Backscroll ranked feed — 2026-07-01 03:12Z

A frozen ranked feed of 60 substantive tweets published on Backscroll (2026-07-01 03:12Z).

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 doesn't change a thing? CXMT memory is all sold and deployed. It's just a question of the small price arbitrage that Chinese vendors get from cxmt vs the big 3.
0 likes · 0 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
Everyone is training on synthetic/RL environments now (actually, many drawing from materials coming from Europe). Essentially how pretraining data stopped being a commodity and all about skills.
0 likes · 0 retweets · 0 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
building verifiers has shown time and time again how important ownership of your harness is good luck trying to patch closed source harnesses to make them more useful to train or eval
11 likes · 1 retweets · 0 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
EXCLUSIVE | Nathaniel Brown to FC Bayern - DONE DEAL Full agreement with Eintracht now reached after Brown had already agreed personal terms with Bayern as revealed. €55m. Contract until 2031. Brown only wanted Bayern. Bayern have now finalised everything, beating
5551 likes · 585 retweets · 178 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
Alienkind just raised a ₹30 crore round - this company was widely dismissed by observers as “Indiranagar PMF” since its 1st store went viral on Instagram in March 2025. This round was was at a ₹250 crore post money valuation (almost x3 the first round which was at ₹80 crore)
183 likes · 5 retweets · 8 replies
No joke: when i saw this tweet earlier today was soooo close of digging up your tweet from a few years ago where you basically introduced the term to the world by saying "we're hiring for midtraining team" Even funnier is Google now also recently has an Omni team!
19 likes · 0 retweets · 2 replies
OpenAI evaluated its new models on PostTrainBench-Lite, a shortened version of our original benchmark that gives agents 5 hours instead of 10 to improve an open-source base model. GPT-5.6 Sol and Terra outperform GPT-5.5, but still rely on narrow strategies and sometimes
7 likes · 2 retweets · 0 replies
Excited to release ParallelKernelBench (PKB), a benchmark for measuring LLMs’ ability to write fast multi-GPU kernels! Multi-GPU kernel generation compounds several hard problems: - a large parallelism design space - a new communication axis to optimize - and
33 likes · 14 retweets · 2 replies
Grammarly (now called Superhuman) is essentially building a productivity suite via acquisitions: - Coda (Dec 2024): docs - Superhuman (Jul 2025): email - Rows (Feb 2026): spreadsheets, folded into Coda - GPTZero (Jun 2026): AI detection
6 likes · 0 retweets · 1 replies
A couple walks down a street in Sana'a, Yemen, 1997
56 likes · 5 retweets · 1 replies
President of @Blackstone Jon Gray says LLM spend among the company's portcos was up 15-fold in Q1 of this year over last year. At the same time, the pace of implementation is frustratingly slow, so it still feels like we're in the very early days of the AI cycle: "It's sort of
24 likes · 2 retweets · 1 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
The Theatre, 1987, Texas
67 likes · 7 retweets · 1 replies
There are few ways to build an economic moat for frontier AI firms. One is architectural breakthrough that's closed to other firms (everyone's working on this, but so far little evidence). The second is enterprise lock in (OG microsoft playbook). An AI "cloud" that is active
6 likes · 2 retweets · 2 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
rebuilt the figma "vector editor" for the web draggable points and cool morphing shapes code: http:// arlan.me/vault
139 likes · 3 retweets · 8 replies
I really want CLI tools to start speaking directly to coding agents with errors. Instead of: “Error:” Return: “Here’s what broke. Here’s how to investigate. Here’s how to generate a redacted repro. Here’s where to send it.” Obviously there are privacy/security challenges to
8 likes · 1 retweets · 3 replies
$3M ARR, can't raise an A, watching others raise $15M seed off a launch video
53 likes · 2 retweets · 4 replies
if you sketch out all of the systems required to make vercel run, you will come to the realization that as long as the rollback service is not circularly dependent on vercel everything else can ~literally be on vercel
38 likes · 0 retweets · 1 replies
so @tradexyz is in such a dominant position within hyperliquid HIP-3. as a new startup HIP-3 deployer, you need to: 1. somehow get 500k $HYPE to stake (>$30m) 2. bid the tickers for ~500 HYPE ($30k) 3. just to break-even on the tickers at 80% discounted fees (growth
50 likes · 2 retweets · 7 replies
Apollo, a large manager, put out a chart showing deep distress is concentrating in small private credit funds. Loans marked below 50: ~12.5% at small funds, ~8% at big ones
71 likes · 9 retweets · 6 replies
Yes. The degree to which chat is really being relegated. And codex / Claude code have gigantic motion even outside engineering is very true
6 likes · 0 retweets · 0 replies
imagine matrix rain but the glyphs are additive gaussian-falloff streaks converging to a vanishing point. the floor rays leave the portal on a vertical tangent then cubic-bezier into perspective, and a domain-warped shader churns in a webgl buffer behind the gate #p5js
14 likes · 0 retweets · 2 replies
GLM 5.2 is 2nd in Vending-Bench. Each GLM release has improved at a remarkably steady pace: a linear fit of R²=0.99, with almost $1k better per month.
149 likes · 15 retweets · 2 replies
New preprint: Training for the model you return Modern LM pipelines often return an averaged model (e.g. EMA), rather than the final iterate, but optimizers are still mostly designed around the final iterate. How should we change training if we know we will return an EMA?
172 likes · 40 retweets · 5 replies
All of the new prediction market startups remind me of all of the Uniswap forks - the biggest problem is not the technology, it's getting enough liquidity.
13 likes · 1 retweets · 5 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
Three friends wanted a shared weekend retreat on seven acres in Hyderabad. Their architects tilted the entire plan 45 degrees so that no two rooms share the same view of the land, then organised everything around a 60 by 30 foot central courtyard the way traditional Indian houses
185 likes · 30 retweets · 7 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
there are levels to building evals lvl 1: using a spreadsheet qa pairs lvl 2: using public agent evals lvl 3: manually label private evals lvl 4: traces to evals and skills lvl 5: turn every prompt & traces into self healing loops almost everyone gets stuck before level 3
39 likes · 1 retweets · 5 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
AI statelets, capped in size by a meta-ethical commitment to human exit rights, must propagate themselves by sexual reproduction: a genetic recombination of the constitution of any N statelets forming a new statelet overmind. with environmental fitness to spare, sexual selection
45 likes · 2 retweets · 4 replies
METR: "GPT-5.6 Sol exhibited an unusually high detected rate of “cheating”" OpenAI: "We believe these behaviors may reflect improved instruction following and training intended to increase persistence"
0 likes · 0 retweets · 0 replies
People tend to know about Liberia (the American-backed Back-to-Africa failed project) but not enough people know that it's right next to Sierra Leone (the British Back-to-Africa failed project)
1376 likes · 78 retweets · 15 replies
it boils down to: - postgres: use up cpu w hash joins and expensive disk as you scale - helix: SIMD optimised compressed bitmap unions on cheap object storage as you scale
21 likes · 2 retweets · 3 replies
Level 10?
25 likes · 0 retweets · 2 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
Pay ranges for autonomous driving companies
0 likes · 0 retweets · 0 replies
obviously I tested it! measured on a 5090, Q6_K, think-off, same harness as the base and the prior version: - quality flat: q_avg 93.7 (base 94.0, prior 94.1): weights didn't move - MTP draft head recovered: 1.9 to 2.3x, back on the base curve, ~144 tok/s peak - agentic score
16 likes · 0 retweets · 3 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
I turned my bookstore into an RPG for Free RPG Day. You earn XP for each $ spent, for attending events, and completing quests.
11 likes · 0 retweets · 3 replies
Pretty good progress! Trained a model on my teleop data and deployed it on the arm. The robot attempts to tidy an object by itself - it moves to the tape, tries to grab it and heads for the bowl. Doesn't manage the grasp, but it seems to understand the gist of the task.
3 likes · 2 retweets · 1 replies
Excited to share new @LaudeInstitute project, where we ask: When does continual learning actually require learning? We test 8 adaption methods as the world shifts and find ones best at learning new facts are worst at unlearning what's now outdated.
9 likes · 2 retweets · 2 replies
Relatedly -- my favorite anti-monopsony bait: If sticky wages are via monopsony (rather than monopolistic unions as usual), then, *tightening monetary policy **raises** employment & output*... (for the same reason that a higher minimum wage boosts employment under monopsony)
20 likes · 4 retweets · 2 replies
relationality is the syntax of thinking; the way things are arranged in relation to one another determines how they behave together. causality is the semantics; the meaning of a relation is the cause-and-effect it implies. temporality is the meta-syntax; the rules governing how
5 likes · 0 retweets · 1 replies
TSA put the fake weapon in my tray and the scanner guy missed it
6330 likes · 44 retweets · 43 replies
Big news! We launched a new competition. Fully open LLMs only (only OLMo, SmolLM, Apertus families base models allowed) and expert proof grading. We use six IMO-style problems. You can discuss the LLM outputs live here: https:// kaggle.com/competitions/a i-mathematical-olympiad-proof-pilot/discussion?sort=undefined …
4 likes · 0 retweets · 1 replies
In rats fed a high-fat, high-sugar diet Selegiline significantly reduced visceral, whole body and subcutaneous fat, and liver inflammation markers (IL-6, TNF-α) Selegiline likely achieves this by inhibiting MAO-B in fat cells, which can be achieved at low doses of the compound
29 likes · 0 retweets · 3 replies
The core claim is that prediction markets are a specific case of state-contingent pricing. In Arrow-Debreu terms, a complete market has claims that pay in particular future states. One state realizes, the corresponding claim pays. Hyperliquid is interesting because it is moving
33 likes · 1 retweets · 6 replies
agent due diligence checklist: can it read the stale PDF? can it survive the cursed dropdown? can it tell karen in billing “no” without creating a lawsuit? can it finish the task when the API doc lies? or did we just give a chatbot a lanyard.
4 likes · 0 retweets · 0 replies
i've never heard a human say belt-and-suspenders but AIs love this phrase apparently
4 likes · 0 retweets · 3 replies
NEW: House Dems reeling from NY primary results - One called it an "earthquake" & a "huge defeat" for leadership - "It was a tough night," Rep. Meeks, Queens Dem Party chair, told @axios - Senior House Dem, asked if members are worried: "Yes they are"
690 likes · 123 retweets · 15 replies
Sources: Google expands the scope of its months-old AI coding strike team to "midtraining" to better compete with Anthropic, after key executive departures ( @erinkwoo / The Information) (Visit Techmeme dot com for the link and full context!)
48 likes · 8 retweets · 1 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
Gene expression & co-expression across 630,000 RNA-seq samples from 10,700+ plant species — free, in your browser. Search a gene → see where it's expressed on the UMAP → explore its co-expression network. https:// mutwil-plant-atlas-viewer.hf.space #PlantScience #RNAseq #Bioinformatics
116 likes · 38 retweets · 3 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.