Backscroll ranked feed — 2026-07-07 15:14Z

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.

Zephyr @zephyr_z9 2026-07-07 14:33Z

I had to track down the original source cited by Reuters. Reuters exaggerated its conclusions and misrepresented China’s consideration of limiting the weights of SOTA models. Tiered plan applies to all open-source technologies, not model weights. links↓ x.com/jukan05/status…

Ryan Brewer @ryanbrewer 2026-07-07 14:36Z

Doesn't count equity my guy

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!

Aseel Swaid @aseelswaid9 2026-07-07 13:28Z

The Singapore International Monetary Exchange, 1997

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

Soren Larson @hypersoren 2026-07-07 11:27Z

what is the TAM for lab inference in discovery where 1 you may not have leverage to force participation in upside of discovery and 2 customers in the biggest businesses of discovery don’t trust you and are trying to bypass you

Lazarz @Laz4rz 2026-07-07 11:05Z

.there is literally no way, after context compaction it thinks that the 32k context was a mistake and corrected 8k checkpoint means 8k context again “I found two plausible local mini configs: the earlier accidental 32k model run and the corrected 8k model run. I’ll compare the

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

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

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

David Sinclair @davidasinclair 2026-07-07 13:47Z

Article is AI slop. Hallmarks of Aging (2013) is just a catalog, nothing original The "reprogramming thesis" rests on work done from 1995-1999 in the Guarente lab @MIT showing aging in yeast cells is epigenetically modifiable http:// cell.com/cell/fulltext/ S0092-8674(00)80219-6 …

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.

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

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

prinz @deredleritt3r 2026-07-07 14:06Z

Reuters: - The CCP has held meetings with top Chinese AI labs (including Alibaba, ByteDance and http:// Z.ai) about potentially restricting overseas access to China's most advanced AI models, including both closed and open-source models. - Under the proposal being

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:

dany @l2000002 2026-07-07 14:16Z

185-205k lmao

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

Warren Pies @WarrenPies 2026-07-07 12:22Z

Despite the nervousness around compute demand ($META), B200 rental rates have shot up over the past week ($5.30 to $5.77 - @Silicon_Data ). This confirms what we have seen in B200 availability, which collapsed back to 0% a few weeks back (excerpt from @3F_Research H2 Outlook).

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

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

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

simp 4 satoshi @iamgingertrash 2026-07-07 13:04Z

Aging societies will move from flow-based economies to capital appreciation based economies The old will stop producing labor, and humanoids are at least a decade away Japan is the first And will kick start a tidal wave in the demand for tokenization of assets 2027

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.

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)

Eugene Y. @eugene_mindset 2026-07-07 11:20Z

design idea for a water tracking app. the fill level is a real fluid sim reacting to your gyroscope built with a webgl fragment shader, damped pendulum physics, refraction and chromatic aberration i can share the file if anyone wants to play with it

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

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

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

XY Han @XYHan_ 2026-07-07 12:59Z

Most misleading part clarified: The “J-space” is not a separate, hidden space. It is an alternative *coordinate system* for intermediate layer activations. Using a Jacobian between the last layer right before unembedding and the intermediate layer, you can “move” rows of the

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

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

AI Security Institute @AISecurityInst 2026-07-07 14:15Z

Every recent model we tried surfaced findings worth patching, and the strongest chained multiple flaws into a real attack path. The takeaway is that capable AI can help defenders harden production systems. Read the new case study unpacking what we found (and fixed):

Aseel Swaid @aseelswaid9 2026-07-07 13:14Z

Haiti, 1973

ak0 @annanay 2026-07-07 12:26Z

Incoming top quantitative intern (litigation desk) at SIG giving me harvard university (extension school) vibes

Gabriel @gabriell_lab 2026-07-07 13:49Z

I love micrographics, but creating them manually every time is painful. So I built a small generator with Claude + Codex. It mixes simple shapes, icons, colors, and layouts into endless random compositions. You can tweak complexity, change the color system, or just hit random.

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

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

Maziyar PANAHI @MaziyarPanahi 2026-07-07 12:00Z

I finally managed to use Unlimited-OCR and Gemma 4 together in OpenMed. A real patient chart: read, de-identified, and mapped to FHIR on one laptop. No cloud, no API key, nothing leaves the machine. All via llama.cpp, all free on @huggingface . What do we do next?

Rosinality @rosinality 2026-07-07 13:18Z

https:// arxiv.org/abs/2607.04763 Multi-turn OPD. Use teacher-generated prefixes and let student generate actions on top of them, and do OPD on it, and give more weight to early step trajectories (because the gap between the teacher and the student is smaller). Prefix RL?

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

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

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

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

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

Opus 25 confirmed

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

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

414 Magyar's Birds @414magyarbirds 2026-07-07 09:09Z

USF Birds played Battleship with the shadow fleet tankers today: The shadow fleet is leaving the chat. The battle for fuel supplies to Crimea in the Sea of Azov: 8 shadow-fleet tankers of the rf were hunted down and lit up by pilots of “Kairos” from the 414th Brigade “Magyar’s

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.

Phillip Yan @PhillipYan2 2026-07-06 17:13Z

Kitchen analogy tracks but the hard part is the expediter role: someone (or something) has to catch when the fast-worker shipped something the planner didn't actually ask for. Multi-model setups fail more on handoff drift than on any single model being weak.

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

Binh Pham @pham_blnh 2026-07-07 12:17Z

n=40 shouldn’t give much info, for a difference of 5% between policies, n should be ~1000 to be correct (off the back of my head, this is based on power analysis) with 40, with your baseline of 10% success rate, you can only faithfully compare down to 20% difference (anything

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

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

@jason @Jason 2026-07-07 11:23Z

two surging markets for tokens have emerged, with philosophical and economic arguments for each the frontier models will have a significant edge in specific verticals and UX/polish, while open source will dominate in extensibility, privacy and cost AI will intelligently route

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

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

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>

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.

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?

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%

Trevor Noren @trevornoren 2026-07-07 13:14Z

Bernstein: "Even if tomorrow you could snap your fingers and build an enterprise-class application, fully feature-rich, and it costs you zero — all the other costs and ramp times are still there, closing the clients, bringing the software on to market. That creates an

Matt Carey @mattzcarey 2026-07-07 09:52Z

a harness is a compiler for prompts, event bus for a distributed system and a orchestrator of not so durable streams.

Alfredo Canziani @ ICML @alfcnz 2026-07-07 08:35Z

Submissions to #ICML2026 doubled this year wrt last year. Yet the work load was handled successfully. Result: 8 poster sessions overlapping with orals. Suggestion: remove the orals. They are pretty useless. Talking at the poster gets you the most out of the authors.