Backscroll ranked feed — 2026-06-30 12:17Z

A frozen ranked feed of 60 substantive tweets published on Backscroll (2026-06-30 12:17Z).

Joel Meyerowitz Cape Cod, 1976
2 likes · 1 retweets · 0 replies
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
1 likes · 0 retweets · 2 replies
he’s joking but this is what the coasean boundaries lead to: a firm size of “everybody”; ultra space communism. if you want to find another way out you need to give it a hard think
101 likes · 3 retweets · 14 replies
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
264 likes · 24 retweets · 12 replies
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
1 likes · 0 retweets · 0 replies
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.
2302 likes · 492 retweets · 224 replies
Most model evals focused on "real-world" settings (ex. TheAgentCompany, APEX-Agents, τ-bench) only target policy compliance and task success. They're useful, but they mainly treat models as tools: sandboxed settings, easy verification and primarily focused on the cost/perf
2 likes · 0 retweets · 2 replies
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
119 likes · 15 retweets · 16 replies
My theory is the relatively high property tax in the U.S. versus extremely low in most of Europe makes dynasties more likely to form in the later than the former.
54 likes · 1 retweets · 4 replies
3 month long interviewing marathon for a person at the end of their PhD (!), from a known uni in the us (!), with a track record (!), network (!) and generally very solid (judging by the blog and notes) Maybe we sometimes really don’t appreciate where we are and high bars are
676 likes · 18 retweets · 8 replies
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
47 likes · 13 retweets · 4 replies
Why was the denazification of West Germany so successful? Arguably, it's because it was antidemocratic. If the Western Allies had immediately held free elections after the war, they might just have produced another fascist government. But instead, the Allies transferred power
40 likes · 1 retweets · 0 replies
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
3 likes · 0 retweets · 1 replies
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
1 likes · 0 retweets · 1 replies
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
177 likes · 15 retweets · 13 replies
Good luck getting GPUs and memory for on prem deployment
10 likes · 0 retweets · 0 replies
There are levels to clustering….
360 likes · 9 retweets · 6 replies
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.
710 likes · 160 retweets · 38 replies
More seriously the main reason it cannot happen and that conditions everything downstream: no tech sector. That’s the critical lifeline for a frontier labs: even after the agentic turn, general economic demand is nowhere near mature enough to cover fixed compute scaling costs.
12 likes · 1 retweets · 2 replies
Blaze and basketball. Spanish Harlem, New York, 1975
38 likes · 6 retweets · 2 replies
@walden_yan and the team cooked with this one! While Devin has always done routing, Devin Fusion is a brand new architecture that delivers SOTA engineering task performance at 35% lower cost
1 likes · 0 retweets · 0 replies
imo every autoresearch is trying to figure out "the scaling law of the scaling law". - scaling law tells you what happens once you have fixed a config. It does not tell you how to train in the first place. - a scaling recipe is the missing map. It says how width, depth, batch
2 likes · 0 retweets · 0 replies
I am quite serious, for those thinking I am joking. Mid training is just too late to learn good features particularly worse when networks have been deep fried.
12 likes · 1 retweets · 2 replies
Mycena leana found a few minutes ago in Altoona, Pennsylvania.
15 likes · 1 retweets · 1 replies
Looks like anti-cheat software (GameGuard) is causing bun to crash. It’s interfering with system calls. https:// github.com/oven-sh/bun/is sues/22747 … GameGuard also causes other software to crash (example https:// issues.chromium.org/issues/40762507)
37 likes · 0 retweets · 4 replies
Some of you guessed right. Owl Alpha on @OpenRouter — that's us. Since going live, it has reached Top 3 globally by daily volume — and #1 on Hermes Agent, #2 on Claude Code, #3 on OpenClaw by monthly volume. Thank you to everyone who tested and used Owl Alpha during stealth
156 likes · 18 retweets · 15 replies
If people get healthier and live longer, the cleanest way to capture that isn't backing any single therapy. It's underwriting the longevity itself. A life insurer ( @meanwhile ) that bend the survival curve and wins twice: fewer claims, longer compounding runway.
5 likes · 0 retweets · 1 replies
i suspect if the cost of SOTA models don’t come down and open weight models continue to progress, we’ll see the use of SOTA models reserved to efforts debated and decided via committee or RFC processes within enterprises and engineering orgs
10 likes · 1 retweets · 1 replies
Turbopack: What's new in Next.js 16.3 Preview • Up to 90% less memory in dev • Up to 5.5x faster builds via Filesystem Cache • Faster HMR and dev startup • Experimental Rust React Compiler • 𝚒𝚖𝚙𝚘𝚛𝚝.𝚖𝚎𝚝𝚊.𝚐𝚕𝚘𝚋 support
55 likes · 11 retweets · 1 replies
auto-research at scale, agent-scale data mining, voting via mixture of agents doing these effectively means your Harness should support dispatching & orchestrating many subagents on massive scale data/tasks “read all 100k traces & experiment logs. find ways to reduce our token
10 likes · 1 retweets · 0 replies
What you guys are referring to as MCP, is in fact a Spec/Protocol, or as I've recently taken to calling it, a Spec plus a Protocol. The spec of an MCP server is flexible, can be represented through lazy loading, code mode, as a CLI whereas the protocol is a fully functioning
64 likes · 3 retweets · 8 replies
It's cool to learn how different warp/warpgroup roles overlap in the kernel, such as FA4. https:// tvm.apache.org/docs/tirx/nati ve_basics/cuda/profiling.html …
26 likes · 2 retweets · 1 replies
Chess engines tell you the best move. But grandmasters are human, they don’t always play it. So I built "Kibitz": a human move predictor for chess broadcasts. I trained this model on my Nvidia RTX 5080. Then I made it run as a business by itself. A channel buys the overlay,
625 likes · 51 retweets · 39 replies
This is the pinnacle of Chinese watchmaking The "Fam al Hut" Möbius This watch blows Richard Mille out of the water.. a 32000$ biaxial tourbillon, double retrograde, jumping hour, just how freaking insane is that... Makes me really happy that they won the GPHG award, because
199 likes · 8 retweets · 12 replies
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
13 likes · 5 retweets · 1 replies
Yan Diomandé has verbally agreed to sign a five year deal at Paris Saint-Germain. Salary and details agreed as @FabriceHawkins reports, he wants PSG. Negotiations starting between clubs in a separate deal than Maghnes Akliouche, both are ongoing.
32375 likes · 2715 retweets · 959 replies
Clear GBC: by me Clear GBC power regulator: by me Screen: by me
852 likes · 71 retweets · 23 replies
deliciously beautiful glass artwork by Jinya Zhao. hypnotic equilibrium of colors, forms, and air
845 likes · 92 retweets · 0 replies
Tomorrow, we’re launching the largest open-source dataset of gaming computer-use data. Gameplay screen recordings + keystrokes/mouse movements. Stay tuned :)
5 likes · 2 retweets · 1 replies
Game Boy Color Artist: @Meillex
1257 likes · 190 retweets · 0 replies
New preprint! Automated interpretability by approximating / replacing NN components (here attention heads) with programs.
15 likes · 0 retweets · 0 replies
EXCLUSIVE: French President Emmanuel Macron has instructed his administration to find new EU-wide taxes to finance the bloc’s next €2 trillion budget, according to five officials familiar with the discussions.
179 likes · 94 retweets · 34 replies
SpaceX and Charter have held executive-level talks about partnering on a consumer mobile phone offering, according to sources
110 likes · 23 retweets · 11 replies
Imagine we use the cheaper model for 200k context. 80% of the context is input, 20% is output. We get a blended cost of $0.16 (in) + $0.2 (out) -> $0.36
0 likes · 0 retweets · 1 replies
$3M ARR, can't raise an A, watching others raise $15M seed off a launch video
53 likes · 2 retweets · 4 replies
"ai-powered software" is like saying "electric computer" the descriptor might stick for a few things (electric toothbrush, electric car) but majority of tech will treat it as given and not even mention it
8 likes · 2 retweets · 2 replies
$70,000/mrr app done $100,000/mrr app next My biggest advice, Don’t listen to 99% of the Slop you hear on X It’s mostly all misleading Just to name a few: > shadow bans don’t exist, your content just sucks > 100 views average is not shadow banned > Ads do work
208 likes · 11 retweets · 22 replies
Protenix-v2 weights available now apples to apples comparison from Lucas Nivon post. Protenix ahead of all; AF3, Boltz, OF3-p2 on a blind test set of protein/ligand. So for Co-Folding use minimum Protenix-v2 https:// huggingface.co/TMF001/proteni x-v2-weights/tree/main …
18 likes · 4 retweets · 0 replies
AI image generators don't "draw"—they follow a compass: the score function, which points toward more probable images. The same compass drives Bayesian sampling and plasma physics. We built DiScoFormer to estimate the score far better when data gets complex.
0 likes · 0 retweets · 1 replies
Sneak peak of a photoshop plug-in I have been working on, halftone and print effects natively integrated, click of a button
24 likes · 3 retweets · 5 replies
The US housing market is now firmly in favor of buyers: 46% of home sellers offered concessions to buyers in May, the highest May reading on record. This percentage has DOUBLED since 2022 and also marks the 2nd consecutive annual increase. This comes as the US now has ~47%
956 likes · 103 retweets · 81 replies
As government restrictions on frontier models tighten, selling frontier AI as a service gets harder. So labs will increasingly use the models they can’t release to capture the value chains themselves. If you can’t sell to Berkshire Hathaway, you become Berkshire Hathaway.
131 likes · 18 retweets · 16 replies
I considered “trusting” as a replacement, but that’s only a good thing insofar as the trust is well-placed, and there are many ways modern westerners shouldn’t trust the society around them. Whereas courage seems much more broadly desirable.
2 likes · 0 retweets · 1 replies
Kanto Map Art #2: Route 1 #pokemon
702 likes · 47 retweets · 2 replies
Hyperliquid finally implementing rules that scale with the number of auction deployments HIP-3 deployers make. I expect this to trickle down across the HIP-3 mechanism design: eyeing dynamic stake requirements.
13 likes · 1 retweets · 3 replies
Rick Rubin already wrote the playbook on tech roles Experimenters 1. Slop cannon prototypers 2. Product builders Finishers 3. Sweepers 4. Growers 5. Maintainers
55 likes · 3 retweets · 5 replies
Excited to introduce TRON, a relighting framework for 3D captures. TRON pairs a neural renderer with 3D Gaussian reconstructions, achieving realistic quality, with 3D, material, & lighting control at interactive frame rates. https:// arxiv.org/abs/2606.11314 https:// research.nvidia.com/labs/sil/proje cts/tron/ …
33 likes · 17 retweets · 2 replies
I hear it everywhere. Speeches, emails, tweets, …
1 likes · 0 retweets · 0 replies
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.
184 likes · 10 retweets · 49 replies
someone sent me a Guardian article about this and asked “where is your principled concern for free speech?” so I started looking into the trial documents and pretty quickly concluded that a whole lot beyond speech was going on for all defendants excessive sentences, though
351 likes · 9 retweets · 17 replies

How it ranks: Backscroll reads my Twitter/X timeline and ranks every tweet for substance with an LLM taste rubric — not raw engagement. Curated by Surya Dantuluri.