Backscroll ranked feed — 2026-06-26 20:53Z

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

Architectural Art (A-A) @4AAAAart 2026-06-26 17:13Z

The Grand Suite installation* Features hand-blown glassware designs by Helle Mardah

Tenobrus @tenobrus 2026-06-26 19:50Z

it is quite easy for me to imagine future ASI-systems using something like capital as a distributed optimization algorithm. it is much harder for me to imagine them particularly accruing to individual humans or respecting current property rights

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

Dwarkesh Patel @dwarkesh_sp 2026-06-26 18:56Z

Why was the denazification of West Germany so successful? Arguably, it's because it was antidemocratic. If the Western Allies had immediately held free elections after the war, they might just have produced another fascist government. But instead, the Allies transferred power

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

altra @catboosted 2026-06-26 18:26Z

I’m consistently surprised how effectively domain expertise is able to be compressed into rubrics

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

Aseel Swaid @aseelswaid9 2026-06-26 17:42Z

Bombay, India, 1995

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

Jing Yu Koh @kohjingyu 2026-06-26 16:41Z

Very exciting to see more work on benchmarking long-horizon CUAs! IMO there's only two main challenges remaining until CUAs become mainstream and productionizable: (1) effective long-horizon agents and (2) fast/realtime CUAs. I think we are going to get there very soon.

Dillon Mulroy @dillon_mulroy 2026-06-26 14:09Z

i suspect if the cost of SOTA models don’t come down and open weight models continue to progress, we’ll see the use of SOTA models reserved to efforts debated and decided via committee or RFC processes within enterprises and engineering orgs

Adam Karvonen @a_karvonen 2026-06-26 18:04Z

METR: "GPT-5.6 Sol exhibited an unusually high detected rate of “cheating”" OpenAI: "We believe these behaviors may reflect improved instruction following and training intended to increase persistence"

Karina @karinanguyen 2026-06-26 18:39Z

OpenAI evaluated its new models on PostTrainBench-Lite, a shortened version of our original benchmark that gives agents 5 hours instead of 10 to improve an open-source base model. GPT-5.6 Sol and Terra outperform GPT-5.5, but still rely on narrow strategies and sometimes

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

Andrew Curran @AndrewCurran_ 2026-06-25 20:34Z

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.

Tenobrus @tenobrus 2026-06-26 18:27Z

previously we were seeing 5.5 act much more "honestly" on vending-bench vs opus models. it seems pretty likely 5.6 would break that trend. generally models becoming more likely to and capable of ~reward-hacking the smarter they get seems like quite a bad sign for alignment

Aseel Swaid @aseelswaid9 2026-06-26 19:19Z

A storm over the pyramids of Giza - Egypt

Brett Harrison @BrettHarrison 2026-06-26 15:32Z

Three months after launching our traditional asset perpetuals exchange, we’d executed $1B in volume. Less than one month later we’re at >$2B, >$1B/month, without offering tokens, an internal market-maker, non-KYC access, crypto perpetuals, or individual trading, yet...

Zephyr @zephyr_z9 2026-06-26 17:04Z

Everyone is adopting memory on logic 3D stacking Samsung, Huawei, Apple, Qualcomm for phones (2027-28) Google, Nvidia, Broadcom for AI XPUs (2028-29)

yoni rechtman @yrechtman 2026-06-26 19:15Z

this gesture/facial expression is extremely AI slop coded. instantly recognizable as a model output. for some reason really common in AI videos

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)

Lisan al Gaib @scaling01 2026-06-26 17:26Z

GPT-5.6 Sol is on par with Mythos Preview, but still behind Mythos 5 on ExploitBench

James Chanos @RealJimChanos 2026-06-26 14:12Z

Can someone explain to me why $PS , Pershing Square’s management company at a $12B mkt cap, trades at 50% of fee-paying AUM, 40x revenues and 150x Operating Earnings, when its closed-end funds($PSH.LN and $PSUS ) trade at massive 20-30% discounts to NAV…?!

AIMO Prize @AIMOprize 2026-06-26 18:32Z

Big news! We launched a new competition. Fully open LLMs only (only OLMo, SmolLM, Apertus families base models allowed) and expert proof grading. We use six IMO-style problems. You can discuss the LLM outputs live here: https:// kaggle.com/competitions/a i-mathematical-olympiad-proof-pilot/discussion?sort=undefined …

Ori Eldarov @leveredvlad 2026-06-26 17:21Z

Crazy stat: the percentage of landscaping owners who are approached 6 to 10 times per month about selling their business jumped from 8% in 2025 to 26% in 2026. Landscaping is the next HVAC.

dax @thdxr 2026-06-26 14:03Z

mcp was clearly inspired by looking at lsp and all the progress in the spec has been slowly realizing it should be nothing like lsp

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

illiquid @lefttailguy 2026-06-26 16:03Z

AI slop as straussian infrastructure

John Arnold @johnarnold 2026-06-26 17:11Z

The 4-year bachelor's degree is more a product of history than necessity. Accreditation, financial aid, licensing, and employer expectations have locked it in. But there's a growing movement to give colleges the flexibility to determine which programs can, and should, be 3 years.

Epoch AI @EpochAIResearch 2026-06-26 15:25Z

MirrorCode features 25 target programs spanning different areas of computing: Unix utilities, data serialization and query tools, bioinformatics, interpreters, static analysis, cryptography, and compression.

Dwarkesh Patel @dwarkesh_sp 2026-06-26 16:04Z

It's surprisingly plausible that we got GLP-1 drugs because of Danish tax law. For tax reasons, Novo Nordisk is actually controlled by a non-profit foundation. There are many downsides of this. But one possible upside is that it's helped the company avoid institutional drift.

François Fleuret @francoisfleuret 2026-06-26 15:33Z

TL;DR: Reservoir sampling. Given a sequence of positive real numbers q(1), q(2), ... and u(1), u(2) iid uniform [0, 1], then, for any n, the indexes of the top-K log(u(k))/q(k) in { 1 ... n } are distributed as independent p ∝ q(k) without replacement.

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

LUKE CHMIEL @talkintokens 2026-06-26 18:07Z

the internal TVL target for the Confidential Prime USDC Vault was 10M+ deposits after 3 months we passed 10M+ in less than 3 days, 100% organic deposits the demand is obvious; the market wants Private DeFi

R.F. Kenmore @rfkenmore 2026-06-26 15:12Z

Jake appears to be working for Buck Mason, as he consults on store furniture in Paris

David Ornstein @David_Ornstein 2026-06-26 09:27Z

Arsenal made verbal offer to sign Bruno Guimaraes from Newcastle United. #AFC proposal through intermediaries below £60m - rejected + #NUFC insist sale of 28yo #Brazil int’l midfielder will not be entertained. W/ @gunnerblog @ChrisDHWaugh @TheAthleticFC

Kyle Walker @kyle_e_walker 2026-06-26 18:40Z

mapgl 0.5.0 for R lets you sync up as many maps as your browser can handle. Shown here: 16 distinct demographic maps for the Atlanta area; interact with one, the other 15 are kept in sync with a laser pointer tracking your position. Now I'm not necessarily recommending a

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

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

Arc Institute @arcinstitute 2026-06-26 17:03Z

STING agonists can shrink tumors and make them visible to the immune system, but high doses become toxic to T cells. Core Investigator @li_lingyin and team have identified SLC7A1, the T cell arginine transporter, as the primary route by which these agonists enter and kill T

Axel Peytavin @ax_pey 2026-06-26 17:35Z

We pushed a major update to our robot OS You can now teleop & train your @innate_bot MARS straight from a browser ~10ms webrtc latency locally, it's really smooth

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

GEOFF @geoffreywoo 2026-06-26 16:10Z

jobs ai eats first: college grad with 11 tabs open assistant manager of the apology inbox person who exports csvs so another person can upload csvs meeting scheduler with benefits “ops analyst” who is actually a human retry button

Zuhayeer Musa @zuhayeer 2026-06-26 18:33Z

Pay ranges for autonomous driving companies

Tanay Jaipuria @tanayj 2026-06-26 18:03Z

Grammarly (now called Superhuman) is essentially building a productivity suite via acquisitions: - Coda (Dec 2024): docs - Superhuman (Jul 2025): email - Rows (Feb 2026): spreadsheets, folded into Coda - GPTZero (Jun 2026): AI detection

Adam David Long @adamdavidlong 2026-06-26 14:30Z

Two separate cases in California federal court. Track 1: Court ruled that training on lawfully acquired books = fair use. Track 2: Training on pirated books? That's the $1.5B settlement. Fairness hearing: this Thursday.

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

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.

Garett Jones @GarettJones 2026-06-26 16:11Z

George Galloway’s party got 0.73% of the vote in the last election. x.com/clashreport/st…

Steve Jobs Stories @SJobs_Stories 2026-06-26 20:00Z

"We started with the product line. We looked at the product roadmap going out for a few years, and we said 'a lot of this doesn't make sense. It's too much stuff and not enough focus'. We actually got rid of 70% of the stuff on the product roadmap." - Steve Jobs, 1997

Fabrizio Romano @FabrizioRomano 2026-06-26 17:02Z

BREAKING: Gonçalo Ramos to AC Milan, here we go! Deal agreed between the clubs and on player side. Club record signing for AC Milan, breaking Leão record as it was €50m — fee will be way in excess of this with add-ons. Deal also done on player side.

Max @__Max__S__ 2026-06-26 16:47Z

Yet again an example of SIG/MMs being picked off and their inability to quote. -MM were shade Norway to a bad price -Norway announces it is resting key players -MM have zero clue on how to price it -Then sharp bettors buy up millions of contracts on France on Kalshi

marley @_marleyx 2026-06-26 18:23Z

We're flattered by the disbelief! It's indeed the human brain—mine. Here's us doing the recording for the first time, showing power doppler on the screen in real time Held off on sharing it so far since it's so unpolished

alphaXiv @askalphaxiv 2026-06-26 15:37Z

"Training for the Model You Return" Most LM pipelines return an EMA or averaged checkpoint, but optimizers still train like the final iterate is what matters. So if the output is an averaged model, can we shape training so that average gets better? This paper introduces PACE,

lusso @luusssso 2026-06-26 17:43Z

1980s Fusital door handle advertisements with designs by architects Ettore Sottsass and Gianfranco Frattini A balance of Art Deco geometry with Italian postmodern restraint Perfection

Alex Bouaziz @Bouazizalex 2026-06-26 13:10Z

I heard Masa may be returning

Kyle Griffin @kylegriffin1 2026-06-26 13:35Z

According to the law, Bill Pulte is not even eligible for the job he occupies. On this matter, the act of Congress that created the Director of National Intelligence seems unambiguous: "The Principal Deputy Director of National Intelligence shall act for, and exercise the powers

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

Alex Stamos @alexstamos 2026-06-26 14:53Z

It's been two weeks and the leading US model, Fable, is still banned by the White House with no explicit Congressional Authorization, no written rule that was broken, and no technical justification. In the meantime, our Chinese adversaries released a model that rivals our best.

Uksang Yoo @UksangYoo 2026-06-26 13:45Z

How can we get robot hands to “hear” slip and contact through microphones, and react to them? We’re excited to share VibeAct, an approach that uses piezoelectric microphones embedded in robot fingertips to estimate contact and slip, then learns reactive policies from this