Backscroll ranked feed — 2026-06-27 01:59Z

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

Bloomberg @business 2026-06-26 22:38Z

SpaceX and Charter have held executive-level talks about partnering on a consumer mobile phone offering, according to sources

Aseel Swaid @aseelswaid9 2026-06-26 22:57Z

Tajás, 1988. São Luís, Maranhão. Luiz Braga.

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

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

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

Andrew Curran @AndrewCurran_ 2026-06-26 23:25Z

Semafor is reporting that The US government has lifted its block on Mythos 5 in a letter this afternoon from Commerce Secretary Howard Lutnick to Tom Brown. Fable is not included. This is US only, and covers 100 US institutions, including major companies and government agencies.

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

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

AI slop as straussian infrastructure

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

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

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

Richard Crowley @rcrowley 2026-06-26 23:40Z

Honest to god, I got an InsufficientInstanceCapacity error for a t4g.nano in a test suite today. The cloud is FULL.

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

Fabio Guzman @FGuzmanAI 2026-06-26 23:14Z

I burned a full Transformer into custom silicon. And Andrej Karpathy liked the project. When someone whose work inspired your own takes a moment to notice it, every late night suddenly feels worth it. Back to building.

rat king @MikeIsaac 2026-06-26 20:34Z

more news: Mark Zuckerberg is telling lieutenants to pursue partnerships with Polymarket and Kalshi with Meta's new prediction market app called Arena, per sources Meta is targeting 18-34 year old users with a goal of 100 million regular "predictors"

Aseel Swaid @aseelswaid9 2026-06-26 20:42Z

Yanomami man and hummingbird. Demini village, Roraima, 1991. Rosa Gauditano.

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

Dwarkesh Patel @dwarkesh_sp 2026-06-26 21:56Z

Suppose an AI model fully automates a $200k job. Its economic value - measured in contribution to GDP - will be both far more and far less than $200k. There are two ways of thinking about the model's economic value: (1) It'll be far more than $200k. After all, the human

Architectural Art (A-A) @4AAAAart 2026-06-26 16:37Z

Calabasas Residence of Maeve Reilly and Zach Quittman by Sarah Solis Calabasas, California, USA

Chelsea Finn @chelseabfinn 2026-06-26 22:53Z

LLM RL optimizes for sequential reasoning We also optimize over the reasoning strategy, incl parallel trains of thought, aggregation of parallel traces, & sequential reasoning This allows the model to better explore & allocate compute at test time https:// arxiv.org/abs/2606.23595

junkbondinvestor @junkbondinvest 2026-06-26 20:44Z

Apollo, a large manager, put out a chart showing deep distress is concentrating in small private credit funds. Loans marked below 50: ~12.5% at small funds, ~8% at big ones

Tenobrus @tenobrus 2026-06-26 18:22Z

btw if u thought ur custom fanned-out data processing multi-agent orchestration harness was a particularly relevant long term moat, sorry, openai just added it as an api parameter

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

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 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"

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.

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

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

nico @nicochristie 2026-06-26 19:06Z

You should not see this and think Mythos is ahead I think this is the most exciting model ever 5.6 Sol is in its own league on the Pareto frontier. Speed and efficiency are measures of intelligence too.

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

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

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)

Thomas H. Ptacek @tqbf 2026-06-26 21:58Z

I don't think I'm ADHD, but I do believe that in 2026, ADHD has become a software developer superpower, if you know how to hold it, because follow-through is now a utility-priced commodity.

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 …

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.

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

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

Isaac King @IsaacKing314 2026-06-26 23:43Z

I see we're doing Pangram discourse again so here's a point against their "low false positive rate". (Yes, I wrote this myself.)

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

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

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

Gary Bernhardt @garybernhardt 2026-06-26 23:56Z

A type & test that GPT 5.5 added, without further comment: export type Root = { url: string; }; it("exports only shared docs domain types", () => { expectTypeOf<Root>().toEqualTypeOf<{ url: string; }>(); [other useless assertions here] })

Matt Slotnick @matt_slotnick 2026-06-26 18:41Z

claude tag access model is confusing... you auth claude once to your systems via global api key, and then any person can tag claude, which now has global system access with no permission or scope enforcement? and i don't see any placeholder for that to exist. what am i missing?

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

Spacesthetic @interiorsuckerr 2026-06-26 17:11Z

Handblown glass pendants by Jamie Harris Studio

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)

Jayden Teoh @jayden_teoh_ 2026-06-26 23:45Z

Nemotron 3 Ultra uses a shared-weight MTP head that allows the same head to be applied recursively for drafting. However, for this to work, it requires a separate "MTP Boosting" phase—on-policy distillation phase—so that the model learns to react to drifts in hidden states not

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

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

Andrew Qu @andrewqu 2026-06-26 22:10Z

if you sketch out all of the systems required to make vercel run, you will come to the realization that as long as the rollback service is not circularly dependent on vercel everything else can ~literally be on vercel

CG Ruthless1990 @cgruthless1 2026-06-26 22:49Z

The 49ers haven't cut Brandon Aiyuk because he is currently on the Reserve/Left Squad list, meaning his contract is effectively frozen and he is not costing the team a roster spot or any salary. Additionally, Aiyuk has yet to submit the reinstatement paperwork required to return

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

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

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

Poore's law

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.

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

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

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…?!

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.”

Aaron Gokaslan @SkyLi0n 2026-06-26 21:42Z

Supply chain risks and tokeonomics. It's that simple.

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