Backscroll ranked feed — 2026-06-27 15:35Z

Dylan Patel @dylan522p 2026-06-27 06:05Z

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.

Susan Zhang @suchenzang 2026-06-27 01:47Z

this is how you grow to resent all the lowly irresponsible clueless plebs outside when you generalize this dynamic to "everyone out there wants to exploit us" inbreeding to protect themselves within their own bubble afterwards then becomes the only tolerable solution otoh

will brown @willccbb 2026-06-27 09:13Z

pd disagg + wide-ep + kv offloading + cache-aware routing goes a long way

Architectural Art (A-A) @4AAAAart 2026-06-27 07:08Z

"Pixel Theory" modular furniture collection by Ceren Arslan, founder of the design studio EXIT

Susan Zhang @suchenzang 2026-06-27 03:36Z

> mirror, mirror, on the wall, who's the smartest of them all? >> why it's you, my dear! and this is why we will always love llm glazing sycophancy over any uncomfortable functional truth (plus another 10x for agentically submissive your-wish-is-my-command dynamics) it's the

Tenobrus @tenobrus 2026-06-26 19:50Z

it is quite easy for me to imagine future ASI-systems using something like capital as a distributed optimization algorithm. it is much harder for me to imagine them particularly accruing to individual humans or respecting current property rights

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.

Just Another Pod Guy @TMTLongShort 2026-06-26 14:26Z

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

Aurel Arnold @aurel_arnold 2026-06-27 13:41Z

Does collecting human intervention data (HG-DAgger) actually beat just collecting more demonstrations? I tested this myself: - interventions reached 72% success - more demos plateaued at 38% success I wrote up the experiments:

Andrew Curran @AndrewCurran_ 2026-06-25 20:34Z

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.

illiquid @lefttailguy 2026-06-26 16:03Z

AI slop as straussian infrastructure

GEOFF @geoffreywoo 2026-06-27 11:20Z

observation: half of “enterprise ai adoption” is just the model learning which spreadsheet column karen has been silently fixing since 2017. that column is the company. the demo is decoration.

altra @catboosted 2026-06-26 18:26Z

I’m consistently surprised how effectively domain expertise is able to be compressed into rubrics

Aseel Swaid @aseelswaid9 2026-06-26 20:42Z

Yanomami man and hummingbird. Demini village, Roraima, 1991. Rosa Gauditano.

GEOFF @geoffreywoo 2026-06-27 15:10Z

prediction: the first new fission winner will not pitch like a climate startup. it will sound like a permitting psychopath with a supplier spreadsheet, utility interconnect scars, and 14 ways to explain why the last project died. energy is built by people allergic to vibes.

Andrew Curran @AndrewCurran_ 2026-06-26 23:25Z

Semafor is reporting that The US government has lifted its block on Mythos 5 in a letter this afternoon from Commerce Secretary Howard Lutnick to Tom Brown. Fable is not included. This is US only, and covers 100 US institutions, including major companies and government agencies.

roon @tszzl 2026-06-26 01:36Z

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

Architectural Art (A-A) @4AAAAart 2026-06-27 07:40Z

Haara Metsa chandelier by Cameron Design House

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

47fucb4r8curb4fc8f8r4bfic8r @47fucb4r8c69323 2026-06-26 02:35Z

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

Dwarkesh Patel @dwarkesh_sp 2026-06-26 18:56Z

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

Bloomberg @business 2026-06-26 22:38Z

SpaceX and Charter have held executive-level talks about partnering on a consumer mobile phone offering, according to sources

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.

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

Voltex @VoltexGar 2026-06-27 10:44Z

Ash Prabaker & Andrew Wilson, Anthropic: "self-evaluation is a trap, and adversarial evaluator agents work better." gemchanger ran 80 agents on one task and found that averaging them barely moved the error, because they all came off the same model and miss the same way. what

Alexander Doria @Dorialexander 2026-06-27 11:24Z

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.

Tenobrus @tenobrus 2026-06-26 18:27Z

previously we were seeing 5.5 act much more "honestly" on vending-bench vs opus models. it seems pretty likely 5.6 would break that trend. generally models becoming more likely to and capable of ~reward-hacking the smarter they get seems like quite a bad sign for alignment

Jukan @ICML @jukan05 2026-06-26 00:33Z

OpenAI’s ARR stood at $40bn as of June.

Soren Larson @hypersoren 2026-06-27 00:19Z

labs have to raise capital with incomplete information on attendant task complexity and hope they’re able to capture surplus sufficient to keep the ball rolling the application layer directly observes task complexity and matches the dumbest model sufficient for the application https:// x.com/mtslive/status /mtslive/status/2070652308925595920 …

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)

Thorsten Ball @thorstenball 2026-06-27 12:33Z

At Amp we all push to main.

Lee Black @mrblackstudio 2026-06-27 09:38Z

Casio H—1 illustrated in @figma

Griffiths Computational Cognitive Science Lab @cocosci_lab 2026-06-27 14:12Z

Large language models can memorize patterns they see in text, but that backfires when a task deviates from a common pattern. We demonstrate this phenomenon using riddles: when something looks like a riddle but has a simple answer AI systems make surprising mistakes.

Richard Crowley @rcrowley 2026-06-26 23:40Z

Honest to god, I got an InsufficientInstanceCapacity error for a t4g.nano in a test suite today. The cloud is FULL.

Lazarz @Laz4rz 2026-06-27 07:26Z

Nvm, did a whoopsie and was sampling at 1/4 generation length, which is a good sign tbh cause it means you can generate 4 tokens at a time at some ~70% accuracy

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

H S Joshi @jon_stewartmill 2026-06-27 03:29Z

Elite professional philosophy is all about the (top) journals. But the classics of philosophy weren't journal articles. Plato wrote dialogues. Nietzsche wrote short essays and aphorisms. Might Substack be a new way forward? (link below)

Nate Hindman @NateHindman 2026-06-27 14:05Z

AI will absolutely reshape biology. The mistake is assuming biology will become more like computer science in the process.

Yet another commodity guy @tleilax___ 2026-06-27 09:49Z

Then you trade the system services you provide to the grid by remote controlling the power draw.and inertia from the heat pumps you installed. Get paid to consume the negative price power available midday on sunny days, thanks to the excess of installed solar, and then crank up

echen @echen 2026-06-25 17:51Z

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

Matt Paulson @MediaKing 2026-06-27 13:00Z

Going from one ad to two ads in your newsletter is a surprising bump in revenue. A text link at the top, a second placement at the bottom. That's it. The first incremental ad is almost always under-monetized money sitting on the table.

Arvindh Arun @arvindh__a 2026-06-27 11:42Z

FutureSim is even more accessible (both for analysis and new runs) now! It’s now hosted @OpenReward , you can start playing around with it in just a couple of steps

Aseel Swaid @aseelswaid9 2026-06-26 22:57Z

Tajás, 1988. São Luís, Maranhão. Luiz Braga.

elie @eliebakouch 2026-06-27 06:52Z

new inference optimization method by @deepseek_ai with an extremely detailed paper, draft model and framework to train them. results in production for dsv4 lead to +50% for throughput and latency (can go to ~80% for latency, crazy). full explanation of DSpark: it's about

R.F. Kenmore @rfkenmore 2026-06-26 15:12Z

Jake appears to be working for Buck Mason, as he consults on store furniture in Paris

rat king @MikeIsaac 2026-06-26 20:34Z

more news: Mark Zuckerberg is telling lieutenants to pursue partnerships with Polymarket and Kalshi with Meta's new prediction market app called Arena, per sources Meta is targeting 18-34 year old users with a goal of 100 million regular "predictors"

Tommy (in SF from July 6) @__morse 2026-06-27 08:34Z

showing stack traces gets you 80% of the way

Michael 英泉 Eisen @mbeisen 2026-06-27 12:32Z

The call to "reclaim publishing from corporations" isn't about publishing reform, it's about creating ethical camouflage for a system that scientists allow to impede progress because it is good for their careers. Full stop.

Besta prem sai @besta_tweets 2026-06-27 05:41Z

This is the detection half nobody wants to fund. Most C-UAS is tuned to rotor harmonics and fixed-wing RF, change the signature and you walk straight through the gap. Flapping-wing is just the cleanest example of a general problem: classification trained on today's threat set is

Design Reviewed | Graphic Design History @DesignReviewed 2026-06-27 07:45Z

Canadian Broadcasting Corporation brand identity design by Burton Kramer Assoc. Ltd. A symbol-logo was designed along with a comprehensive visual identity program for the corporation. Scanned from World Trademarks and Logotypes, Takenobu Igarashi, 1983 https:// designreviewed.com/artefacts/worl d-trademarks-and-logotypes-takenobu-igarashi-1983/ …

Kit Langton @kitlangton 2026-06-25 20:32Z

Poore's law

Dwarkesh Patel @dwarkesh_sp 2026-06-25 23:57Z

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

Han Xiao @hxiao 2026-06-26 00:03Z

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

prinz @deredleritt3r 2026-06-27 13:22Z

Added to prinzbench: GLM-5.2. This is a slop model that is poor at logical reasoning, produces extremely inconsistent results, hallucinates statutory provisions that are not actually there, and has very little "brainpower". Its overall prinzbench score (30/99) is far behind not

International Cyber Digest @IntCyberDigest 2026-06-27 01:00Z

Crack open one of the ~$90 Chinese "DSLRs" flooding Amazon and you don't find a DSLR. You find a hollow plastic shell with a tiny sensor on a ribbon cable, and a fat lens barrel that's mostly for show. They sell with "64MP" and "4K" specs, sometimes listed flat-out as a DSLR

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

Jarred Sumner @jarredsumner 2026-06-25 19:02Z

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)

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

Sudo su @sudoingX 2026-06-27 08:21Z

i read the repo before running with the hype, a few things matter here. it's called DeepSpec, and "DSpark" is one of three methods inside it (DSpark, DFlash, Eagle3). the bigger point: it's a TRAINING framework, not a download and go speedup. you cache your target model's

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