Backscroll ranked feed — 2026-06-26 02:04Z

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

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

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

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

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

The Theatre, 1987, Texas

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 …

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

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)

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

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

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

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

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

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

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

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

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

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.

dax @thdxr 2026-06-26 01:58Z

the news about 5.6 is being phrased in a certain way but i think it's going to be a thing where it's in a review period for a bit during which access is restricted (because it's unreviewed) and then it'll be made public like any other model

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

John Ketcham @JKetcham91 2026-06-26 01:06Z

The Rent Guidelines Board just voted to freeze rent-stabilized rents for one- and two-year leases—despite its own evidence showing operating costs rose 5.3% over the past year. As RGB member and @ManhattanInst senior fellow @arpitrage explains in @cityjournal ’s Substack,

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

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!

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

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.”

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.

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

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.

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

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

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.

Timothy Luong (Chongz) @chongz 2026-06-25 21:47Z

What in the decel shit is this

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

Poore's law

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.

Susan Zhang @suchenzang 2026-06-25 14:13Z

instead of indulging in all this fear-mongering frontier-stealing narrative slop, @Alibaba_Qwen might as well release all "29 million Claude exchanges" for the benefit of open-source research and call it a day let the public decide for themselves how "dangerously critical" these

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)

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.

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

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

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 …

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

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

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

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

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

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

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,

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.

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

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

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.

Rudolf Laine @LRudL_ 2026-06-25 22:59Z

With nuclear, we effectively banned civilian use and thus all benefits, while the actually dangerous weapons stuff continued unabated behind closed doors “Minimize the benefits and maximize the harms”

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

A woman harvesting roses - Oman

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 !

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

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.

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

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

Algeria can’t risk a loss because of GD

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

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!!!

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.

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

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

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

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

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