Backscroll ranked feed — 2026-06-25 21:41Z

Andrew Curran @AndrewCurran_ 2026-06-25 20:26Z

SynthID goes in everything it seems. It's actually inside Gemini's text as well. Gemini uses a secret rulebook to slightly affect the percentages of equally good word options when generating the next probable word. This creates an invisible pattern within the text.

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

sam @samgoodwin89 2026-06-25 20:40Z

This was a large refactor driven by my experiments at integrating coding agents and building my own harness in Alchemy. It used to be hard to write a Layer like this that included resources+bindings and slot it into a Durable Object. Now, it's dead simple.

Architectural Art (A-A) @4AAAAart 2026-06-25 17:54Z

Tribeca Triplex by Jae Joo New York, New York, USA

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!

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

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

Ara Kharazian @arakharazian 2026-06-25 18:28Z

This is the chart I would show to argue against AI bubble concerns. Yes AI is the fastest growing spend category ever. It's also still, even for the top quartile of firms, fairly small...<1% of total spend per Ramp data. In @azeem 's great report on the state of the AI economy.

Lucas Beyer (bl16) @giffmana 2026-06-25 19:33Z

That's pretty cool! They decompose the weight matrices into interpretable subsets (that part needs a lot more than 4 tokens iiuc) and then basically tune down the subset for German, which destroys German performance while keeping almost everything else intact (unlike lora)

Brendan Falk @BrendanFalk 2026-06-25 20:50Z

Today we are excited to launch Branches in @useHercules Branches allow you to create unlimited isolated snapshots of your app to develop on. Each Branch gets its own 1) git branch, 2) database clone, 3) backend environment, and 4) development machine. Branches makes it

Ronith @ronithhh 2026-06-25 17:55Z

What’s crazy is that buoy can do all of these things… without a human! It can just directly use DoorDash, uber, and whatever site needed to book dinner reservations. One of the big perks of building outside imessage is we have browser use that actually works

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

Level 10?

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.

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.

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

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,

Jerry Neumann @ganeumann 2026-06-25 20:55Z

I've had founders do this, and even seemingly disappear from the face of the planet. To the point where I stop caring about the investment and start asking potential mutuals about their well-being https:// x.com/maiab/status/2 /maiab/status/2069891233904988222 …

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

Dylan Patel @dylan522p 2026-06-25 20:10Z

People don't quite understand how many behind the meter power generation assets are being built for datacenters because the US Grid sucks, despite the higher cost and complexity.

Matt Rickard @mattrickard 2026-06-25 02:34Z

Doing most of my code review locally with /review now. PR review bots feel wrong Have to pull the changes down and run them through the harness anyways. Suggestion is pre-CI. Too much noise, not enough signal, GitHub's fault too Problem is all the context gets lost for others

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."

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

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,

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

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

Simone Canc @simonecanciello 2026-06-25 17:17Z

whoever built this app is a genius. an app where your guests record your event and create a shared memory album. $20k/month, 7.5M views. people spending $30k on a wedding won’t think twice about another $10. tldr: don’t build for broke people.

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

what even is midtraining?

George @georgecurtiss 2026-06-25 19:24Z

AI memory will be solved by the models getting better.

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

Baobab Waterfall by Ahmad Eghtesad

Theo Jaffee @theojaffee 2026-06-24 15:59Z

Incredibly, San Francisco's local politics are now distinctly less radical/socialist than New York City's

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.

Enze Xie @xieenze_jr 2026-06-25 02:31Z

Sol Video Inference Engine is here! An agent-native, training-free full-stack accelerator for video diffusion. It auto-tunes cache + sparse attn + token pruning + quant + kernel fusion for any model/hardware/config. >2× end-to-end speedup on 64B Cosmos3-Super, 22B LTX-2.3

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,

marley @_marleyx 2026-06-25 19:02Z

I could never convey the feeling of this discovery: it’s a random Tuesday and, all of a sudden, we saw a branching, treelike structure. Moreover, it’s my own brain. A brain was looking at itself, while also knowing that it was the first time anyone in the world had seen this

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

Tanay Jaipuria @tanayj 2026-06-24 02:47Z

I was at the game. Senegal fans had a section (albeit smaller numbers) and were loud throughout the night

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

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

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.

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

Peter Sterne @petersterne 2026-06-24 20:44Z

I don't think young progressives even conceive of the idea that members of Congress are supposed to "bring money back" to their districts. They think in terms of national policy, like providing universal healthcare or childcare, not winning funds for specific pet projects.

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

Cel Shaded Assembly Viewer

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

Hanming Fang @HanmingF 2026-06-25 00:39Z

High-speed rail may have helped fuel China’s EV boom by easing range anxiety and spurring charging rollout while amplifying its impact. HSR may account for up to one-third of China’s EV market-share growth, highlighting complementarities between intercity transit and EV adoption.

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

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

Jenia Jitsev @JJitsev 2026-06-24 23:09Z

There is a lot of room to push open-source models. Eg using Qwen3-32B base with our composed data, we score on Terminal Bench 2.0 25.60% using only 100k samples and no RL, SFT only, outperforming Nemotron-3-Nano and coming close to Nemotron-3-Super that uses way more compute.

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. . .

Architizer @Architizer 2026-06-24 21:01Z

The Living Room at Parque del Espolón reimagines the typology of the kiosk as an open, traversable, and yet intimate space ─ a site of urban domesticity ─ rather than a stage raised from the ground. Project by: Palmyra PLLC Details: https:// hubs.la/Q04mqplv0

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

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

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

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"

th0rgal @Th0rgal_ 2026-06-24 07:59Z

Just reverse engineered my Oura Ring 5 so I can control my computer like a wizzard. @ouraring please send my love to whoever buried a feature to stream live accelerometer data

Ted Nyman @tnm 2026-06-25 05:54Z

the fact git has lasted this long (20+ years) is bonkers. statistically, something you wouldn’t expect to see in your lifetime. the messi of software. a testament to good engineering practices. an extremely careful set of maintainers & most of all: the C programming language.

YCS Yang @YCS_Yang 2026-06-24 16:54Z

Hard drive suffering from widespread pad detachment; repaired via manual jumper wiring.

3X Long Labubu @labubu_trader 2026-06-24 21:00Z

Thanks to the Chinese government, $MU’s ER was stable. Chinese HFs used to be the best “sell the news” players to shake out retails and amplify volatility. Now they have to hold and “can’t sell high buy low,” as selling would mean losing the leverage to regain exposure.

witcheer @witcheer 2026-06-23 11:18Z

obviously I tested it! measured on a 5090, Q6_K, think-off, same harness as the base and the prior version: - quality flat: q_avg 93.7 (base 94.0, prior 94.1): weights didn't move - MTP draft head recovered: 1.9 to 2.3x, back on the base curve, ~144 tok/s peak - agentic score

derek guy @dieworkwear 2026-06-25 00:36Z

If you look at photos of how actual American students dressed in the 20th century, you will find a much more eccentric and diverse expression of classic American collegiate style than what's popularly presented through Ralph Lauren ads (which are then recycled through Twitter).

Kai Cabero @kai_cabero 2026-06-24 09:00Z

I know a guy who pretends to be a venture capitalist to get his competitors to voluntarily hand over their entire client list, revenue numbers, and churn rate... He is not a VC. He has never invested in anything. He drives a 2019 Civic But he registered an LLC called "[City

Sidhant Sibal @sidhant 2026-06-24 10:41Z

Passport is a travel document, not document of citizenship, say MEA officials