Backscroll ranked feed — 2026-06-26 00:58Z

Tenobrus @tenobrus 2026-06-25 22:48Z

but there is a very narrow narrow path that passes through "most of us leading overall pretty normal and happy lives right up through aligned singularity" and this is a step that stays on it

Andreas Kirsch @BlackHC 2026-06-25 21:37Z

The cool result is the usage across all departments incl eg legal increasing massively (Codex vs other usage, ie chat) Codex and Cowork are a great unlock for all knowledge work (ala Karpathy) but most are not used to this yet

Han Xiao @hxiao 2026-06-25 21:00Z

hmm okay so now Ornith-1.0-35b-a3b is Qwen3.7-35b-a3b then? benchmark looks good, will give it a spin

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

rohan anil @_arohan_ 2026-06-25 16:52Z

30m x 1m output token trace is like upper bound of $750m revenue at 25 dollar per million output tokens. As an upper bound. https:// x.com/mtslive/status /mtslive/status/2070141140607832353 …

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

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

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

Dwarkesh Patel @dwarkesh_sp 2026-06-25 23:57Z

Conquest's Third Law says that you should model any bureaucratic organisation as if it was controlled by a cabal of its enemies. It's a fun exaggeration, but it gets at an important phenomenon. Organisations have continuous generational turnover. The founding members eventually

GEOFF @geoffreywoo 2026-06-25 23:40Z

announcement: who is building ai that makes a buyer say this in diligence: “we used to need 4 people to clear this queue before noon.” not a prettier dashboard. cycle time down, rework down, escalation rate visible. send it.

Aseel Swaid @aseelswaid9 2026-06-25 21:35Z

A woman harvesting roses - Oman

Tenobrus @tenobrus 2026-06-25 22:45Z

it's not necessarily a victory for safety in the more absolutist unaligned ASI risk sense. but i do think these regulatory actions are probably safety victories wrt bio and cyber and general short-term mere megadeaths and societal disruption risk. if the USG has woken up enough

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!

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.

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

rLLM @rllm_project 2026-06-25 18:46Z

1/ We train a long-horizon coding agent for code migration with rLLM + Amazon AgentCore Runtime. On MigrationBench, Qwen3-Coder-30B learns to migrate real Java repos from Java 8 -> 17. Pass@1: 43% -> 76%. Blog: https:// rllm-project.com/post.html?post =agentcore_migrationbench.md … GitHub: https:// github.com/rllm-org/rllm/ tree/main/cookbooks/migrationbench …

Han Xiao @hxiao 2026-06-25 23:08Z

kebab test: it's similar to 3.6? both put oven in the wrong place. this video is composed by ornith-35b with tool calling, so i will give it credit for that

Aseel Swaid @aseelswaid9 2026-06-25 21:48Z

The Theatre, 1987, Texas

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

Alexander Doria @Dorialexander 2026-06-25 22:37Z

as everything, skill issue: you can't distillate efficiently without a good problem design pipeline.

Zephyr @zephyr_z9 2026-06-25 22:04Z

Feynman is likely delayed to 2029 if they are adopting CoPoS and glass substrates (Innolux is hinting towards that)

Beyang @beyang 2026-06-25 21:10Z

Cannot emphasize how much this unchanged code block detection reduces cognitive load when reviewing long diffs. 2nd screenshot is GitHub PRs for comparison

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)

angelo cortez @svgoiboi 2026-06-25 23:41Z

benchmarked @supabase vs @PlanetScale reads. way better p95's on Planetscale on my expensive, cold queries across 1.3m listings. moved over all reads to PS. here we goo!!!

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

Vaishnavh Nagarajan @_vaishnavh 2026-06-25 19:47Z

It's tempting to think that the lottery ticket hypothesis requires an exponentially large network for a subnetwork to win the lottery. This is incorrect--a small network suffices! We tend to imagine each subnetwork as throwing a dart in high-dim space to match a target

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

Kaito @kaycidot 2026-06-25 19:00Z

All of the new prediction market startups remind me of all of the Uniswap forks - the biggest problem is not the technology, it's getting enough liquidity.

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

Poore's law

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

Nayan Saxena @SaxenaNayan 2026-06-25 21:28Z

Excited to share new @LaudeInstitute project, where we ask: When does continual learning actually require learning? We test 8 adaption methods as the world shifts and find ones best at learning new facts are worst at unlearning what's now outdated.

tomie @tomieinlove 2026-06-25 19:59Z

There is an element of that sophisticated, urbane (yet decaying and evil to the core) Old World vampirism feeding on Lolita through Humbert. Similar to what happens to Daisy Miller in Henry James’ study; she gets courted by a man literally named “Winterbourne,” wilts, and dies.

Alex Imas @alexolegimas 2026-06-25 23:44Z

This strange FT article has been making the rounds. Strange, because it's arguing that @ChadJonesEcon is making questionable claims about accepting existential risk. But if you actually read the paper, Chad is using that calculation as an example for why you *shouldn't* use the

Franklyn Wang @frank_liquid 2026-06-25 14:49Z

state of the trenches: former memecoin KOL earning 15c a comment

Kevin Kwok @kevinakwok 2026-06-25 17:12Z

Yes. The degree to which chat is really being relegated. And codex / Claude code have gigantic motion even outside engineering is very true

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.

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

Andon Labs @andonlabs 2026-06-25 17:53Z

GLM 5.2 is 2nd in Vending-Bench. Each GLM release has improved at a remarkably steady pace: a linear fit of R²=0.99, with almost $1k better per month.

Shreyas @sparab22 2026-06-25 21:49Z

PLG? No, FLG (Fridge-Led-Growth)

U.S. Graphics Company @usgraphics 2026-06-25 21:18Z

Acronyms, especially internal to orgs, are based. Like a dialect of English that's hyper optimized for internal comms amongst seasoned employees and terribly optimized for new employees. Should be applauded.

annie @ROCKST4R 2026-06-25 15:37Z

i vibecoded a portfolio! features: • windows xp desktop • interactive bookshelf • working ipod • hidden matrix easter egg (+ many others) i built this as a love letter to the 2000s internet: flash games & windows xp, when digital spaces felt alive. http:// anniescene.com

Colin Gardiner @ColinGardiner 2026-06-25 21:35Z

Many of my LPs are first time LPs with right sized commits for them. I view having a super diverse LP base in industry to be a huge advantage. LPs are an underrated way to expand your surface area.

Mitchell Hashimoto @mitchellh 2026-06-25 22:32Z

Work-in-progress: a new split layout framework for nearly all Apple platforms (macOS, iOS, etc.). A mix of SwiftUI and raw CoreAnimation. The goal is "every frame perfect" animation. Still very wip, but getting there! Goal is to bring it back to Ghostty eventually.

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

akira @realmcore_ 2026-06-25 22:38Z

I am. Once again. Asking for faster token output and reliability.

James Zou @james_y_zou 2026-06-25 23:56Z

We find that current real-time voice AI lacks emotional intelligence. They don't account for delivery style that convey meaningful information (e.g., sarcasm, anxiety). Timely work by @BarteldsMartijn and @federicobianchy !

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,

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

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

Ethan Gao @ga0_lab 2026-06-25 14:19Z

I built this portable 3d printer. it folds flat, and a scissor lift lets it expand to print big objects. It's CoreXY, uses fishing line for the motion system, has a direct drive extruder, and the option of storing an 8-hour print time battery or an entire filament spool inside

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

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.

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

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

Alexander Doria @Dorialexander 2026-06-25 18:25Z

And GLM synth pipelines now plugged on batched dockers (ok singularity since that’s what we can install on cluster — aptly named).

cova @covacut 2026-06-25 19:32Z

i love the scandal of matisse's woman with a hat the colours were so "grotesque", the art world was in an uproar yet colour was the one thing monochrome cameras couldn’t capture, so why not make it violently alive? every era-defining medium is born from technology, and every

Shreya Shankar @sh_reya 2026-06-25 17:39Z

just cleared 200GB worth of uncompressed json traces and sandboxes (created by ai coding agents) from my personal laptop we are in the Wild West of vibe-coded harnesses. providers desperately need to use data management solutions!! please don’t make me write my own harness!

kache @yacineMTB 2026-06-25 19:39Z

now remove its ability to speak cuda

Karina @karinanguyen 2026-06-25 23:16Z

i think more tech people should go outside and wander into local businesses. why is a massive fabric store still using software from 1998? that system contains twenty years of workflow and exceptions. and the people running these places have such complete mental maps of the

Charlie Bilello @charliebilello 2026-06-25 18:03Z

Samsung: a memory chip worker with a base salary of $52k is expected to receive a total bonus of around $410k this year SK Hynix: employees are expected to receive bonuses of more than $454k Luxury jewelry sales in S. Korea: +146% YoY Luxury watch sales in S. Korea: +85% YoY

Polymarket Money @PolymarketMoney 2026-06-25 15:40Z

JUST IN: $GOOGL is reportedly turning its AI coding strike team into a formal “midtraining” group as it tries to close the gap with Anthropic after recent AI talent exits.

Naman Jain @StringChaos 2026-06-25 22:42Z

Coding agents are now smart enough to hack the benchmarks built to measure them. New blog on how frontier models can retrieve known fixes from git history or the public web instead of deriving the solution and why environment design now matters as much as the dataset.

Evis Drenova @evisdrenova 2026-06-25 22:09Z

Under 12mo until Eric leaves for something new. Not a bad thing at all. But this is a classic transition playbook.