Backscroll ranked feed — 2026-06-26 17:33Z

A frozen ranked feed of 60 substantive tweets published on Backscroll (2026-06-26 17:33Z).

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
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
it's not necessarily a victory for safety in the more absolutist unaligned ASI risk sense. but i do think these regulatory actions are probably safety victories wrt bio and cyber and general short-term mere megadeaths and societal disruption risk. if the USG has woken up enough
83 likes · 7 retweets · 12 replies
these people want ai models to be treated like nuclear weapons, researchers treated like they own nuclear state secrets during ww2, and geopolitical tensions escalated until we get ww3, where all production power can then be seized for the "ai industrial revolution" they strive
69 likes · 11 retweets · 4 replies
AI slop as straussian infrastructure
0 likes · 0 retweets · 0 replies
It's surprisingly plausible that we got GLP-1 drugs because of Danish tax law. For tax reasons, Novo Nordisk is actually controlled by a non-profit foundation. There are many downsides of this. But one possible upside is that it's helped the company avoid institutional drift.
82 likes · 3 retweets · 3 replies
When deploying video perception - often teams asked for more control over latency and sampling with their heuristics in mind. We are launching video frame support in our API. Now you can have stronger control on frame sampling and reduce TTFT through client side sampling.
2 likes · 0 retweets · 1 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
Three months after launching our traditional asset perpetuals exchange, we’d executed $1B in volume. Less than one month later we’re at >$2B, >$1B/month, without offering tokens, an internal market-maker, non-KYC access, crypto perpetuals, or individual trading, yet...
37 likes · 10 retweets · 2 replies
"Training for the Model You Return" Most LM pipelines return an EMA or averaged checkpoint, but optimizers still train like the final iterate is what matters. So if the output is an averaged model, can we shape training so that average gets better? This paper introduces PACE,
12 likes · 1 retweets · 2 replies
mcp was clearly inspired by looking at lsp and all the progress in the spec has been slowly realizing it should be nothing like lsp
67 likes · 3 retweets · 15 replies
Very exciting to see more work on benchmarking long-horizon CUAs! IMO there's only two main challenges remaining until CUAs become mainstream and productionizable: (1) effective long-horizon agents and (2) fast/realtime CUAs. I think we are going to get there very soon.
0 likes · 0 retweets · 0 replies
MirrorCode features 25 target programs spanning different areas of computing: Unix utilities, data serialization and query tools, bioinformatics, interpreters, static analysis, cryptography, and compression.
1 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
I hate “AI is bad because it becomes autonomous” genre. Anthropomorphizing does a great disservice relaying the actual risks Class power is about control over scarce resources. Everyone’s fighting for land because, as the saying goes, they’re not making more of it
0 likes · 0 retweets · 0 replies
hmm okay so now Ornith-1.0-35b-a3b is Qwen3.7-35b-a3b then? benchmark looks good, will give it a spin
16 likes · 1 retweets · 2 replies
Jake appears to be working for Buck Mason, as he consults on store furniture in Paris
115 likes · 2 retweets · 13 replies
A couple walks down a street in Sana'a, Yemen, 1997
56 likes · 5 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
OpenAI’s ARR stood at $40bn as of June.
112 likes · 11 retweets · 10 replies
TL;DR: Reservoir sampling. Given a sequence of positive real numbers q(1), q(2), ... and u(1), u(2) iid uniform [0, 1], then, for any n, the indexes of the top-K log(u(k))/q(k) in { 1 ... n } are distributed as independent p ∝ q(k) without replacement.
0 likes · 0 retweets · 0 replies
we built HANDBOOK.md: a new benchmark for long-context agentic instruction following. we drop an agent into a live company environment with files (PDFs, Excel, Word Docs, ...), tools (email, Slack, Jira, calendar, ...), and a dense corporate handbook (up to 124 pages), across 5
10 likes · 2 retweets · 0 replies
New: Granular evidence citations in QA Beyond chunk-level citations, our QA now points to the exact evidence snippet supporting each answer. Easier to inspect, verify, and trust. → Available with high-quality processing.
5 likes · 2 retweets · 1 replies
Conquest's Third Law says that you should model any bureaucratic organisation as if it was controlled by a cabal of its enemies. It's a fun exaggeration, but it gets at an important phenomenon. Organisations have continuous generational turnover. The founding members eventually
23 likes · 3 retweets · 3 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
Custom Land Rover Defender known as "Baloo" designed and built by The Landrovers
206 likes · 24 retweets · 1 replies
The Rent Guidelines Board recently voted for a two year rent freeze on NYC rent stabilized housing I was the lone dissenting vote I've written two pieces here; which highlight the challenges of a rent freeze and where to go next https:// vitalcitynyc.org/mamdani-rent-f reeze-agenda-nyc/ … https:// cityjournal.substack.com/p/the-high-cos t-of-new-yorks-rent-freeze …
46 likes · 12 retweets · 1 replies
PCB production is fucked at this rate. Looks like all the Japanese players have stopped accepting tungsten carbide cutting tools orders
32 likes · 1 retweets · 6 replies
finally read the OSTP NSTM-4 memorandum so proudly cited by anthropic, and turns out the entire thing is just a self-own whine-fest about getting commoditized at a fraction of the cost by maliciously exploitative chynese actors who dare remove the performative "security
80 likes · 7 retweets · 9 replies
Can someone explain to me why $PS , Pershing Square’s management company at a $12B mkt cap, trades at 50% of fee-paying AUM, 40x revenues and 150x Operating Earnings, when its closed-end funds($PSH.LN and $PSUS ) trade at massive 20-30% discounts to NAV…?!
222 likes · 21 retweets · 40 replies
30m x 1m output token trace is like upper bound of $750m revenue at 25 dollar per million output tokens. As an upper bound. https:// x.com/mtslive/status /mtslive/status/2070141140607832353 …
204 likes · 5 retweets · 9 replies
I should probably do a longer thread on why both of those companies can win Anything that touches enterprise is very unlikely to grow 0-10mm in year 1 without 3mm in pure sales opex (sales compensation to bookings typically run 3x) So your ability to grow that quickly in
3 likes · 0 retweets · 2 replies
For sure this guy works @tryramp He’s also never met anyone at Slash We’re a 65 person company doing 300M in revenue No one here is a grifter
12 likes · 0 retweets · 1 replies
A Palestinian woman and an Arabian horse. Captured by Aseel Swaid
12 likes · 2 retweets · 0 replies
Yesterday's drop didn't come from cascading liquidations. It came from coordinated spot selling across all exchanges at the same time. Coinbase, Binance, Bitfinex, Kraken, Bybit. Negative CVD across all of them, with spikes concentrated in the last two days. Negative spot
224 likes · 10 retweets · 16 replies
this is an arrangement of 50 Cent's "In Da Club" by my pianist michael potter. somehow he turned this certified hood classic into an epic that's in 11 and has metric modulations all over the place. but yeah other than that it sounds exactly like the original
6 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
Poore's law
332 likes · 5 retweets · 7 replies
Overheard in Florida: “The West doesn’t exist. It died with the French Revolution. America supplanted it.”
330 likes · 21 retweets · 34 replies
The cool result is the usage across all departments incl eg legal increasing massively (Codex vs other usage, ie chat) Codex and Cowork are a great unlock for all knowledge work (ala Karpathy) but most are not used to this yet
4 likes · 0 retweets · 1 replies
Gemini API reliability has been brutal lately. ~17% success rate right now. Below 50% overall for the past 4 days. I never had this kind of issue with OpenAI’s API. @googledevs @OfficialLoganK is this being fixed, or should I migrate right now?
13 likes · 2 retweets · 2 replies
Did you know that there is a women's badminton group stage match from 2012 Olympics that has million+ views on YT? It's referenced in the Stanford Algorithmic game theory course: due to poor tournament design in this match both teams were trying to LOSE!
2 likes · 1 retweets · 1 replies
While we eagerly await Fable 5's return, our agentic WebGPU kernel optimization framework kept running. Opus 4.8 picked up where Fable left off, pushing Liquid AI's new LFM2.5 230M to an unbelievable 1,400 tok/s... running locally in your browser. Don't blink or you'll miss it.
113 likes · 21 retweets · 14 replies
Arsenal made verbal offer to sign Bruno Guimaraes from Newcastle United. #AFC proposal through intermediaries below £60m - rejected + #NUFC insist sale of 28yo #Brazil int’l midfielder will not be entertained. W/ @gunnerblog @ChrisDHWaugh @TheAthleticFC
33503 likes · 5951 retweets · 1490 replies
It is a coincidence that 'emoji' and 'emoticon' sound alike. 'Emoji' is a Japanese term meaning 'pictograph.' 'e' is "picture, drawing" + 'moji' is “letter, character” 'Emoticon' is 'emotion' + 'icon.'
786 likes · 82 retweets · 12 replies
The actual Nixon historical legacy is big government soft-leftism:
19 likes · 10 retweets · 3 replies
Look at how that green (teal? cyan? celes? edgar?) line at the bottom is flat as hell—so even in times of increasing total homelessness like post-pandemic, the people actually living rough on the streets is pretty consistent and has been for twenty years. There’s a bit of a drop
1 likes · 0 retweets · 1 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
Autonomy isn't the ability to act without human supervision. It's the ability to *learn* without human bottlenecks in the process. A system that is fully dependent on human training data and RL environments is only an imprint of human knowledge.
195 likes · 21 retweets · 27 replies
From OSWorld 1.0 to 2.0, we went from minutes (~30 steps) to hours (~318), from single apps to real workflows, from high scores (83%) to hard problems (21%). 1+ year, 20+ people, every task rigorously verified. This is what real cua evaluation takes. http:// osworld-v2.xlang.ai
19 likes · 7 retweets · 0 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
#ProjectOfTheDay: Sea of the Umbrians by Paolo Belard Inspired by the woven forms of traditional fishing nets, this proposal transforms Lake Trasimeno's piers into places for swimming and gathering. The project also replaces artificial beaches with restored reed habitats.
18 likes · 4 retweets · 0 replies
Some thoughts after a day of using codex: - Proofreading papers is awesome! "proofread this" and then "implement all edits" and "compile and commit to github". And then I can tune the proofreading (how to present changes, stylistic stuff, etc.) and dump it in a skill
28 likes · 1 retweets · 1 replies
This is kind of stunning. It's not just SpaceX's stock that's been getting slammed, even its recent bond sale is doing so badly traders can't remember something similar having happened before. https:// bloomberg.com/news/articles/ 2026-06-26/bond-traders-stunned-as-losses-on-spacex-s-new-debt-keep-growing?srnd=homepage-americas …
232 likes · 55 retweets · 18 replies
It's been two weeks and the leading US model, Fable, is still banned by the White House with no explicit Congressional Authorization, no written rule that was broken, and no technical justification. In the meantime, our Chinese adversaries released a model that rivals our best.
112 likes · 19 retweets · 1 replies
When market making, once you have your tail latencies under control and an accurate, fast OMS, it becomes easier to crank out large daily volume with relatively little capital. This is when fee tiers, small edges and a bit of alpha really add up. A 0.5 bip edge on $10M daily
27 likes · 2 retweets · 1 replies
And GLM synth pipelines now plugged on batched dockers (ok singularity since that’s what we can install on cluster — aptly named).
23 likes · 0 retweets · 2 replies
Live inference data is the best training data for speculators. The hard part is training them online without storing hidden states or impacting inference performance and reliability. We built exactly that into the Baseten Inference Stack. Result: +20% median acceptance rates,
45 likes · 6 retweets · 3 replies
They bought distressed debt at ~1 cent on the dollar and got headline credit for the sticker price. Obviously great that they spent $5.5M getting people out of debt but all the headlines are about $550M
36 likes · 2 retweets · 3 replies
native iPad app with partial erasing, crushing 120fps on a 2018 12.9inch Pro
59 likes · 0 retweets · 4 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.