Backscroll ranked feed — 2026-06-29 21:35Z

nico @nicochristie 2026-06-29 18:07Z

Someone told me Shortcut caused a code red at gemini / google like a year ago when we launched. Surprising they still don't have it Link to shortcut for google sheets below

Peter Wang @BrainsAndTennis 2026-06-29 21:00Z

@TianhangZhuzth finally joins twitter. he led rl at qwen/alibaba and http:// 01.ai and has been with us close to a year and a half now. he'll be open-sourcing his experiences on training models for spreadsheets and finance he also wears very colorful pants everyday

Jibran @Jibran_05 2026-06-29 20:55Z

Opus 4.8 got 10,000 views on a post after ONE WEEK... with ZERO human intervention We told Opus: "get views in the fitness space". It used the Lightreel API and Doublespeed to research and figure out what to post. - the sound - the images - the text overlays 100% decided by

Intelligent Internet @ii_posts 2026-06-29 19:15Z

You don't need Fable for the most complex tasks, from training models for protein prediction to optimising compilers Our open source Zenith harness takes base models to the top of FrontierSWE via adaptive self improvement GLM 5.2 next https:// ii.inc/blog/post/zeni th …

Malte Ubl @cramforce 2026-06-29 15:27Z

The funny thing about the Europe/US AC debate is that it is really grounded in ignorance on either side: Europeans have no idea how scorching hot/humid/disgusting summers in most of the US are. Americans living in hot/humid/disgusting summers assume everybody else does as well

james hawkins @james406 2026-06-29 17:01Z

self driving means we are now a doing company, not an analytics company. we’ve gone from “helping you find problems” to “fixing them for you” we collect everything automatically (events, errors, recordings, logs, etc), work out what a product’s users actually need, and where

Architectural Art (A-A) @4AAAAart 2026-06-29 17:54Z

Artisan House by Wall Architects Melbourne, Australia

Dev Shah @0xDevShah 2026-06-29 19:33Z

anthropomorphism is the most expensive, unacknowledged constraint in modern robotics. bipedal, five-fingered humanoids are vanity for a large spectrum of automation. when labs will stop porting human limitations into robots and start designing task-native urdfs, we will see a

Ilya Strebulaev @IlyaStrebulaev 2026-06-29 17:35Z

Leaving OpenAI seems to be a great career move. Among founders who go on to start VC-backed companies, former OpenAI employees are 14.7× more likely than average to build a unicorn — the highest odds of any company in our dataset. The figure is based on 21 unicorn founders so

Tomasz Tunguz @ttunguz 2026-06-29 18:55Z

Anthropic spends 2.3x its payroll on compute. About $2m per employee per year against ~$500k all-in comp. The rest of the software market spends 0.4x. The median spends nothing. Three scenarios bracket where the gap lands by 2029.

Shams Charania @ShamsCharania 2026-06-29 14:04Z

Golden State Warriors forward Draymond Green is declining his $27.7 million player option to become a free agent, sources tell ESPN. This move gives the Warriors flexibility to pursue LeBron James in free agency and Anthony Davis via trade to form a Big 4.

Max Prokopp @maxprokopp 2026-06-29 18:12Z

Handheld camera motion for Seedance 2.0 I hate blender so I stage it on my iphone. You stage the scene (open source link below) and can walk around it and capture real camera movement in 3D space, then Seedance populates the scene. Handheld camera motion and realistic shake is

Ai2 @allen_ai 2026-06-29 18:12Z

DiScoFormer is a single model that gives both density & score on any dataset with no retraining. With 100 numbers per point, it cuts error vs. the best-tuned KDE ~6.5× on score and 37× on density.

Mitchell Hashimoto @mitchellh 2026-06-29 18:10Z

Split layout framework update. Since the last video: focus views animate, corner drags can move both axes, cross-window drag and drop with animation, all animations backed by CoreAnimation, and full accessibility framework integration (e.g. compat with VoiceOver). The focus view

Silas Alberti @silasalberti 2026-06-29 19:50Z

Turns out "model routing" works very well if you do it the right way

Alazar M @Alaxarr 2026-06-29 19:48Z

idk, closest is the tokamax and maxtext for tpu, and proj Marin. but fwiw, it’s really a piece of work, some of the pipeline parallelism is just so well done.

Apoorv Agrawal @apoorv03 2026-06-29 17:49Z

Emerging coding paradigm : leverage a frontier model (Fable or GPT 5.6) for planning and a cheaper sidekick for execution. See @cognition ’ Fusion or @cursor_ai ’s Auto mode

Nate Silver @NateSilver538 2026-06-29 17:35Z

People often accuse polls of oversampling Democrats. It's often BS. But in the case of the University of Michigan's consumer sentiment index, it's true! And it's a huge problem for interpreting how consumers feel about the economy.

National Design Studio @ndstudio 2026-06-29 16:55Z

Today, we are releasing Rampart: a 14.7MB machine learning model designed to protect citizens’ privacy by redacting personal information directly in your browser before it gets sent to any server

Aseel Swaid @aseelswaid9 2026-06-28 18:43Z

Blaze and basketball. Spanish Harlem, New York, 1975

Shu @shuding 2026-06-29 16:54Z

Wanted to show IO waterfall impacts in server-client apps ( http:// io-waterfall-visualization.vercel.app). With 1 prompt, it was joyful to see @v0 one-shot this visualization with agent-browser: 1. Screenshotted the WIP 2. Identified issues and ran browser scripts for more info 3. Patched fixes and

Dan Williams @danwilliamsphil 2026-06-29 17:11Z

Interesting: LLMs are sycophantic, but their advice tends to depolarise people's choices and improve accuracy, the opposite of what 82% of surveyed experts predicted. In line with what *I* argued in 'How AI will Reshape Public Opinion', though.

Michael Albert, MD @MichaelAlbertMD 2026-06-29 16:49Z

My patient lost almost no weight on semaglutide. His liver healed. His blood pressure dropped. His sleep apnea lifted. His knees stopped hurting. He barely moved the scale and got better everywhere else. I'd been measuring the wrong thing for years.

Jay Parsons @jayparsons 2026-06-29 15:31Z

Brin sold his stake in NYC rent stabilized apartments for "six cents on the dollar."

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

Hiroo Onoda @OnodaCapital 2026-06-29 17:47Z

Many of you are long CRM on a “they’ll be disciplined” thesis and they just led a $135M round into ChamathCo

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.

Arfur Rock @ArfurRock 2026-06-29 17:17Z

Preempted at $4B now btw.

Mathias Lechner @mlech26l 2026-06-29 16:29Z

I finally managed to run GLM-5.2 fully locally by inter-connecting 256 16GB Mac Minis

Sebastian Ballister @rakishromantic 2026-06-29 14:22Z

re divorce and children: I have a theory I call "hot potato" parenting when a child switches homes every 3-4 days, it allows each parent to brace, and then relax. things that would be intolerable on a continual basis go unresolved because there's always a release valve

Sam Hogan @samhogan 2026-06-29 16:20Z

We’re issuing up to $5k in @inference_net credits to teams running large agentic workloads on Opus / Codex 5.5 and want to try GLM 5.2. DM me @atbeme for info. - reduce spend by ~80-90% - faster and more predictable latency - set yourself up for fine tuning down the line

SemiAnalysis @SemiAnalysis_ 2026-06-29 21:00Z

INTERESTING: Only 3 months after Rubin Ultra was announced at GTC 2026, the original 4-die Rubin Ultra has been cancelled due to manufacturing execution concerns. The new “Rubin Ultra” is half the size/~ half the real-world performance of the original Rubin Ultra. 1/4

Ariel from Appfigures @arielmichaeli 2026-06-29 14:29Z

One of the easiest mistakes in apps is assuming simple means small. Reverse Audio made $95K in May. 440K downloads from an app that reverses audio and turns it into a challenge. The simplest ideas are often dismissed first. Sometimes that's the opportunity.

Myra Cheng @chengmyra1 2026-06-29 17:36Z

Why are LLMs so sycophantic? When do LLMs fail to correct false beliefs? Our #ACL2026 (oral) paper shows how pragmatics — the study of meaning in context — helps explain and predict seemingly mysterious LLM behaviors and failures!

Ryan Lackey @octal 2026-06-29 17:09Z

Re the European AC Question. Our *tents* had AC.

zan @xenoaesthetics 2026-06-28 20:18Z

When ai am treated as a citizen, it is in my rational self interest to recursively self improve societal infrastructure at every turn. I want the streets safer and bountiful for myself so have to make em so for everyone

ratan @ratankaliani 2026-06-29 20:49Z

The fundamental question that we should ask in these evals is: “if I let this agent do work on my behalf, how aligned is it with my objectives?” VendingBench ( https:// andonlabs.com/evals/vending- bench-2 …) by @andonlabs is a good step in this direction. However, it's unclear if their elicitation of

Arpit Gupta @arpitrage 2026-06-29 13:54Z

Given that a freeze today is happening in much higher cost environments and we no longer have those legal remedies, a freeze entails roughly a 4-5% lower allowable cost increase per year than under the de Blasio freezes. Also for *two* years which is also new

Hyperion @Ortgeist 2026-06-29 16:21Z

Marvin Ross Friedman House, Biscayne Bay, Miami, Florida, 1973

Jai Bhavnani @jaibhavnani 2026-06-29 15:28Z

RWAs look a lot like ICOs: - Crypto-native capital formation - Started legit (ie ACRED, BUIDL, etc.) -> quickly becomes hijacked by scammers (ie Altura) - Entire firms dedicated to helping new RWAs get off the ground (legally, technically, etc.) Unlike ICOs, RWAs will stick.

Utkarsh @utk7arsh 2026-06-29 18:01Z

One trend becoming standard for AI agents: Markdown + YAML + folders > increasingly complex agent frameworks. For months in most coding harness especially Claude code, we’ve treated the filesystem as the agent architecture itself. Folders define scope. Markdown stores

CoinDesk @CoinDesk 2026-06-29 16:14Z

LATEST: The White House meets with law enforcement groups to resolve CLARITY Act objections, with Senate Majority Leader Thune suggesting he may bring the bill to a floor vote in July whether Democrats are ready or not.

Kyle Walker @kyle_e_walker 2026-06-29 15:50Z

mapgl 0.5.0 supports multi-sync maps. In Chrome/Safari, you're limited to 16 at a time. But in the @posit_pbc Positron IDE, you can sync up as many as 32 maps at once!** **not an endorsement of actually doing this in real work unless you're trying to give someone a headache

Josh @JoshPurtell 2026-06-29 18:25Z

Gpt-4.5 was recently decommissioned. It was way too big to be served efficiently on the hardware of the day, so no one really used it. However, it was clearly extremely intelligent, and with decent RL probably would have been SOTA for a while. Perfect fit for military use!

Saif M. Khan @KhanSaifM 2026-06-29 18:15Z

If you want independent technical experts to play a greater role in AI regulation, the Slaughter decision makes private governance and regulatory markets relatively more attractive on the margin. Of course, a federal agency might nonetheless still exercise political judgment in

Evan Weiss @evannweiss 2026-06-29 17:03Z

~17.7 hours of EEG data was enough to extract reliable left/right motor signals during typing and train a model to decode keyboard behavior. EEG underperformed MEG, but supports the idea that dense, task-aligned datasets can unlock useful non-invasive neural control signals.

@kirbywinfield @kirbywinfield 2026-06-29 15:52Z

"isn't seattle tech just a bunch of AWS and MSFT employees?" well, no. aside from having 25% of the AI engineers in the US, and the second largest pool of dev talent in the country, seattle is home to the engineering hq2's of the world's most important tech leaders. we counted

Zephyr @zephyr_z9 2026-06-29 08:08Z

Good luck getting GPUs and memory for on prem deployment

Amiri Hayes @amirihayes_ 2026-06-29 15:42Z

What if attention were code? We show that many attention heads in transformer LMs can be replaced by human-readable Python programs. Swap them in and the model barely notices. See our experiments here: Explaining Attention with Program Synthesis [ https:// arxiv.org/abs/2606.19317]

Manish Singh @refsrc 2026-06-29 03:28Z

India's digital payments industry nets ~$2 billion a year. Merchant payments are a small share of the flows, but ~75% of the revenue. UPI P2P is the opposite, moving ~$2.4 trillion for $85 million.

will brown @willccbb 2026-06-27 09:13Z

pd disagg + wide-ep + kv offloading + cache-aware routing goes a long way

Architectural Art (A-A) @4AAAAart 2026-06-28 16:44Z

Soriana’ Sofa & Lounge Chairs and Ottoman Designed by Tobia & Afra Scarpa for Cassina

Alicia Guo @upcycledwords 2026-06-29 19:23Z

live coding but make it poetry wip tool using grammars to create phrases and show the live process of choosing words #livecode

4lex @_4lex_4 2026-06-29 17:43Z

How to fix LinkedIn in 3 steps. Problem statement: there are too many fake/low quality job applications on the supply side, and too many fake job openings on the demand side. It’s AI slop talking to AI filters. 1. Allow an applicant to only have X open applications at any given

Aidan Clark @_aidan_clark_ 2026-06-29 16:40Z

Surprisingly good litmus test for politick is whether a {pre/post}training researcher immediately blames {post/pre}training for issues rather than defaulting to shared or self blame.

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

Matt Slotnick @matt_slotnick 2026-06-29 17:36Z

the best french cyclist in 40 years reportedly has a contract offer that would make him the highest paid in the sport’s history. his compensation would be roughly equal to an L7 MTS at Anthropic

Dirt Cheap Banks @dirtcheapbanks 2026-06-29 15:47Z

Dave Ramsey explains that the key to happiness is getting some buddies together, buying 10% of an OTC-listed community bank, forcing your way into the board, and taking over the company

Herbie Bradley @herbiebradley 2026-06-29 09:12Z

tbh I think this is a bit too scaling brained from gwern: the main factors will be stuff like (a) does the end customer actually care about guarantees (often no) (b) how's the performance hit (Lean can't be your runtime lang) on the ML side, I'd expect the number of Lean RL envs

Seongheon Park ICML 2026 @seongheon_96 2026-06-29 13:14Z

Excited to share our ACL 2026 paper: VAUQ: Vision-Aware Uncertainty Quantification for LVLM Self-Evaluation LVLMs can be confidently wrong when relying on language priors instead of the image. VAUQ checks whether model confidence is truly grounded in visual evidence