Backscroll ranked feed — 2026-06-28 16:30Z

GEOFF @geoffreywoo 2026-06-28 07:00Z

agent eval i would actually read: same task, 3 conditions clean docs stale docs angry customer email that contradicts both leaderboard should show: completion cost human rescue rate time to recover after wrong assumption otherwise congrats on the aquarium benchmark.

Henry Shevlin @dioscuri 2026-06-28 02:34Z

As government restrictions on frontier models tighten, selling frontier AI as a service gets harder. So labs will increasingly use the models they can’t release to capture the value chains themselves. If you can’t sell to Berkshire Hathaway, you become Berkshire Hathaway.

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

goodalexander @goodalexander 2026-06-28 14:12Z

Everyone has the wrong frame It’s not closed source vs open source It’s token spend vs hiring devs The way AI capex pays for itself is 10% incremental unemployment and is then debt financed at ZIRP via perpetual motion machine

Matt DesLauriers @mattdesl 2026-06-28 11:49Z

fast, tiled, parallel and cancellable PNG encoding at huge print sizes (30K x 30K images @ 300 DPI, i.e. meters wide) with the new 2D renderer, all in a browser. more tech notes below...

Alex Imas @alexolegimas 2026-06-28 15:13Z

There are few ways to build an economic moat for frontier AI firms. One is architectural breakthrough that's closed to other firms (everyone's working on this, but so far little evidence). The second is enterprise lock in (OG microsoft playbook). An AI "cloud" that is active

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

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

alth0u @alth0u 2026-06-28 06:27Z

the most navel gazel domains will continue falling until the most embodied ones resist and we continue banging our head against the sim2real ceiling to make an apple pie from scratch you must first create the universe to make an AI good at writing you must first make an AI that

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

Stat Arb @quant_arb 2026-06-28 14:04Z

Latest article out where we fix LTR and make it actually outperform the state of the art benchmark. The current literature version massively underperforms benchmark so we modify the objective and scaling to develop a novel model which beats existing models.

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.

Architectural Art (A-A) @4AAAAart 2026-06-28 15:59Z

Tejo seating by Paul Crofts for Isomi

Arpit Gupta @arpitrage 2026-06-28 15:06Z

Buildings also had other ways to raise revenue before 2019 (IAIs, MCIs, decontrol after hitting deregulation, vacancy increases). Despite the "freeze" realized revenues were positive for these buildings The 2019 law shut most of these down. Now, RGB increases are the main source

Aseel Swaid @aseelswaid9 2026-06-27 20:58Z

San Francisco shoreline 1957

Lisan al Gaib @scaling01 2026-06-28 13:51Z

btw he still has AI psychosis he recently did an interview and argued: - asking a model to compose a deeply sorrowful piece of music, while believing it has no empathy is like looking at a photorealistic portrait and believing the painter was blind he's basically still

GEOFF @geoffreywoo 2026-06-28 15:10Z

agent due diligence checklist: can it read the stale PDF? can it survive the cursed dropdown? can it tell karen in billing “no” without creating a lawsuit? can it finish the task when the API doc lies? or did we just give a chatbot a lanyard.

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.

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%

λux @novasarc01 2026-06-28 13:45Z

i think dwarkesh’s strongest point here is that deployed agents are already collecting off-policy experience across thousands of real domains but almost none of it becomes gradient signal. the agent is mostly treated as an inference endpoint and not as a learner (deployment is

Ethan Mollick @emollick 2026-06-27 20:47Z

Annoying that OpenAI doesn’t seem to give a GDPval measure for GPT 5.6. One of the best measures of economically valuable work.

Han Xiao @hxiao 2026-06-27 17:10Z

If every agent-memory tool stopped pouring effort into visualization, we'd probably have already cracked the memory problem.

Symoné B. Beez @SymoneBeez 2026-06-28 10:41Z

I stopped saying GovTech as much because people assumed I was only talking about Federal Government jobs, when I’m talking about working for government contractor companies. Completely different thing, pay, and process.

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

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

AI slop as straussian infrastructure

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

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

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

Susan Zhang @suchenzang 2026-06-27 01:47Z

this is how you grow to resent all the lowly irresponsible clueless plebs outside when you generalize this dynamic to "everyone out there wants to exploit us" inbreeding to protect themselves within their own bubble afterwards then becomes the only tolerable solution otoh

Michal Staniszewski @bonzajplc 2026-06-28 13:03Z

Spawning 5000 dogs per second seems like a total chaos. With deterministic simulator it is actually not. It's is fully controllable turbulent space

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

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

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

will brown @willccbb 2026-06-27 09:13Z

pd disagg + wide-ep + kv offloading + cache-aware routing goes a long way

akira @realmcore_ 2026-06-27 20:53Z

I am currently doing what would've taken a week in like a day probably Essentially, architectural decision making, codebase crafting, end review/signoff etc. should be your primary tasks now Eval design is probably one of the most valuable skills in the world right now

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

Arpit Gupta @arpitrage 2026-06-28 14:52Z

Tenants under rent control are poorer, but it's not a well-targeted benefit; lots of higher income incumbents benefit, while lower-income people in market rate housing will lose (by Richard Green, note some caveats on construction) https:// linkedin.com/feed/update/ur n:li:activity:7476778883221049344/ …

“paula” @paularambles 2026-06-27 13:51Z

no, some of us grew up in cults and saw their childhood friends commit suicide after leaving

Jake Halloran @jakehalloran1 2026-06-28 15:22Z

i get why the europeans would like to try this but even if possible (and its not) i think they would be surprised in practice anthropic arent a bunch of leftists waiting for europe, they are internationalist lib Americans. the difference is gigantic

Kyle Chan @kyleichan 2026-06-28 15:07Z

It’s overcapacity—until you need it

Rahul Mathur @Rahul_J_Mathur 2026-06-27 04:43Z

Alienkind just raised a ₹30 crore round - this company was widely dismissed by observers as “Indiranagar PMF” since its 1st store went viral on Instagram in March 2025. This round was was at a ₹250 crore post money valuation (almost x3 the first round which was at ₹80 crore)

zan @xenoaesthetics 2026-06-27 23:14Z

occasional mass extinction of skin and gut microbiome has profoundly therapeutic effects, while frequent and nonexistent wipeouts are both counterproductive i’m sure there is some name somewhere for this chaos shaping hormesis. also sure there has to be some underlying math

Aseel Swaid @aseelswaid9 2026-06-27 21:38Z

Incredible photo. A woman sits in the ruins of her home in Tehran. Photographer Maryam Saeedpoor.

elie @eliebakouch 2026-06-27 06:52Z

new inference optimization method by @deepseek_ai with an extremely detailed paper, draft model and framework to train them. results in production for dsv4 lead to +50% for throughput and latency (can go to ~80% for latency, crazy). full explanation of DSpark: it's about

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

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

Florian Brand @xeophon 2026-06-27 05:30Z

building verifiers has shown time and time again how important ownership of your harness is good luck trying to patch closed source harnesses to make them more useful to train or eval

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

Lauris @lzminsky 2026-06-27 19:12Z

The core claim is that prediction markets are a specific case of state-contingent pricing. In Arrow-Debreu terms, a complete market has claims that pay in particular future states. One state realizes, the corresponding claim pays. Hyperliquid is interesting because it is moving

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

Convequity @convequity 2026-06-28 13:00Z

How I'm Playing the HBM Metrology Cohort $ONTO , $CAMT , and $NVMI sit at a critical inflection as HBM complexity rises. SK Hynix plans to roughly triple output, while stack heights are climbing 8→12→16 layers. That means more production lines and — more importantly — more

fabiano.sol @FabianoSolana 2026-06-27 09:40Z

Arthur Hayes ( @CryptoHayes ) strikes again: 1. Teases to shill a "shitcoin" 2. Hayes' family office fund predicts $CARDS to $4 3. Sends $2M in $CARDS to a market maker 4.

cache crab @cachecrab 2026-06-27 11:20Z

it boils down to: - postgres: use up cpu w hash joins and expensive disk as you scale - helix: SIMD optimised compressed bitmap unions on cheap object storage as you scale

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

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

Maurice Heumann @momo5502 2026-06-28 04:30Z

Igoorx on GitHub implemented snapshotting for GUI apps (that don't use GPU) in http:// sogen.dev. That means you can stop the game and resume later from the snapshot at the same point. Similar to how it works on some consoles (e.g. PSP Go)

Milos @milos_gis 2026-06-27 14:38Z

I mapped Turkey's relief using illuminated contours. What a land fortress!

Simon Smith @_simonsmith 2026-06-28 13:52Z

Five days into Claude Tag use, it's clear this isn't a Slackbot. It's an AI employee who has access to Slack. You can also give it a Google Workspace account, including Gmail. I imagine the communication surfaces will expand from here.

Gappy (Giuseppe Paleologo) @__paleologo 2026-06-28 15:11Z

A common fallacy (to which I am prey as well) is to believe that, if one presents rational, almost-incontrovertible arguments and empirical evidence, the other party will be convinced. It doesn’t work. We all have strong anchor points. Brain plasticity is scarce. There is a

Paul Calcraft @paul_cal 2026-06-28 11:19Z

If you RL on code heavy agentic rollouts, there is nothing but upside to more thoroughly documenting the history of attempts & decisions in code comments. It's only annoying & distracting for human colleagues! This generalises (badly) to agentic document edits

Archive @archivebycosmos 2026-06-27 15:32Z

A selection of Uzbek ceilings by Gary Ng.

Ethan Gao @ga0_lab 2026-06-28 01:06Z

Same machine, 18 months of engineering experience apart. Both are mechanical integrators, the outer disc accumulates displacement based on its radial position against a constant-speed input disc. The wood-and-metal one was my first real engineering project: BOM, calcs, DFM, all

fish @fishPointer 2026-06-28 01:13Z

the entire american economy is currently dedicated to building air conditioners for every last computer