify, and trust. → Available with high-quality processing.
5 likes · 2 retweets · 1 replies
When a quantized reasoning model fails, we usually assume it just lost the capability to solve the problem. But we surprisingly show that in up to 52% of failures, the quantized model actually reaches the correct answer but then does not commit to it. 1/8
9 likes · 4 retweets · 1 replies
Plots of empty desert could soon be worth more than the most fertile farmland in human history. @CJHandmer ran through the back-of-the-envelope math. 10 acres of solar is enough for 1 MW of H100s. And the FLOPS done by those H100s is roughly equivalent to 1000 human brains. If
450 likes · 60 retweets · 36 replies
> get super stoked > 328 tps > wow that's sick > sign up for account > run quick benchmark bc I'm not a trusting person > it's 58 tps the same as everyone else thanks for playing https:// x.com/Yuchenj_UW/sta /Yuchenj_UW/status/2070166719839326396 …
556 likes · 13 retweets · 30 replies
JUST IN: $GOOGL is reportedly turning its AI coding strike team into a formal “midtraining” group as it tries to close the gap with Anthropic after recent AI talent exits.
258 likes · 11 retweets · 13 replies
30m x 1m output token trace is like upper bound of $750m revenue at 25 dollar per million output tokens. As an upper bound. https:// x.com/mtslive/status /mtslive/status/2070141140607832353 …
204 likes · 5 retweets · 9 replies
rebuilt the figma "vector editor" for the web draggable points and cool morphing shapes code: http:// arlan.me/vault
139 likes · 3 retweets · 8 replies
Many institutions are now wondering if trading is going to move onchain. The crypto industry itself might be a tell tale. Onchain DEX trading is now 25% of all trading of crypto assets. Up from practically 0 back when we started tracking this 8 years ago with @Dune
7 likes · 0 retweets · 1 replies
announcement: who is building ai that makes a buyer say this in diligence: “we used to need 4 people to clear this queue before noon.” not a prettier dashboard. cycle time down, rework down, escalation rate visible. send it.
2 likes · 0 retweets · 1 replies
And GLM synth pipelines now plugged on batched dockers (ok singularity since that’s what we can install on cluster — aptly named).
23 likes · 0 retweets · 2 replies
The cool result is the usage across all departments incl eg legal increasing massively (Codex vs other usage, ie chat) Codex and Cowork are a great unlock for all knowledge work (ala Karpathy) but most are not used to this yet
4 likes · 0 retweets · 1 replies
Excited to share new @LaudeInstitute project, where we ask: When does continual learning actually require learning? We test 8 adaption methods as the world shifts and find ones best at learning new facts are worst at unlearning what's now outdated.
9 likes · 2 retweets · 2 replies
Cannot emphasize how much this unchanged code block detection reduces cognitive load when reviewing long diffs. 2nd screenshot is GitHub PRs for comparison
296 likes · 5 retweets · 9 replies
Weirdly, nobody has figured out how to have Modernization without Westernization. Not Showa Japan, not Nasser’s Egypt. Nobody. As far as we can tell, if you set up a modern state and lay the groundwork for industry, a few decades later you’ll have young urban women wearing jeans
1074 likes · 46 retweets · 32 replies
I am. Once again. Asking for faster token output and reliability.
9 likes · 0 retweets · 0 replies
We can learn a model that provides shaped "process rewards" for robotic RL, that evolves automatically as the policy gets better. This improves performance on benchmarks, and works in the real world! Some fun new work with Raymond Tsao & @ajwagenmaker
1 likes · 0 retweets · 1 replies
A Palestinian woman and an Arabian horse. Captured by Aseel Swaid
12 likes · 2 retweets · 0 replies
This is probably the most favorable policy for China, opening the door to international markets for Alibaba's PPU and Huawei's Ascend. China never planned to make money from large models in the first place; their strategy is to dump data center hardware to address the issue of
15 likes · 0 retweets · 7 replies
GLM 5.2 is 2nd in Vending-Bench. Each GLM release has improved at a remarkably steady pace: a linear fit of R²=0.99, with almost $1k better per month.
149 likes · 15 retweets · 2 replies
Meh $800k ~40k a month take home for two people no kids is when you can start living very comfortably in NYC. $10-12k housing, $8-10k food, drinks, random fun stuff, $5k vacation accrual, $10-20k savings. Also assumes on continued upward finance trajectory / random carry check
64 likes · 0 retweets · 9 replies
It's tempting to think that the lottery ticket hypothesis requires an exponentially large network for a subnetwork to win the lottery. This is incorrect--a small network suffices! We tend to imagine each subnetwork as throwing a dart in high-dim space to match a target
12 likes · 0 retweets · 1 replies
jira is like a harness for an organization
61 likes · 4 retweets · 13 replies

How it ranks: Backscroll reads my Twitter/X timeline and ranks every tweet for substance with an LLM taste rubric — not raw engagement. Curated by Surya Dantuluri.