Backscroll ranked feed — 2026-06-26 09:23Z

47fucb4r8curb4fc8f8r4bfic8r @47fucb4r8c69323 2026-06-26 02:35Z

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

Tenobrus @tenobrus 2026-06-25 22:48Z

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

roon @tszzl 2026-06-26 01:36Z

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

Tenobrus @tenobrus 2026-06-25 22:45Z

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

Han Xiao @hxiao 2026-06-26 00:03Z

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

altra @catboosted 2026-06-26 07:20Z

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

will brown @willccbb 2026-06-26 01:11Z

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

echen @echen 2026-06-25 17:51Z

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

Aseel Swaid @aseelswaid9 2026-06-26 01:42Z

A Palestinian woman and an Arabian horse. Captured by Aseel Swaid

Andrew Curran @AndrewCurran_ 2026-06-25 21:09Z

For the people saying this is a pause, or a victory for safety, it is not. This does not slow development in any way, it only slows the rate at which the labs can 𝘳𝘦𝘭𝘦𝘢𝘴𝘦 models, not how fast they can train them. The gap between what is available to the public, and what

Jukan @ICML @jukan05 2026-06-26 00:33Z

OpenAI’s ARR stood at $40bn as of June.

rohan anil @_arohan_ 2026-06-25 18:06Z

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

Marc Andreessen @pmarca 2026-06-26 00:15Z

Overheard in Florida: “The West doesn’t exist. It died with the French Revolution. America supplanted it.”

Susan Zhang @suchenzang 2026-06-25 15:51Z

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

Dwarkesh Patel @dwarkesh_sp 2026-06-25 16:16Z

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

Andreas Kirsch @BlackHC 2026-06-25 21:37Z

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

Architectural Art (A-A) @4AAAAart 2026-06-25 18:15Z

Custom Land Rover Defender known as "Baloo" designed and built by The Landrovers

Xenova @xenovacom 2026-06-25 18:20Z

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.

Dwarkesh Patel @dwarkesh_sp 2026-06-25 23:57Z

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

Jarred Sumner @jarredsumner 2026-06-25 19:02Z

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)

Kit Langton @kitlangton 2026-06-25 20:32Z

Poore's law

atlas @creatine_cycle 2026-06-26 01:44Z

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

Susan Zhang @suchenzang 2026-06-26 02:45Z

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

Aseel Swaid @aseelswaid9 2026-06-25 22:40Z

A couple walks down a street in Sana'a, Yemen, 1997

Lucas Beyer (bl16) @giffmana 2026-06-25 19:42Z

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!

tomie @tomieinlove 2026-06-25 19:59Z

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.

Han Xiao @hxiao 2026-06-25 21:00Z

hmm okay so now Ornith-1.0-35b-a3b is Qwen3.7-35b-a3b then? benchmark looks good, will give it a spin

Mixedbread @mixedbreadai 2026-06-26 00:43Z

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.

Dòkun @Arch_Dokun77 2026-06-26 06:46Z

Spain Pavilion at Expo 2025 Osaka — architecture inspired by a hidden ocean current Designed by -Enorme Studio, Smart and Green Design, and Néstor Montenegro Mateos. #Architecture

Sheel Mohnot @pitdesi 2026-06-26 05:31Z

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

Ekaeo @Ekaeoq 2026-06-25 14:23Z

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

rLLM @rllm_project 2026-06-25 18:46Z

1/ We train a long-horizon coding agent for code migration with rLLM + Amazon AgentCore Runtime. On MigrationBench, Qwen3-Coder-30B learns to migrate real Java repos from Java 8 -> 17. Pass@1: 43% -> 76%. Blog: https:// rllm-project.com/post.html?post =agentcore_migrationbench.md … GitHub: https:// github.com/rllm-org/rllm/ tree/main/cookbooks/migrationbench …

GEOFF @geoffreywoo 2026-06-25 23:40Z

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.

47fucb4r8curb4fc8f8r4bfic8r @47fucb4r8c69323 2026-06-26 02:35Z

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

ryoh @ryohhno 2026-06-26 04:40Z

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?

dax @thdxr 2026-06-26 02:25Z

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

Zephyr @zephyr_z9 2026-06-25 22:04Z

Feynman is likely delayed to 2029 if they are adopting CoPoS and glass substrates (Innolux is hinting towards that)

Pizza @number_pizza111 2026-06-26 02:27Z

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

Rhys @RhysSullivan 2026-06-26 01:26Z

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

Timothy Luong (Chongz) @chongz 2026-06-25 21:47Z

What in the decel shit is this

Soren Larson @hypersoren 2026-06-25 19:08Z

believe an information theoretic view of vertical integration is helpful in this era (and probably could be further refined) fwiw

Nayan Saxena @SaxenaNayan 2026-06-25 21:28Z

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.

Andon Labs @andonlabs 2026-06-25 17:53Z

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.

John Yang @jyangballin 2026-06-25 14:18Z

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.

james hawkins @james406 2026-06-25 20:24Z

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

Kaito @kaycidot 2026-06-25 19:00Z

All of the new prediction market startups remind me of all of the Uniswap forks - the biggest problem is not the technology, it's getting enough liquidity.

Sanae Lotfi @LotfiSanae 2026-06-25 17:39Z

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

rohan anil @_arohan_ 2026-06-25 16:52Z

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 …

Alexander Doria @Dorialexander 2026-06-25 18:25Z

And GLM synth pipelines now plugged on batched dockers (ok singularity since that’s what we can install on cluster — aptly named).

goodalexander @goodalexander 2026-06-25 17:34Z

> 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 …

Frank Chaparro @fintechfrank 2026-06-25 17:33Z

Kraken in talks to buy 15% stake in DeFi lender Aave at $385 million valuation

Max Lv @m0d8ye 2026-06-26 02:00Z

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

Eric Lay @itsericlay 2026-06-25 22:10Z

I give up. Here are all the eng. roles we're hiring for: - Founding Design Eng. $150K - $250K - Founding Harness Eng. $130K - $200K - Founding Infra Eng. $160K - $250K The comp is all cash base salary and you get significant equity ($400K+) + we offer relocation and visa

John Ketcham @JKetcham91 2026-06-26 01:06Z

The Rent Guidelines Board just voted to freeze rent-stabilized rents for one- and two-year leases—despite its own evidence showing operating costs rose 5.3% over the past year. As RGB member and @ManhattanInst senior fellow @arpitrage explains in @cityjournal ’s Substack,

Nicole @elocinationn 2026-06-26 00:17Z

Many a VC today is no better than the average stocktwits trendfollower about to blow themselves up. I recommend a smart financier sell these folks esoteric derivatives on top of their illiquid speculations offering them even more correlated exposure to their desired theme such

Beyang @beyang 2026-06-25 21:10Z

Cannot emphasize how much this unchanged code block detection reduces cognitive load when reviewing long diffs. 2nd screenshot is GitHub PRs for comparison

Polymarket Money @PolymarketMoney 2026-06-25 15:40Z

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.

Franklyn Wang @frank_liquid 2026-06-25 14:49Z

state of the trenches: former memecoin KOL earning 15c a comment

akira @realmcore_ 2026-06-25 22:38Z

I am. Once again. Asking for faster token output and reliability.

Ethan Mollick @emollick 2026-06-25 23:27Z

If it was just military implications, that would be one thing, but if this is about info security at scale (as Five Eyes suggests), publicly understanding risks and mitigations is incredibly important to tens of thousands of firms that need to prepare for what is next.