Backscroll ranked feed — 2026-07-01 01:28Z

Aseel Swaid @aseelswaid9 2026-06-25 21:48Z

The Theatre, 1987, Texas

Florian Brand @xeophon 2026-06-27 05:30Z

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

Dwarkesh Patel @dwarkesh_sp 2026-06-26 16:04Z

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.

dax @thdxr 2026-06-26 14:03Z

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

rohan anil @_arohan_ 2026-06-25 17:33Z

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.

Steve Faulkner @southpolesteve 2026-06-27 01:11Z

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

Alexander Doria @Dorialexander 2026-06-27 09:08Z

Yeah vllm/sglang do work quite well know, batching ensures a good theoretical throughout but then everyone want to Claude Code and you have to manage X parallel sessions, each with varying latency.

goodalexander @goodalexander 2026-06-25 17:34Z

> get super stoked > 328 tps > wow that's sick > sign up for account > run quick benchmark bc I'm not a trusting person > it's 58 tps the same as everyone else thanks for playing https:// x.com/Yuchenj_UW/sta /Yuchenj_UW/status/2070166719839326396 …

Tanay Jaipuria @tanayj 2026-06-26 18:03Z

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

TBPN @tbpn 2026-06-24 19:30Z

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

Adam Karvonen @a_karvonen 2026-06-26 18:04Z

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"

Ekaeo @Ekaeoq 2026-06-25 14:23Z

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

Kevin Kwok @kevinakwok 2026-06-25 17:12Z

Yes. The degree to which chat is really being relegated. And codex / Claude code have gigantic motion even outside engineering is very true

Dylan Patel @dylan522p 2026-06-27 06:17Z

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

Rhys @RhysSullivan 2026-06-26 01:26Z

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

Karina @karinanguyen 2026-06-26 18:39Z

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

Aseel Swaid @aseelswaid9 2026-06-25 22:40Z

A couple walks down a street in Sana'a, Yemen, 1997

Nathan @asplencmnt 2026-06-23 20:24Z

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

Florian Plettenberg @Plettigoal 2026-06-25 17:39Z

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

Dillon Mulroy @dillon_mulroy 2026-06-26 14:09Z

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

Nayan Saxena @SaxenaNayan 2026-06-25 21:28Z

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.

Lucas Beyer (bl16) @giffmana 2026-06-25 19:42Z

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!

donn @tzedonn 2026-06-28 11:47Z

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

Frogo @Dubflip 2026-06-27 12:05Z

TSA put the fake weapon in my tray and the scanner guy missed it

Hadley Harris @Hadley 2026-06-24 12:32Z

$3M ARR, can't raise an A, watching others raise $15M seed off a launch video

Rahul Mathur @Rahul_J_Mathur 2026-06-27 04:43Z

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)

Adam Block @adam_block66985 2026-06-25 13:05Z

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?

David Ornstein @David_Ornstein 2026-06-26 09:27Z

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

Wenting @zephray_wenting 2026-06-29 06:26Z

Might sound a bit cursed, but I turned the M5Stack PaperS3 into a GameBoy emulator:

Aleksandr @alexfeeel 2026-06-23 17:11Z

Build Small's bet: no giant model, no data center — all under 32B, on hardware you own. I took it on a real one: a chest X-ray second-opinion tool. Two 4B models + plain logic. No cloud API. ~30s audit on a hosted GPU. @Gradio @huggingface @NVIDIA https:// huggingface.co/spaces/build-s mall-hackathon/cxr-draft-auditor …

cache crab @cachecrab 2026-06-27 11:20Z

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

Arlan @arlanoska 2026-06-25 16:00Z

rebuilt the figma "vector editor" for the web draggable points and cool morphing shapes code: http:// arlan.me/vault

SHAV★ @shavnyuy 2026-06-26 17:43Z

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

Brett Harrison @BrettHarrison 2026-06-26 15:32Z

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...

Prathamesh @invinciDesigns 2026-06-25 18:16Z

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

Yash Bhalgat @ysbhalgat 2026-06-25 08:47Z

Level 10?

Epoch AI @EpochAIResearch 2026-06-26 15:25Z

MirrorCode features 25 target programs spanning different areas of computing: Unix utilities, data serialization and query tools, bioinformatics, interpreters, static analysis, cryptography, and compression.

thebes @voooooogel 2026-06-24 01:49Z

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

Warden of the East @ChampSantiago 2026-06-28 13:13Z

Most Black suburban families are literally 1 generation removed from the hood…Yall fiend for our class divisions to mirror old-money white America when our wealth isn’t insulated enough for “the suburbs” & “the hood” to be fully separate worlds

witcheer @witcheer 2026-06-23 11:18Z

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

QC @QiaochuYuan 2026-06-29 16:48Z

so on the one hand i think social interactions that explicitly ban phones are genuinely valuable and likely to become more popular. on the other hand taking away people’s phones also gives you meaningful power over them which is why it’s also a cult thing

Alex Bouaziz @Bouazizalex 2026-06-26 13:10Z

I heard Masa may be returning

Kaito @kaycidot 2026-06-25 19:00Z

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.

Zuhayeer Musa @zuhayeer 2026-06-26 18:33Z

Pay ranges for autonomous driving companies

Andon Labs @andonlabs 2026-06-25 17:53Z

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.

Nick Dobos @NickADobos 2026-06-29 16:41Z

Rick Rubin already wrote the playbook on tech roles Experimenters 1. Slop cannon prototypers 2. Product builders Finishers 3. Sweepers 4. Growers 5. Maintainers

Xiangyi Li @xdotli 2026-06-23 05:37Z

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

FlightTrace @FlightTraceData 2026-06-29 05:24Z

$NVS Novartis flew Basel → San Francisco nonstop tonight — a Sunday evening in late June. That route is rare: just 4 times in 3+ years, every one between January and February — 3 of them in early January (ostensibly) tied to JPM Healthcare Conference in SF. This fleet is

GEOFF @geoffreywoo 2026-06-28 15:10Z

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.

Andrew Solender @AndrewSolender 2026-06-24 03:46Z

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"

산왕 @mount_king_gg 2026-06-24 09:59Z

서울시 보도자료 중 정비사업지역 전자총회 지원 관련 자료인데요.. 보시면 3년 내 착공 가능할 것으로 보이는 정비사업 지역 list가 나와 있습니다 정비사업 입주권에 관심 있으시다면 가능한 아해 지역을 알아보시기를 바랍니다 [재개발사업] - 36곳 (강북 지역) 한남2, 한남3, 갈현1, 증산5,

marek mutwil @LabMutwil 2026-06-24 09:38Z

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

fabiano.sol @FabianoSolana 2026-06-27 09:40Z

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.

gaezv @gaezpeat 2026-06-29 09:22Z

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

sk @compchemm 2026-06-29 06:47Z

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 …

junkbondinvestor @junkbondinvest 2026-06-26 20:44Z

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

Iyan Moon @iyanmoonyang 2026-06-29 07:39Z

first ee project: I put a bitmap running horse on a raspberry pi pico using codex! I couldn’t get the snout right so it looks a bit stubby thank you @covacut for the ingredients

Lauris @lzminsky 2026-06-27 19:12Z

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

th0rgal @Th0rgal_ 2026-06-24 07:59Z

Just reverse engineered my Oura Ring 5 so I can control my computer like a wizzard. @ouraring please send my love to whoever buried a feature to stream live accelerometer data

Techmeme @Techmeme 2026-06-25 13:20Z

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!)