Backscroll ranked feed — 2026-06-26 07:05Z

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

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

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

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

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

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

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 2026-06-25 21:35Z

A woman harvesting roses - Oman

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

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

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

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

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

Jukan @ICML @jukan05 2026-06-26 00:33Z

OpenAI’s ARR stood at $40bn as of June.

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

atlas @creatine_cycle 2026-06-26 01:44Z

this is an arrangement of 50 Cent's "In Da Club" by my pianist michael potter. somehow he turned this certified hood classic into an epic that's in 11 and has metric modulations all over the place. but yeah other than that it sounds exactly like the original

Marc Andreessen @pmarca 2026-06-26 00:15Z

Overheard in Florida: “The West doesn’t exist. It died with the French Revolution. America supplanted 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

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.

Aseel Swaid @aseelswaid9 2026-06-26 01:42Z

A Palestinian woman and an Arabian horse. Captured by Aseel Swaid

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.

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

Poore's law

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

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

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)

Tanay Jaipuria @tanayj 2026-06-25 22:35Z

Algeria can’t risk a loss because of GD

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

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.

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 …

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).

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.

dax @thdxr 2026-06-26 02:25Z

the reason we excluded frontier models from our data page is they are artificially low in usage and didn't want people concluding this most people using them use them directly, they don't buy them through us we wouldn't be surprised if they're 90%+ of total token spend

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)

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

goodalexander @goodalexander 2026-06-25 17:34Z

> get super stoked > 328 tps > wow that's sick > sign up for account > run quick benchmark bc I'm not a trusting person > it's 58 tps the same as everyone else thanks for playing https:// x.com/Yuchenj_UW/sta /Yuchenj_UW/status/2070166719839326396 …

Ekaeo @Ekaeoq 2026-06-25 14:23Z

This is the pinnacle of Chinese watchmaking The "Fam al Hut" Möbius This watch blows Richard Mille out of the water.. a 32000$ biaxial tourbillon, double retrograde, jumping hour, just how freaking insane is that... Makes me really happy that they won the GPHG award, because

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

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 …

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.

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!

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.

Sheel Mohnot @pitdesi 2026-06-26 05:31Z

They bought distressed debt at ~1 cent on the dollar and got headline credit for the sticker price. Obviously great that they spent $5.5M getting people out of debt but all the headlines are about $550M

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

Nicole @elocinationn 2026-06-26 00:17Z

Many a VC today is no better than the average stocktwits trendfollower about to blow themselves up. I recommend a smart financier sell these folks esoteric derivatives on top of their illiquid speculations offering them even more correlated exposure to their desired theme such

Sergey Levine @svlevine 2026-06-26 04:12Z

We can learn a model that provides shaped "process rewards" for robotic RL, that evolves automatically as the policy gets better. This improves performance on benchmarks, and works in the real world! Some fun new work with Raymond Tsao & @ajwagenmaker

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.

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,

Max Lv @m0d8ye 2026-06-26 02:00Z

This is probably the most favorable policy for China, opening the door to international markets for Alibaba's PPU and Huawei's Ascend. China never planned to make money from large models in the first place; their strategy is to dump data center hardware to address the issue of

Eric Lay @itsericlay 2026-06-25 22:10Z

I give up. Here are all the eng. roles we're hiring for: - Founding Design Eng. $150K - $250K - Founding Harness Eng. $130K - $200K - Founding Infra Eng. $160K - $250K The comp is all cash base salary and you get significant equity ($400K+) + we offer relocation and visa

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.

Mixedbread @mixedbreadai 2026-06-26 00:43Z

New: Granular evidence citations in QA Beyond chunk-level citations, our QA now points to the exact evidence snippet supporting each answer. Easier to inspect, verify, and trust. → Available with high-quality processing.

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

derek guy @dieworkwear 2026-06-26 02:08Z

Absolutely: — Focus on fabric weight and weave, as lightweight fabrics made with a porous weave will breathe better — Consider linen shirts, as these will wick moisture from your skin and keep you dry — Advocate for affordable housing in walkable neighborhoods, as this will

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

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

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

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

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

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

will brown @willccbb 2026-06-26 03:22Z

jira is like a harness for an organization

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

The Theatre, 1987, Texas

Ethan Mollick @emollick 2026-06-25 23:27Z

If it was just military implications, that would be one thing, but if this is about info security at scale (as Five Eyes suggests), publicly understanding risks and mitigations is incredibly important to tens of thousands of firms that need to prepare for what is next.