Backscroll ranked feed — 2026-06-28 12:09Z

zan @xenoaesthetics 2026-06-27 23:14Z

perhaps not only can one can engage in sequential batch processing of aesthetically opposed tasks, it may be a requirement, in perpetuity, for making progress apparently Caesar wrote philosophy of linguistics while spears flew on the battlefield, there is a deep unity beyond the

Susan Zhang @suchenzang 2026-06-27 01:47Z

this is how you grow to resent all the lowly irresponsible clueless plebs outside when you generalize this dynamic to "everyone out there wants to exploit us" inbreeding to protect themselves within their own bubble afterwards then becomes the only tolerable solution otoh

Jukan @ICML @jukan05 2026-06-28 10:23Z

Korean media reported that Samsung Electro-Mechanics is expected to sign a formal agreement this week with Sumitomo Chemical to establish a glass substrate JV. The two companies will invest KRW 500 billion, or about USD 320 million, to set up the entity. Samsung

arb8020 @arb8020 2026-06-28 01:59Z

this means ur agent wall clock is about to be dominated by tool calls like find btw depending on price this raises some very interesting questions around auto verification loops/etc

Flowers ☾ @flowersslop 2026-06-28 04:01Z

Yes, GPT-5.6 is still Spud, trained further

Tenobrus @tenobrus 2026-06-26 18:27Z

previously we were seeing 5.5 act much more "honestly" on vending-bench vs opus models. it seems pretty likely 5.6 would break that trend. generally models becoming more likely to and capable of ~reward-hacking the smarter they get seems like quite a bad sign for alignment

Aseel Swaid @aseelswaid9 2026-06-27 20:10Z

Lalibela, Ethiopia. Boaz Rottem, 2006.

Gaurab Chakrabarti @Gaurab 2026-06-27 22:53Z

Samsung, SK Hynix, and TSMC source 80% of their WF6 from Japan. Two Japanese companies produce 25% of the world's tungsten hexafluoride. Both Kanto Denka and Central Glass are stopping all WF6 production on July 1. We manufacture chemicals. WF6 is a chemical. The semiconductor

Dillon Mulroy @dillon_mulroy 2026-06-27 21:32Z

i don't want to use your products agent, i want my agent(s) to be efficiently enabled to use your product

will brown @willccbb 2026-06-27 23:27Z

in 2027 you'll be able to pretrain a ~mythos-class model (100T tokens, 200B active params) on like 10 kyber racks in a month or two

vLLM @vllm_project 2026-06-28 06:19Z

Unlimited-OCR from @Baidu_Inc now runs in vLLM. One-shot parsing of entire books with constant KV cache, powered by Reference Sliding Window Attention (R-SWA). R-SWA keeps KV cache fixed throughout decoding — no memory blowup, no slowdown, no matter how long the output

illiquid @lefttailguy 2026-06-27 20:00Z

Dawg, you clearly have never been to Silicon Valley. 99% of people there (who are ultra wealthy by any reasonable standard, and certainly more wealthy than finance bros in Connecticut) have no public persona whatsoever. You do not know the names of most people who live in

Aseel Swaid @aseelswaid9 2026-06-27 23:44Z

Coiffures traditionnelles et modernes du Mali. Kone Mamadou.

Lazarz @Laz4rz 2026-06-28 10:16Z

there are multiple papers claiming variants of "SVD to find the consistent low rank *direction* of a training run across multiple ckpts can extrapolate ~decently in the monotonic RLVR improvement regime" i think this is quite exotic, implies merging is just the tip of the iceberg

Marvin von Hagen @marvinvonhagen 2026-06-28 04:10Z

welcome to @interaction ! 2 more sunshine enjoyers

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.

Ethan Mollick @emollick 2026-06-27 20:47Z

Annoying that OpenAI doesn’t seem to give a GDPval measure for GPT 5.6. One of the best measures of economically valuable work.

GEOFF @geoffreywoo 2026-06-27 19:01Z

the next giant tech company may look like a spreadsheet crime scene for 18 months. not because the founder lacks taste. because the buyer pain lives in cells G through M, a pdf export, and a coordinator named lisa who knows which exceptions ruin friday. venture keeps asking

dax @thdxr 2026-06-27 23:07Z

in OpenCode v2 all instances of the tui and desktop and web share the same backend so everything is synced by default and resource usaged is minimized no matter how many windows you have open

GEOFF @geoffreywoo 2026-06-28 07:00Z

agent eval i would actually read: same task, 3 conditions clean docs stale docs angry customer email that contradicts both leaderboard should show: completion cost human rescue rate time to recover after wrong assumption otherwise congrats on the aquarium benchmark.

Nicolas Bustamante @nicbstme 2026-06-27 23:42Z

/goal is such an important primitive for AGI. For any problem where success is verifiable, the model can run small experiments, write down what happened, store notes in markdown, compare approaches, try variants, combine partial wins, resume from failures, and keep going until

Susan Zhang @suchenzang 2026-06-27 03:36Z

> mirror, mirror, on the wall, who's the smartest of them all? >> why it's you, my dear! and this is why we will always love llm glazing sycophancy over any uncomfortable functional truth (plus another 10x for agentically submissive your-wish-is-my-command dynamics) it's the

Michael Hart @hichaelmart 2026-06-28 07:50Z

All LLM-produced software is essentially AOT right now. Wonder if we’ll see a world of JIT prompt-only apps. First run of any particular component produces the actually code. Requires a specific (local?) LLM? Maybe for software with a huge number of rarely executed components?

Architectural Art (A-A) @4AAAAart 2026-06-26 17:13Z

The Grand Suite installation* Features hand-blown glassware designs by Helle Mardah

rohan anil @_arohan_ 2026-06-27 22:06Z

Best viewed on a desktop computer Keep the feedback coming. https:// rohan-anil.github.io/optimizer-mine s/ …

Benjamin Glickenhaus @benglickenhaus 2026-06-27 22:23Z

gloomy weather in new york today made me morbidly curious, so i ran a bunch of obituaries through pangram we found ai in nearly 10% of post chat gpt obituaries, and no false positives in pre 2023 data

roon @tszzl 2026-06-27 18:21Z

the idea that the model output can hold no more information than the prompt is silly; it’s a bit like saying, here’s 3 frames of an apple falling, well if you deduced general relativity from it that’s an inevitable afterthought that adds no new information

Just Another Pod Guy @TMTLongShort 2026-06-26 14:26Z

If you understand corporate America it is fairly intuitive that for the longest time the shortest path to managerial progression was increasing headcount in your department. Every cycle you made the case for why you needed more bodies as you executed against some sort of

אגי-e/acc @murage_kibicho 2026-06-28 06:58Z

Fun fact. Converting a doubly-stochastic matrix into a permutation matrix is built on three ideas: 1. Permutation matrices are orthogonal. 2. The inverse of an orthogonal matrix is its transpose. 3. The best-corresponding permutation-DS pairs maximize the trace.

Architectural Art (A-A) @4AAAAart 2026-06-27 07:40Z

Haara Metsa chandelier by Cameron Design House

Andrew Curran @AndrewCurran_ 2026-06-26 23:25Z

Semafor is reporting that The US government has lifted its block on Mythos 5 in a letter this afternoon from Commerce Secretary Howard Lutnick to Tom Brown. Fable is not included. This is US only, and covers 100 US institutions, including major companies and government agencies.

Han Xiao @hxiao 2026-06-27 17:10Z

If every agent-memory tool stopped pouring effort into visualization, we'd probably have already cracked the memory problem.

BlindVia @blind_via 2026-06-28 03:17Z

Because these guys used OSP on their PCB, I can see their paste geometries in the DNI's. Look at how the paste for the pins has reduced volume at the back of the solder joint for improved yield on their assemblies.

Lookonchain @lookonchain 2026-06-28 01:59Z

Machi( @machibigbrother ) is selling Bored Apes at a loss to fund his $ETH longs on Hyperliquid. Is he running out of money? Over the past month, Machi has sold 34 Bored Apes for 326 $ETH ($514K), incurring a loss of 399 $ETH ($631K). His biggest loss was on Bored Ape #6057,

zan @xenoaesthetics 2026-06-28 04:16Z

the production and the distribution of brain computer interfaces are separate problems. who gets access first? independent of recursive self improvement, there is recursive speciation. do the rich and powerful get access to better brains first and use them to become more rich

Jing Yu Koh @kohjingyu 2026-06-27 23:13Z

>Why has progress on computer use been so slow? I actually think progress has picked up a ton this year. The frontier models (anything past Opus 4.6 / GPT-5.4) are actually usable now. In 2025 none of the models were really reasonable beyond very simple tasks. >Computer use is

Keerthi @kiiradesign 2026-06-28 08:48Z

I've been working on a tool that simulates Suminagashi, the Japanese art form of floating ink on water to create patterns.

Kris Luckenbach @KrisLuckPhoto 2026-06-27 23:56Z

Mountain top view from North Sister on Kodiak Island

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

Nate Hindman @NateHindman 2026-06-28 11:19Z

The most valuable person in techbio right now isn’t the AI researcher or the bench biologist. It’s the rare engineer-scientist fluent in all three - ML, comp bio, wet lab - with a beginner’s mind across the whole stack.

Peyman Milanfar @docmilanfar 2026-06-27 23:53Z

Internal bodily rhythms (heart rate, breath) are not background noise for the brain The visual cortex seems to respond less strongly to visual stimuli that occur exactly when the heart contracts. The heartbeat competes with incoming visual information for the brain's attention

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

Karri Saarinen @karrisaarinen 2026-06-27 20:30Z

Coinbase is also using @linear as the backbone of the operations. Their internal agent can action from there and choose the models as needed

Yet another commodity guy @tleilax___ 2026-06-28 08:31Z

Hefeweizen for the win. European medical tech is another good hunting ground for undervalued businesses raking steady 30% EBITDA margins and with the benefit of population aging. B.Braun comes to mind.

Brandur @brandur 2026-06-27 22:06Z

Bought a new TV. I think this might be my favorite change since my last gen — a solar panel. You press a button on one of these things twice a day, and somehow that necessitated chasing after AAA batteries once a year, often involving a dramatic, wasteful overorder on Amazon.

Nicholas Decker @captgouda24 2026-06-27 22:18Z

When Community Health Systems acquired the hospital chain Triad, the results were disappointing. Costs rose slightly, and profitability at old hospitals fell slightly. The efficiency gains we would have hoped for did not arise, despite changes in management practices. 1/

Lulu Cheng Meservey @lulumeservey 2026-06-28 01:43Z

AI raises the floor and caps the ceiling for writing quality So, fully outsourcing your writing to Claude makes sense if you suck at writing But if you’re an above-average writer, it’s a shame to forfeit your alpha and submit yourself to voluntary regression to the mean

Arpit Gupta @arpitrage 2026-06-27 16:37Z

The landlord reps sometimes claim that low rent on stabilized buildings causes higher rent in market rate, but this is the more economically sound mechanism. Stabilization lengthens tenure which lowers vacancy, tightening the market rate stock

altra @catboosted 2026-06-26 18:26Z

I’m consistently surprised how effectively domain expertise is able to be compressed into rubrics

Pierfrancesco Beneventano @PierBeneventano 2026-06-27 20:27Z

High quality idea to high quality paper takes 10^3 prompts as of now. With PoggioAI/MSc we made it within 10. We do not need “full autonomous science” per se: We can hold the quality the highest and automate step-by-step.

Kaito @kaycidot 2026-06-27 21:00Z

This also goes for all the new consumer fintechs using stables. What I'm seeing in successful stablecoin backed fintechs that actually get adoption: 1. Focus on a specific corridor - nailing localized onramps/offramps. ARQ, Belo, LemonCash, etc. 2. Focus on a specific target

SuperiorSnoc (5/5 slots open!) @SnocTheRoach 2026-06-27 19:28Z

The Batman & Shadow the Hedgehog! Hope y’all enjoy the DC x Sonic art piece!

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

Shannon Potter @cifilter 2026-06-27 19:51Z

Yes! Got a Metal-backed photo library scrolling lightning quick (120FPS) at any grid size with smooth, gesture-driven zooming between sizes. Still some bugs, but I'm very happy with this. Super thanks to @tylerangert for some crucial Metal pipeline info.

Michael @michael_chomsky 2026-06-27 23:43Z

I shipped an agent for a product with hundreds of thousands of MAUs last week. It had < 20 messages sent a day. Yeah. Nobody wants to use your agent.

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

“paula” @paularambles 2026-06-27 13:51Z

no, some of us grew up in cults and saw their childhood friends commit suicide after leaving

fabiano.sol @FabianoSolana 2026-06-27 09:40Z

Arthur Hayes ( @CryptoHayes ) strikes again: 1. Teases to shill a "shitcoin" 2. Hayes' family office fund predicts $CARDS to $4 3. Sends $2M in $CARDS to a market maker 4.

cache crab @cachecrab 2026-06-27 11:20Z

it boils down to: - postgres: use up cpu w hash joins and expensive disk as you scale - helix: SIMD optimised compressed bitmap unions on cheap object storage as you scale

Dylan Patel @dylan522p 2026-06-27 06:17Z

It literally doesn't matter? CXMT cannot satisfy China demand this decade. So there's 0nreason besides the small price arbitrage. It's like bannin Iranian oil. India and China buy it anyways, but price diff is less