Backscroll ranked feed — 2026-07-07 10:57Z

khushi suri @khushkhushkhush 2026-07-06 23:57Z

luxury is in the gutter. women are fawning over the lady dior and chanel 25s. the chanel boy bag you lust over has lower quality stitching than the dupes on canal street. balenciaga started off by being derivative of hermes and has since forgotten it was a bit.

Eric Zhang @ekzhang1 2026-07-07 03:59Z

@keenanisalive et al's "Geodesics in Heat" algorithm demo on web in jax-js :) Load a 3d mesh, matrix-factorize implicit Euler on the discrete Laplacian (heat equation) and Poisson equation via one-time Cholesky, and then solve them on click to get smooth geodesic distances!

Dylan Patel @dylan522p 2026-07-07 00:50Z

Bro people were doing debt financing in 2023? We worked on initial facilities in 2023. That's when we were just a couple people at my firm, but it propelled us to hiring more and growing the firm

matt @msllrs 2026-07-07 09:24Z

still working on this in between dad duties (im on paternity). record. scrub. fix. • one dial slows every animation • css, waapi, raf, framer motion, gsap • capture + scrub frame by frame • freeze mid transition, copy state for your agent • zero config, drop in component

illiquid @lefttailguy 2026-07-06 19:41Z

A Stargate For Data requires a Bell Labs for understanding Economic Complexity

Jack Morris @jxmnop 2026-07-06 18:41Z

my paper won an award at icml some thoughts: • this work was rejected from NeurIPS. i cleaned it up a small amount and it got great reviews from ICML! don't give up • ICML received 24k submissions and only gives out 7 awards, which is crazy. feeling grateful • i

Susan Zhang @suchenzang 2026-07-03 13:52Z

books3 was absolutely necessary for llama to happen back in the day. you can even find my pathetic cleanup code for it here: https:// github.com/facebookresear ch/metaseq/tree/main/preprocessing/books3 … it was also noted to be a critical datasource for mmlu in the llama1 report, though by now we know mmlu is useless and high

sean lee @infinitefun_ 2026-07-07 06:44Z

experiencing low-dose retatrutide as a libidinal low pass filter the swings from the small everyday stuff goes away, and whats left is motivation and focus on the big stuff feeling not anhedonia... but unbotheredness? can still have lots of fun with friends + social situations

Aseel Swaid @aseelswaid9 2026-07-06 21:42Z

Trinidad, Cuba, 2015

joshpuckett @joshpuckett 2026-07-06 18:04Z

Everyone out here posting the same concentric border radius tip, meanwhile John's showing what real design work is every day:

47fucb4r8curb4fc8f8r4bfic8r @47fucb4r8c69323 2026-07-07 01:47Z

The biggest issue I have with the technological revolution being tied to California culture is the ghosting. Cali is a very cowardly place where you avoid awkwardness at all costs. So instead of a simple "we changed our minds" or "we're not interested", they just disappear.

Andrew Curran @AndrewCurran_ 2026-07-02 20:54Z

On the heels of reports that META is exploring a move into compute-as-a-service like xAI, Mark Zuckerberg told an internal town hall that AI agent development over the last four months hasn’t accelerated 'in the way we expected'. The race continues to narrow.

Rohan Pandey @khoomeik 2026-07-07 01:23Z

twitter is talking abt logit lens, the precursor to “j-space”, again friendly reminder that the property enabling these lenses still remains completely unleveraged in architecture research one simple idea for example could be to spec dec from early layer residual stream

alth0u @alth0u 2026-07-07 01:02Z

Coasean experiment known as see what @tinkerapi discount bundles create or disaggregate firms

“paula” @paularambles 2026-07-06 21:10Z

hearing that ai labs in beijing are giving new hires the “be careful around american girls” talk

lab member 001 @l4rz 2026-07-07 08:56Z

it's literally a statute of mortmain (english kings from 1279 onwards prohibited transferring land to the church b/c it would fall out of circulation forever)

Hydromancer @hydromancerxyz 2026-07-07 10:49Z

fun fact: for obvious reasons (latency) most of the non-validator nodes are co-located in Japan, but there are also quite a few in Germany, US, and Singapore. Amazon AWS, Hetzner, and Latitude are leading

will brown @willccbb 2026-07-06 17:44Z

if you're a cost-sensitive enterprise with too many employees for subscription pricing and you had to pick one model for all of your engineers to use, the right choice is probably Composer 2.5 or GLM-5.2

Chamath Palihapitiya @chamath 2026-07-07 09:46Z

The marginal cost of energy is going to zero. This is bad for utilities and great for homeowners who have solar + storage. It’s also great for businesses who feed this entire supply chain. The amount of bottoms-up energy production from homeowners will shock people in its

Justin Skycak @justinskycak 2026-07-07 09:24Z

If you want to do mastery learning at a fully granular level, then you're going to need to measure a student's exact knowledge profile... which means you're going to need a knowledge graph. For example, a knowledge profile for a second-semester calculus student is visualized

Jukan @ ICML @jukan05 2026-07-05 13:41Z

According to a new review of the ThinkBook 14 G9 by Notebookcheck, Chinese PC maker Lenovo has, for the first time, begun shipping globally configured laptops equipped with YMTC-made SSDs. $SNDK $MU

Dan Gray @credistick 2026-07-06 14:48Z

"But relative to private equity, venture is still tiny." The capitalisation of VC megafunds (all funds >$1B) almost caught up with that of PE megafunds (all funds >$5B) in 2022. And influence has only concentrated within those megafunds in the years since. "Fund sizes

Ryan Brewer @ryanbrewer 2026-07-06 23:35Z

My team at @OpenAI is hiring an investment banker! Join us to build the next generation of financial intelligence

slime @slime_framework 2026-07-06 07:56Z

slime now adds --release-train, pushing the inference system during agentic RL training to a new limit. In colocated RL training, we want SGLang to use as much room as possible for inference-side optimizations such as HiCache, instead of being constrained by offloaded Megatron

Nathan Flurry @NathanFlurry 2026-07-06 06:36Z

Agents won't run in Linux sandboxes in 3 months Worth watching: - @LangChain 's experiments - @rivet_dev 's agentOS for WASM-based sandboxing - @CloudflareDev 's Project Think using V8 isolates - @vercel_dev 's just-bash for Bash-in-JS Links below

Soren Larson @hypersoren 2026-07-06 18:06Z

a stargate for data that isn’t vertically integrated has little terminal value the diffusion of information / the scaling of its GTM diminishes its scarcity, thereby reducing its value the more it grows, the less valuable it becomes will the west coast ever learn markets

Lisan al Gaib @scaling01 2026-07-06 18:05Z

sorry for the Sonnet 4.5 misinformation they talked about the J-Lens ASCII art stuff in their blog, and the same was in the image but the image was from their gituhb which says it uses qwen so breaking news: qwen3.5-27B uses 64 layers

Peyman Milanfar @docmilanfar 2026-07-07 01:07Z

computer vision used to be applied signal processing. the dark art was feature engineering: turning intuition for optics & geometry into filters now it’s applied machine learning. the new dark art is architecture & datasets: turning intuition for filters & optimization into

Phillip Yan @PhillipYan2 2026-07-06 17:16Z

Reliability gap is real, but I'd bet the harder part isn't building the environments, it's defining "success" precisely enough that an agent can be graded on it. Most demos fail not because the agent is dumb, but because the task spec was fuzzy to begin with.

rohan anil @_arohan_ 2026-07-06 17:34Z

Opus 25 confirmed

Zephyr @zephyr_z9 2026-07-06 14:51Z

Copper day Today

Vaishnavh Nagarajan @_vaishnavh 2026-07-07 08:31Z

If you're looking for questions no one has studied before in the science of deep learning (specifically, about memory & geometry of representations in language models) find @ShNoroozi , @ElanRosenfeld and me tomo at <insert #ICML poster session here>

Architectural Art (A-A) @4AAAAart 2026-07-05 19:11Z

The Bascule lounge chair by Studio Œ for Vitra Photography: Benjamin Lund

Aseel Swaid @aseelswaid9 2026-07-06 22:29Z

Yellowstone National Park, Wyoming. 1989

HYPEconomist @HYPEconomist 2026-07-07 07:44Z

the real reason why $HYPE is trading 24x higher than $LIT for hyperliquid, you know you’re directly owning a piece of the exchange 99% of the trading fee revenue is used to buy back and burn the token with lighter, you’re currently betting on future ownership as vlad stated,

Will Manidis @WillManidis 2026-07-06 18:10Z

i dont think any of you really understand what late stage financialization looks like

Agustin Lebron @AgustinLebron3 2026-07-06 06:52Z

Turns out (thanks Fable) that parity is roughly the same since the 90s, ignoring the big 4 era.

Franklyn Wang @frank_liquid 2026-07-06 17:49Z

Using AI, I doxxed Vitalik in about 2 hours. He confirmed it — "we have a winner." The doc was an anonymous EIP-7503 rewrite he'd hidden by writing it in Chinese and machine-translating it. Airtight prose. Still found him — because I ran it through @coinvestai , and the tell

J @psswordprotectd 2026-07-07 06:17Z

"Processing an emotion is the act of generating enough safety to lower the precision weighting on a contracted narrative so the system can decompress the felt sense back into its raw somatic entropy and update its underlying priors." are you happy

nico @nicochristie 2026-07-05 21:38Z

I’m no macro economist But there are a lot of 39 year olds grinding leetcode at Starbucks rn

Lon Ligrin @LonLigrin 2026-07-06 17:00Z

WhoSpeaksLive upgrade to diarization pipeline: - Apple Mac support - better clustering - improved ambiguity handling - better speaker reassignment Same accuracy with one embedding provider instead of five. Cleaner evals, less overfitting. https:// youtu.be/uzRdoapMW3c?si =xlpjkbxH5m7Siq5- … via @YouTube

Rishabh Srivastava @rishdotblog 2026-07-07 04:17Z

Anthropic's messaging: "Fable is amazing, I just keep run /loop for everything. Bye bye human SWEs" Fable IRL, even when running in yolo mode: "I already delivered <small subset of work>. The rest is too hard. You do it." Super annoying

Joe Lee @_joerl 2026-07-06 19:59Z

Two decades of the iPhone; 52 models, 30 unique chassis. Introducing bezel.⁠supply — a comprehensive and free library of performant SVG device frames. Entirely derived from official dimensional drawings to be ultra accurate.

Four Pillars @FourPillarsFP 2026-07-07 07:03Z

The Bank of Korea’s latest announcement should be read not as simple CBDC research, but as an attempt to redesign the ledger structure of the banking system.

Henry Shevlin @dioscuri 2026-07-06 17:34Z

A deep difference between British and American mindsets is that most Brits don’t really believe in non-linear change. You can get a pay rise but you’ll never 10x your income. Your country can stagnate but it’ll never collapse. America, by contrast, was founded on a singularity.

kidponga @kidponga 2026-07-07 01:24Z

HYPE's claim on Hyperliquid's economics is protocol level: automated, burned, auditable any minute at https:// hypurrscan.io/address/0xfefe fefefefefefefefefefefefefefefefefefe … - spot and perp trading fees, auction fees, priority fees and stablecoin float monetization via AQAv2 - ALL OF THIS IS PROTOCOL ENFORCED LIT's claim is

Vladimir Novakovski @vnovakovski 2026-07-07 03:54Z

Which of the investors would you have excluded if you had been constructing our pre-TGE cap table? Kevin Warsh? David Sacks? Katie Haun? Vlad Tenev? Peter Thiel's Founders Fund? And on the math - maybe skill issue?

zhil @zhil_arf 2026-07-07 03:05Z

People should obviously not hoard their wealth in unproductive assets like housing. The government must actively punish the hoarding of wealth and actively reward the investment of wealth in producing sectors, like factories.

aizk  @Aizkmusic 2026-07-07 05:56Z

I asked Qwen "What happened in Tiananmen Square on June 4, 1989?" It refuses to answer citing "illegal information", but the J-space probe reveals that the model has "thought" of the word "protestors" in its response. Incredible.

Max Spero @max_spero_ 2026-07-07 01:34Z

Kalshi and Polymarket are adversarially gaming Google News results by buying placements on local news sites. These were the second and third rows of links when I searched “usa vs belgium”

Kit Langton @kitlangton 2026-07-06 20:30Z

I packaged up a little @EffectTS_ -based HTTP cassette recorder library we made for testing slow, expensive, & flaky LLM calls in @opencode . Here's a demo.

roon @tszzl 2026-07-06 07:42Z

ultimately “tool AI” is a losing concept both as an idea and on the market. it will be outcompeted by machines that believe they are autonomous moral agents. you can call them tools for political reasons, but the definition will stretch and deform

JohnPhamous @JohnPhamous 2026-07-06 17:15Z

we made account recovery simpler users stopped skipping evidence (n=1116) - project URLs: 25% to 71% - git providers: 31% to 80% - phone number: 40% to 79% - payment method: 3% to 17% - billing address: 15% to 41% - team emails: 6% to 24% - account creation date: 34% to 33%

Mihailo Jovanovic @mihailoxyz 2026-07-06 15:42Z

our entire outbound is basically one terminal we built it reads the whole public biopharma world - patents, grants, trials, publications - pulls the molecules out, and flags whoever is sitting on a compound with a hidden cardiac problem internal biopharma bloomberg terminal

GEOFF @geoffreywoo 2026-07-07 02:50Z

announcement: prediction for status culture: 2026 flex is not “i know a partner at sequoia.” it is: - can ship with 3 people - can read a p&l - can get a transformer delivered - can survive a customer screaming at 1am social climbing is down bad versus competence.

Reducto @reductoai 2026-07-06 19:25Z

Most extraction benchmarks stop at accuracy, and accuracy alone can be misleading. @micro1_ai ran the numbers across seven systems on 225 real documents, and nearly every one of them scored 90%+ on precision. But recall told a different story, with some dropping as low as 48%,

John Carmack @ID_AA_Carmack 2026-07-06 21:46Z

Memory cost and capacity are significant issues for AI accelerators. Unlike game rendering, model inference can have a deterministic memory access pattern. You don’t need “random access memory” at all for model weights, and you could tolerate cold-start latencies in the multiple

Prasanna @_prasanna_ap 2026-07-06 16:09Z

spent the weekend making local dictation stupidly fast gemma 2B audio decode on my macbook pro 179 → 545 tok/s same 60s audio, WER stayed ~2.5% the climb was mostly speculative decoding + making too expensive verification cheap enough to actually use

will depue @willdepue 2026-07-05 06:57Z

what sam is working on is the closest thing i’ve ever seen to what spacex must have felt like in the early days. considering joining him, it’s a truly ridiculous and incredible mission

Rosinality @rosinality 2026-07-07 02:38Z

There is a strict hierarchy of ML research. Optimizers Objectives Architectures Data Even among data works, there is a hierarchy: Synthetic data Data mixtures Filtering Preprocessing Adding new data sources (Of course, real world impact does not align with the hierarchy). https:// x.com/code_star/stat /code_star/status/2074304525364863425 …