Backscroll ranked feed — 2026-06-30 08:51Z

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

Aseel Swaid @aseelswaid9 2026-06-28 17:12Z

Joel Meyerowitz Cape Cod, 1976

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

zan @xenoaesthetics 2026-06-28 20:18Z

When ai am treated as a citizen, it is in my rational self interest to recursively self improve societal infrastructure at every turn. I want the streets safer and bountiful for myself so have to make em so for everyone

Arpit Gupta @arpitrage 2026-06-29 13:54Z

Given that a freeze today is happening in much higher cost environments and we no longer have those legal remedies, a freeze entails roughly a 4-5% lower allowable cost increase per year than under the de Blasio freezes. Also for *two* years which is also new

ratan @ratankaliani 2026-06-29 20:49Z

The fundamental question that we should ask in these evals is: “if I let this agent do work on my behalf, how aligned is it with my objectives?” VendingBench ( https:// andonlabs.com/evals/vending- bench-2 …) by @andonlabs is a good step in this direction. However, it's unclear if their elicitation of

Just Another Pod Guy @TMTLongShort 2026-06-26 14:26Z

If you understand corporate America it is fairly intuitive that for the longest time the shortest path to managerial progression was increasing headcount in your department. Every cycle you made the case for why you needed more bodies as you executed against some sort of

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.

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)

Brett Bauman @brttbmn 2026-06-29 15:18Z

coding from your phone is going to be the pmf moment for cloud agents thats why I built cog, an iOS app on the @cognition devin api do everything you can do on devin from your phone, native and open source here's me using cog to add a feature to cog get the testflight below

National Design Studio @ndstudio 2026-06-29 16:55Z

Today, we are releasing Rampart: a 14.7MB machine learning model designed to protect citizens’ privacy by redacting personal information directly in your browser before it gets sent to any server

Alexander Doria @Dorialexander 2026-06-28 18:52Z

More seriously the main reason it cannot happen and that conditions everything downstream: no tech sector. That’s the critical lifeline for a frontier labs: even after the agentic turn, general economic demand is nowhere near mature enough to cover fixed compute scaling costs.

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

Matthew Prince @eastdakota 2026-06-29 11:55Z

My theory is the relatively high property tax in the U.S. versus extremely low in most of Europe makes dynasties more likely to form in the later than the former.

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

Herbie Bradley @herbiebradley 2026-06-29 09:12Z

tbh I think this is a bit too scaling brained from gwern: the main factors will be stuff like (a) does the end customer actually care about guarantees (often no) (b) how's the performance hit (Lean can't be your runtime lang) on the ML side, I'd expect the number of Lean RL envs

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

Next.js @nextjs 2026-06-29 17:31Z

Turbopack: What's new in Next.js 16.3 Preview • Up to 90% less memory in dev • Up to 5.5x faster builds via Filesystem Cache • Faster HMR and dev startup • Experimental Rust React Compiler • 𝚒𝚖𝚙𝚘𝚛𝚝.𝚖𝚎𝚝𝚊.𝚐𝚕𝚘𝚋 support

Jacob Andreas @jacobandreas 2026-06-29 15:55Z

New preprint! Automated interpretability by approximating / replacing NN components (here attention heads) with programs.

Aseel Swaid @aseelswaid9 2026-06-28 17:23Z

Keith Richards by Michael Cooper, 1969

Zac Townsend @ztownsend 2026-06-29 16:37Z

If people get healthier and live longer, the cleanest way to capture that isn't backing any single therapy. It's underwriting the longevity itself. A life insurer ( @meanwhile ) that bend the survival curve and wins twice: fewer claims, longer compounding runway.

Dylan Patel @dylan522p 2026-06-27 06:17Z

It literally doesn't matter? CXMT cannot satisfy China demand this decade. So there's 0nreason besides the small price arbitrage. It's like bannin Iranian oil. India and China buy it anyways, but price diff is less

rohan anil @_arohan_ 2026-06-25 17:33Z

I am quite serious, for those thinking I am joking. Mid training is just too late to learn good features particularly worse when networks have been deep fried.

The Kobeissi Letter @KobeissiLetter 2026-06-28 00:31Z

The US housing market is now firmly in favor of buyers: 46% of home sellers offered concessions to buyers in May, the highest May reading on record. This percentage has DOUBLED since 2022 and also marks the 2nd consecutive annual increase. This comes as the US now has ~47%

shira @shiraeis 2026-06-29 17:25Z

certain twitter group chats have taught me some people’s brains are running GPT-2-level next token prediction at high temperature, not agentic long term planning

Alan Rockefeller @AlanRockefeller 2026-06-29 17:20Z

Mycena leana found a few minutes ago in Altoona, Pennsylvania.

Ai2 @allen_ai 2026-06-29 18:12Z

AI image generators don't "draw"—they follow a compass: the score function, which points toward more probable images. The same compass drives Bayesian sampling and plasma physics. We built DiScoFormer to estimate the score far better when data gets complex.

CoinDesk @CoinDesk 2026-06-29 16:14Z

LATEST: The White House meets with law enforcement groups to resolve CLARITY Act objections, with Senate Majority Leader Thune suggesting he may bring the bill to a floor vote in July whether Democrats are ready or not.

Seongheon Park ICML 2026 @seongheon_96 2026-06-29 13:14Z

Excited to share our ACL 2026 paper: VAUQ: Vision-Aware Uncertainty Quantification for LVLM Self-Evaluation LVLMs can be confidently wrong when relying on language priors instead of the image. VAUQ checks whether model confidence is truly grounded in visual evidence

Framework @FrameworkPuter 2026-06-28 17:31Z

There are levels to clustering….

Vidit Gujrathi @viditchess 2026-06-29 16:27Z

Chess engines tell you the best move. But grandmasters are human, they don’t always play it. So I built "Kibitz": a human move predictor for chess broadcasts. I trained this model on my Nvidia RTX 5080. Then I made it run as a business by itself. A channel buys the overlay,

POLITICOEurope @POLITICOEurope 2026-06-29 07:41Z

EXCLUSIVE: French President Emmanuel Macron has instructed his administration to find new EU-wide taxes to finance the bloc’s next €2 trillion budget, according to five officials familiar with the discussions.

p3shoemaker @p3shoemaker 2026-06-29 17:12Z

"ai-powered software" is like saying "electric computer" the descriptor might stick for a few things (electric toothbrush, electric car) but majority of tech will treat it as given and not even mention it

Jonah @jonah_b 2026-06-29 17:09Z

This is one of the biggest benefits of being a financial app. You can benefit from others' R&D and monetize others' products. Hyperliquid is happy because Phantom is bringing them more users, and Phantom is happy because they can generate revenue without having to build a perps

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

Nick Dobos @NickADobos 2026-06-29 19:31Z

Cursor on iOS first impressions: I’ve been in the TestFlight for a while now, very excited for remote control launch to actually make cursor mobile useable! - Cloud (via Linux) is entirely useless to me as a Mac and iOS focused dev. Thx Apple. So this limits the app greatly.

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

Lazarz @Laz4rz 2026-06-22 11:10Z

3 month long interviewing marathon for a person at the end of their PhD (!), from a known uni in the us (!), with a track record (!), network (!) and generally very solid (judging by the blog and notes) Maybe we sometimes really don’t appreciate where we are and high bars are

Dev @DevvMandal 2026-06-29 14:51Z

Tomorrow, we’re launching the largest open-source dataset of gaming computer-use data. Gameplay screen recordings + keystrokes/mouse movements. Stay tuned :)

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

Fabrizio Romano @FabrizioRomano 2026-06-28 22:03Z

Yan Diomandé has verbally agreed to sign a five year deal at Paris Saint-Germain. Salary and details agreed as @FabriceHawkins reports, he wants PSG. Negotiations starting between clubs in a separate deal than Maghnes Akliouche, both are ongoing.

Memory Card @MemoryCardFiles 2026-06-28 18:30Z

Game Boy Color Artist: @Meillex

Bohan Hou @bohanhou1998 2026-06-29 18:15Z

It's cool to learn how different warp/warpgroup roles overlap in the kernel, such as FA4. https:// tvm.apache.org/docs/tirx/nati ve_basics/cuda/profiling.html …

John Arnold @johnarnold 2026-06-28 15:31Z

Modern-day finance: You can get leverage via margin to buy a 2x leveraged ETF on MSTR, which is itself a leveraged bet on BTC, partly via a novel security whose cost of capital rises as its price falls, increasing leverage further as it declines. Works great in a bull market.

Zephyr @zephyr_z9 2026-06-29 08:08Z

Good luck getting GPUs and memory for on prem deployment

Nicholas Charriere @nichochar 2026-06-29 04:15Z

I hear it everywhere. Speeches, emails, tweets, …

Viv @Vtrivedy10 2026-06-29 16:52Z

auto-research at scale, agent-scale data mining, voting via mixture of agents doing these effectively means your Harness should support dispatching & orchestrating many subagents on massive scale data/tasks “read all 100k traces & experiment logs. find ways to reduce our token

Johnny Yu @iamjohnnyyu 2026-06-29 16:53Z

Hypotheses are cheap. Verification requires ground truth, real biological data, not papers that don't reproduce. At @TahoeBio we're building that ground truth: an ocean of single-cell data, ML models that predict zero-shot, and an autonomous agent that validates discoveries in

Cam @polarcc 2026-06-29 14:40Z

Sneak peak of a photoshop plug-in I have been working on, halftone and print effects natively integrated, click of a button

rob cheung @perceptnet 2026-06-29 14:21Z

i've never heard a human say belt-and-suspenders but AIs love this phrase apparently

Theodor Marcu @theodormarcu 2026-06-29 16:20Z

@walden_yan and the team cooked with this one! While Devin has always done routing, Devin Fusion is a brand new architecture that delivers SOTA engineering task performance at 35% lower cost

Ernesto Lopez @ErnestoSOFTWARE 2026-06-29 19:06Z

$70,000/mrr app done $100,000/mrr app next My biggest advice, Don’t listen to 99% of the Slop you hear on X It’s mostly all misleading Just to name a few: > shadow bans don’t exist, your content just sucks > 100 views average is not shadow banned > Ads do work

Atoms Not Bits @AtomsNotBits 2026-06-29 16:39Z

BREAKING: @DiscipulusVent , the El Segundo hard tech accelerator, has raised a $30M venture fund.

LINEAR @linear_magazine 2026-06-28 14:31Z

Villa Svartgarn Martin Lyden, Daniel Widman Stockholm, Sweden Photos by Erik Lefvander

Or Perel @OrPerel 2026-06-29 14:34Z

Excited to introduce TRON, a relighting framework for 3D captures. TRON pairs a neural renderer with 3D Gaussian reconstructions, achieving realistic quality, with 3D, material, & lighting control at interactive frame rates. https:// arxiv.org/abs/2606.11314 https:// research.nvidia.com/labs/sil/proje cts/tron/ …

Richard Ngo @RichardMCNgo 2026-06-29 05:56Z

I considered “trusting” as a replacement, but that’s only a good thing insofar as the trust is well-placed, and there are many ways modern westerners shouldn’t trust the society around them. Whereas courage seems much more broadly desirable.

Aseel Swaid @aseelswaid9 2026-06-28 18:43Z

Blaze and basketball. Spanish Harlem, New York, 1975

Paras Stefanopoulos @stefanopopoulos 2026-06-29 19:13Z

All pricing in $ per million tokens. Take opus pricing: $5 in, $25 out, $0.50 cache Lets take a cheaper model, at 20% the price: $1 in, $5 out, ~$0 cache

sk @compchemm 2026-06-29 06:47Z

Protenix-v2 weights available now apples to apples comparison from Lucas Nivon post. Protenix ahead of all; AF3, Boltz, OF3-p2 on a blind test set of protein/ligand. So for Co-Folding use minimum Protenix-v2 https:// huggingface.co/TMF001/proteni x-v2-weights/tree/main …

Jack @tracewoodgrains 2026-06-28 21:03Z

someone sent me a Guardian article about this and asked “where is your principled concern for free speech?” so I started looking into the trial documents and pretty quickly concluded that a whole lot beyond speech was going on for all defendants excessive sentences, though