Backscroll ranked feed — 2026-06-25 22:47Z

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

xjdr @_xjdr 2026-06-25 21:34Z

around feb / starting with gpt 5.2, model capability stopped being my main constraint (which is why more of _my_ cycles are spent on working with these models effectively these days vs model capability). there are still lots of things that frontier models cant do, but they are

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

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.

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

dalibali @dalibali2 2026-06-25 18:10Z

Funny anecdote from a friend: His company ended 2025 with less than 30M annualized AI spend Took out 200M of dev headcount spend Encourages use of codex and Claude By end of May spend balloons to 500M annualized. A single person spent 10M and got fired when they realized he

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.

Kevin Kwok @kevinakwok 2026-06-25 16:59Z

Ramp is introducing an interesting thought of if trust was high enough. How many people could you load into co-ceo role and make room for more exec suite

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!

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

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

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

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

Ivan Burazin @ivanburazin 2026-06-25 20:09Z

If you're optimizing for raw inference throughput, you're betting on the wrong curve. Look at how humans think. There's inference in our heads, but most of what we actually do is fire deterministic tools. A search query, a route home, a card swipe at checkout, etc. Each one of

wavefnx @wavefnx 2026-06-25 17:28Z

Just remembered I got a video editor in a branch where you can use GLSL and utilizes my MP4 Parser (moov) The seeking part being so fast and with such little memory used is incredible and result of my lib (moov) For now, the MP4-Parser will be used to run on embedded/STM32s,

Works in Progress @WorksInProgMag 2026-06-25 17:28Z

"Every major American city has a land shortage. But we have more than enough shallow water, dredging capacity, and market incentive to make more land, just like we did 150 years ago. The only obstacle is our own choice to make making land illegal." Read the full story in Issue

elie @eliebakouch 2026-06-25 19:35Z

very interesting report by openai on the internal and external use of codex also very interesting that research team token consumption growth is by far the highest, and that oai uses skills and concurrent agents much more compared to other orgs

Arlan @arlanoska 2026-06-25 16:00Z

rebuilt the figma "vector editor" for the web draggable points and cool morphing shapes code: http:// arlan.me/vault

marley @_marleyx 2026-06-25 21:40Z

Today we’re sharing the most detailed scan ever taken of a living human brain. None of us were ultrasound scientists before this. We worked backwards from a desire for brain interfaces and taught ourselves physics, ultrasound, electromagnetism. This is a story four years in the

Aidan Clark @_aidan_clark_ 2026-06-25 18:14Z

what even is midtraining?

Evil Rabbit @evilrabbit_ 2026-06-25 17:45Z

Working on Lifeline. A timeline for your personal, work, and social life. Add links, alternate timelines, and project your future. Fast. Super lean. Just a few KB. Planning to open source it.

Reka @RekaAILabs 2026-06-25 20:39Z

Training world models needs egocentric video and dense action signals, synchronized. That data is genuinely hard to find. We built it from Counter-Strike 2 demos. CS2-10k: 600K+ player-round videos, 10K+ hours, per-frame annotations — keyboard state, mouse delta, 3D position,

swyx @aiDotEngineer @swyx 2026-06-25 19:48Z

on their @latentspacepod we covered how @pimdewitte accidentally made the PERFECT world model data collection business by collecting the world's largest dataset of trainable (video,action) pairs. turning the attention economy into the attention industry. congrats Pim!! link

Harvey @harvey 2026-06-25 17:03Z

Inference-time model routing based on legal practice area dramatically improves agent performance. So do other types of "blended intelligence": - collaborative model teams, - advisor-executor patterns, - and model routing based on inferred user preference. While these methods

nikhil @uninsightful 2026-06-25 18:14Z

over and over in technology cycles we’ve seen vertical integration win early in a market and then modularization take hold we believe we are potentially seeing the beginning of that turning point in AI — and with the agentic stack in particular such an exciting moment. more

Yash Bhalgat @ysbhalgat 2026-06-25 08:47Z

Level 10?

Guillermo Rauch @rauchg 2026-06-25 20:23Z

How we imbue coding agents with our design standards

Tsung Xu @tsungxu 2026-06-25 02:34Z

Our development process will save hundreds of thousands of dollars in proprietary but manual aerospace tools, said an engineer who has worked at one of these companies. Huge and widening variance in what toolchains companies are using right now

Alex Imas @alexolegimas 2026-06-25 18:29Z

Not quite, what you get is market segmentation: commodity segment and quality/premium segment. You see this in essentially every market that has any degree of automation: food, clothing, creative output. Here you'll likely see the same thing: most writing will be AI or

gabriel @gabriel1 2026-06-25 20:50Z

AI is so bad at business decisions like - who should we hire - what product should we stock - what's the biggest bottleneck probably because there is close to zero long trajectory data of decisions being made and their outcomes. maybe that's agi

Vatsal Baherwani @vatsalbaherwani 2026-06-25 16:09Z

Scaling laws predict an LLM's pretraining loss, but not its capabilities. Abilities like in-context learning emerge abruptly and only past a certain scale. Our new paper traces this to one bottleneck: learning which tokens attention should focus on. https:// arxiv.org/abs/2606.25010

Tobias Katsch @TobiasKatsch 2026-06-25 18:18Z

shovel > gold

Evis Drenova @evisdrenova 2026-06-25 20:48Z

Amazing what jemalloc, fixing a memory leak and removing a recursive function from the hot path can do for your infra

Jay @jayendra_ram 2026-06-25 16:12Z

I wonder if OpenAI hiding thinking caused it to not be the de facto choice for distillation, which in turn caused Anthropic’s revenue to explode due to API demand.

Houda Nait El Barj @Houda_nait 2026-06-25 18:03Z

AI can interpret my grief. It cannot grieve. It can describe a wine, but it has never tasted that wine. I wrote for @NoemaMag about why, as intelligence gets cheap, the thing we will crave from one another is raw, unoptimized, shared reality. The stuff machines can describe,

Agu Seguí @agusegui 2026-06-25 19:43Z

tonight we launch the early beta of the Paper Mono type specimen only 80 copies exist... think you're lucky enough to grab one? P.S. rainbow not included... @paper HQ just caught some colors

Gabriella Garcia @reallygabriella 2026-06-25 19:55Z

in an age of AI, we wanted to do something real. so we attempted to recreate these famous NASA images - built a 20ft x 10ft chalkboard wall, filmed our launch video stopmotion style. one all-nighter. 6 hours in a williamsburg warehouse. many bruises later. wouldn't have done it

Ronnie Chatterji @RonnieChatterji 2026-06-25 16:21Z

The biggest shift in AI for work may be the move from chat to delegation. In a new paper, we study how Codex usage evolved over the past year - especially within @OpenAI - as agentic tools became more capable. We saw an expansion of users and usage in this time.

Sam D'Amico @sdamico 2026-06-25 21:50Z

hyperstitioning the "electric stack" as a venture category and "american shenzhen" as the goal

Will Quist @wquist 2026-06-25 20:41Z

1H 26 is going to be brutal comp for the labs in 27

Nabeel S. Qureshi @nabeelqu 2026-06-25 15:06Z

It’s too easy to write with LLMs and most people don’t care very much (either as readers or writers); *and* the LLMs will get better. People take the easy route for most things. So the conclusion seems inevitable: over time, the majority of the text we read will be AI-written.

Andrea Miotti @andreamiotti 2026-06-25 15:53Z

Chad Jones, economist who'll be joining Anthropic: "In other words, with log utility it is optimal to take a 1 in 3 chance of ending human existence in exchange for a 2/3 chance of dramatically raising living standards by a factor of 55."

Dramz2(람지)/fevercell @nurik2ami 2026-06-24 10:43Z

Hatsune Miku Racing Queen 2020 >> 2026 I drew some fan art of Racing Miku! Hatsune Miku Racing Queen Fan Art #初音ミク #HatsuneMiku #レーシングミク #RacingMiku #ミクイラスト #MikuFanArt

Eric Ho @ericho_goodfire 2026-06-25 16:30Z

we're getting closer to being able to hand edit a weight and know exactly what it will do. kind of wild that this was done in one day using silico (our ai neuroscientist)

stefan @wasserstein_rao 2026-06-25 04:20Z

Cel Shaded Assembly Viewer

Simon Willison @simonw 2026-06-25 18:28Z

Also not a great advertisement for a sandboxing product: @daytonaio effectively saying they don't trust the security of their product enough to expose the source code

ParametricArchitecture @parametricarch 2026-06-24 21:14Z

Baobab Waterfall by Ahmad Eghtesad

산왕 @mount_king_gg 2026-06-24 09:59Z

서울시 보도자료 중 정비사업지역 전자총회 지원 관련 자료인데요.. 보시면 3년 내 착공 가능할 것으로 보이는 정비사업 지역 list가 나와 있습니다 정비사업 입주권에 관심 있으시다면 가능한 아해 지역을 알아보시기를 바랍니다 [재개발사업] - 36곳 (강북 지역) 한남2, 한남3, 갈현1, 증산5,

Lewis Fausett @lewiswf 2026-06-24 15:41Z

This is why you are poor. $10m / year brand w/ 65% Gross margin @ 2.0 MER Cogs - $3.5m Ads - $5m Klaviyo + Shopify + SMS - $174k Misc Opex - $326k ($27k / month) To have $1m ACCRUAL net income. Which mean's you're likely not paying yourself. . .

Nathan @asplencmnt 2026-06-23 20:24Z

Excited to release ParallelKernelBench (PKB), a benchmark for measuring LLMs’ ability to write fast multi-GPU kernels! Multi-GPU kernel generation compounds several hard problems: - a large parallelism design space - a new communication axis to optimize - and

Hyperliquid @HyperliquidX 2026-06-25 03:53Z

Portfolio margin on Hyperliquid has been live for six months and is now in beta with increased limits. Users with account value <$25M can use BTC and HYPE as collateral to trade perps, spot, and outcome markets with greater capital efficiency. See the Docs for borrow and supply

Tathagat @tathagat_ 2026-06-24 19:46Z

Built a better RAG with Proposition Indexing. Instead of embedding raw chunks, I embed key facts and retrieve full context later. 80% → 90% hit rate 895ms → 814ms latency 100% local stack. GitHub: https:// github.com/tathagat-git/p roposition-rag … #RAG #LLM #AI @rishabh10x @kirat_tw

thebes @voooooogel 2026-06-24 01:49Z

AI statelets, capped in size by a meta-ethical commitment to human exit rights, must propagate themselves by sexual reproduction: a genetic recombination of the constitution of any N statelets forming a new statelet overmind. with environmental fitness to spare, sexual selection

sam @samgoodwin89 2026-06-25 19:05Z

Alchemy's Effect IaC abstractions are starting to compound. I integrated Vercel's AI SDK 7 into Alchemy CF Containers so you can deploy harnesses like OpenCode, Claude Code, Codex. Here's OpenCode in a Cloudflare Containers as an importable Resource Layer.

LanceDB @lancedb 2026-06-24 15:30Z

Updating a label on a row that contains a 50MB video shouldn't require Spark to read that video. Lance Blob V2 in Lance Spark keeps blob columns as descriptors through the plan. Bytes materialize only on write. A blob column stays a column. The SQL stays the same. What changes

Saarth Shah @saarth_ 2026-06-25 17:04Z

We built an AI agent that traces child predators. A 41 year old man has been commenting on teenage girls' posts for 5 years. Our agents traced him from his Instagram handle. This kind of predatory behavior is happening at scale right now across Instagram, Discord, TikTok,

Jibran @Jibran_05 2026-06-23 17:27Z

We've been running @viralbench_ai for the last few days. These models will autonomously get millions of views without any of our guidance. We're open sourcing the entire experiment so everyone can follow along.

Sergio @SergioGarc20223 2026-06-25 20:15Z

Hi, I'm Sergio Garcia. Head of Operations at Corgi. I built Dataroom for Corgi after we got tired of paying premium prices for basic data room functionality while running our own startup. I created this X account specifically for this, but you can also reach me on LinkedIn if

Andrew Solender @AndrewSolender 2026-06-24 03:46Z

NEW: House Dems reeling from NY primary results - One called it an "earthquake" & a "huge defeat" for leadership - "It was a tough night," Rep. Meeks, Queens Dem Party chair, told @axios - Senior House Dem, asked if members are worried: "Yes they are"

Jason Mast @Jasonmmast 2026-06-23 11:43Z

RA Capital, RTW others put $230 million behind a new Chinese-founded competitor to $Beam Therapeutics

Josiah Gulden @jgulden 2026-06-25 17:08Z

let’s talk about this. Figma keeps using this phrase — “code is a material for design”— and if that’s your working mental model, the execution of code layers makes sense. executable code sits inside frames on the canvas, alongside other “materials” like boolean vectors, bitmap