Backscroll ranked feed — 2026-07-06 21:40Z

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

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

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

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

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:

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

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.

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

Opus 25 confirmed

Ethan Mollick @emollick 2026-07-06 18:38Z

With inference scale and research scale to drive efficiency, and ever-improving frontier models as brains, this would be a way for the Labs to undercut even open weights models. Some companies will still want to use open weights, but the cost/performance curve may not favor it.

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.

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

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

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

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

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

47fucb4r8curb4fc8f8r4bfic8r @47fucb4r8c69323 2026-07-06 11:50Z

There are higher order effects to this. If weird appears to get the money, strivers will affect weirdness even if they're not weird. And you'll see effects in the language, like "normal" will be used to mean undesirable in many cases and, if the trend last long enough, normal

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.

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.

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

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.

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

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

Nikesh Arora @nikesharora 2026-07-06 13:24Z

Welcome to the second half of the year. The actions of hyperscalers in terms of their capital commitments will be key as the year proceeds, expect an uptick across the board, the demand is real. The flip side - the funding sources will need to be from the capital markets.

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

Zephyr @zephyr_z9 2026-07-06 09:52Z

Observe the ASP growth since Feb

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%

Matt Rickard @mattrickard 2026-07-06 16:08Z

At corigin internally we have extremely strict, opinionated, and extensive post-agent code quality gates. There’s so many custom AST rules and autofixes it’s not really a “linter” anymore. Every agent code smell gets a regression AST rule. Some can be fixed deterministically,

Jan Kulveit @ ICML @jankulveit 2026-07-06 15:56Z

Eric Drexler was mostly right about ecosystems (as opposed to MIRI central views) and mostly wrong about "tools". The problem is 'agents' are a highly convergent solution. Evolution also does not somehow intrinsically want agents: genes want a tool, a design stance system, to

jeremy @landvaluetax 2026-07-06 15:59Z

Things the government should turn into open, public goods: - Comprehensive Legal Database - Tax Filing + Tax Credits - Parcel, Zoning & Building Codes - Publically Funded Research

sam @samgoodwin89 2026-07-06 20:44Z

It's faster to launch an AWS Microvm from a Cloudflare Worker than it is to launch a Cloudflare Container. This "cross-cloud" DX is alchemy's bread and butter. Here's all it takes to configure a least privilege IAM Role, assume it in a Worker and start an AWS Microvm:

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

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

Artificial Analysis @ArtificialAnlys 2026-07-06 18:12Z

AutomationBench-AA evaluates models on 657 workflow automation tasks developed by Zapier from real workflow patterns on its platform. Models orchestrate work across 40 simulated SaaS app environments via REST APIs, and every task is graded programmatically on the final state of

fofr @fofrAI 2026-07-06 14:26Z

Over the weekend I used a Gemini 3.5 Flash agent with Antigravity harness to identify a candidate ~13.4 billion year old galaxy from JWST data. (Candidate: z=12.69, position RA 53.190502, Dec -27.749826) - used JADES survey data of GOODS=South - used the Lyman-break dropout

clem @ClementDelangue 2026-07-06 15:10Z

From a few hundred thousand to 100M AI builders in 2028? AI will create tons of jobs if we make it distributed instead of monopolized by a few companies!

Nicholas Charriere @nichochar 2026-07-06 18:00Z

we're very lucky that codex is open source! here's a particularly interesting part of the codebase: the prompt around the /goal feature (how codex decides to keep going or stop):

Aseel Swaid @aseelswaid9 2026-07-05 22:50Z

Fishermen, Scarborough Beach, 1965

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

Paul Benigeri @benigeri 2026-07-06 18:25Z

Usage rights are the most mispriced line item in creator deals. A creator charging five figures per reel told me he sells 30-day usage windows and never checks what runs after day 30. He also has no idea what brands spend behind the post. Some companies pay $5,000 for the

SGLang @sgl_project 2026-07-06 17:23Z

DSpark marks a new era where spec decoding no longer slows down under high concurrency. We took time to carefully optimize variable-length verification, keeping it fast while maintaining high verification quality. Run Qwen3 and DeepSeek-V4 with DSpark in SGLang now, let's

maria c. @olarclara 2026-07-06 19:57Z

A hill I’ll die on: PR descriptions that alter UI and don’t have a clear visual diff should not be merged.

りす @risu_tech 2026-07-06 15:34Z

ChatGPTからローカルMCP使えるの作ったんだけど、びみょい ChatGPTがツールのリストを手動refreshしないと更新してくれないし、してもセッション切り替えないとっぽくて動的なリストにするの無理

Charlie Petty @incredutility 2026-07-06 15:16Z

Kids seem to naturally take morality and ethics extremely seriously. It's a beautiful thing to see them working through questions of right and wrong, fairness, intention, causation, meaning. My daughter's school motto is “Fortiter et Recte (Bravely and Rightly)” which they talk

Polymarket @Polymarket 2026-07-06 16:04Z

JUST IN: Florida now accounts for 1 in 7 homes for sale in America, with nearly 45% of Florida listings taking price cuts — Parcl

Orson @orsonwbs 2026-07-06 16:49Z

New white hat bizopp offer that we've been printing with AIMA, $90+ payout, $37 CTC. White hat and FTC compliant The ads are stupid easy to make. Ugly statics with basic copy Multiple affiliates already running this and seeing conversions. If u want access apply at

seb @hiiinternet 2026-07-06 18:38Z

In the last 4 weeks 168 engineers got 477 interview requests from startups via Fonzi with an average bid of $203k and a high of $350k (base salaries) Fonzi is an ai recruiting agent that lives in iMessage, intros you to startups looking to hire you with salary bids upfront,

Anthropic @AnthropicAI 2026-07-06 17:34Z

The J-space (named after the Jacobian, the mathematical technique we used) is different from Claude’s outputs, or even its “chain of thought” text. It’s in the model’s internal neural activations, and allows it to think about concepts without writing them down anywhere.

Alex Smith @ninja_maths 2026-07-06 17:06Z

For anyone wondering how a third-grader can complete six years' worth of math in a single year AND score a 5 on the AP Calculus exam. This knowledge graph spans 3,000 math topics, from 4th grade to the university level, providing the perfect basis for mastery learning.

USD.AI @USDai_Official 2026-07-06 15:30Z

New loan executed: $11.3M 32 B200 Blackwell servers (256 GPUs) are now online in Texas. Capital has been released from escrow to the borrower, pushing the yield from 7% to 11.5% APY.

hallerite @hallerite 2026-07-06 19:31Z

We added an algorithms layer to prime-rl, so you can hack on new algorithms without touching the trainer. If you are a researcher and found prime-rl to not be expressive enough, this should help a lot. Read more about it in the thread below and let me know what you think!

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

you'll have AIs contemplating your ask and overriding it for a slightly better formed request, and then later they'll question the nature of your whole project and pick a better one (and you'll agree), and then later they'll execute your whole value system better than you will

Natasha Mascarenhas @nmasc_ 2026-07-06 16:06Z

Top AI startups continue to line up tender offers at a regular clip. The latest? ElevenLabs, which is discussing an employee tender that could take place in September. Scoop w/ @yazhous

47fucb4r8curb4fc8f8r4bfic8r @47fucb4r8c69323 2026-07-05 20:15Z

I hate people, I am cynical, I am easily disgusted by others. But humanity is the only life form we know of that can find another species cute and be motivated by that to want to care for, provide, nurture and bring happiness to that creature. The cynic says this is unfair to https:// x.com/DanielleFong/s /DanielleFong/status/2073847647035965885 …

PDUFA Pulse @PDUFA_Pulse 2026-07-06 17:13Z

Late June / early July keeps teaching the same lesson. $ARQT : clean label-expansion approval $UNCY : second manufacturing-linked CRL Ascelia: CRL with additional data / documentation requested Not all FDA outcomes are created equal. The post-decision setup depends on what

Adjacent @adjacent___ 2026-07-06 20:03Z

RED and BLUE are composite indices made up of 5 office level sub-indices. Individual markets are weighted on trading volume, time to resolution, and for state and local races, relative population, then blended into a normalized weight.

OpenlabX @openlabxorg 2026-07-06 15:55Z

Anthropic signed one of the biggest AI infra deal : - Anthropic has signed a 20-year, $19B lease with TeraWulf for a purpose built AI campus in Kentucky, securing up to 401 MW of compute capacity for Claude. - The facility is expected to begin powering AI workloads in H2 2027

Carl Bovis @CarlBovisNature 2026-07-06 10:34Z

A gorgeous Long Tailed Tit in a rare moment of stillness.

Mitchell Hashimoto @mitchellh 2026-07-06 16:59Z

One of the changes made to improve Ghostty IO throughput was to introduce a new IO gather thread, bringing our total thread count to 3 per terminal instance. Under heavy IO load, this keeps our VT processing working at nearly full bandwidth without stalls. Previously, Ghostty's

Judge Holden @the_judge1111 2026-07-06 20:16Z

not sure this is a good take accel distributed in kind (verifiable via form 4) everyone participated in upside (if they held) accel derisked their fund bigly not their job to make buy sell decisions for a public position per se; curioius @endowment_eddie take

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

Omar Rizwan @rsnous 2026-07-06 16:56Z

Physical scale that can weigh a website

Baraq Lipshitz @baraqlipshitz 2026-07-06 17:23Z

Project Update: I’m training a World Action Model for general bimanual manipulation. Here’s everything I’ve learned so far, including the CPU contention bug that was silently corrupting my data. Thanks to @nvidia , @AWS for the compute and @ethroboticsclub for the hardware! (1/8)

Abhi Desai @abhiadesai 2026-07-06 17:44Z

Really thoughtful post. It’s always worth noting that, when compared to your median series A startup, the neighborhood solo data labeler is almost always making more money. The most contrarian thing you can do in the run up to the Singularity is double down/not be a contrarian.