Backscroll ranked feed — 2026-06-27 13:21Z

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.

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

Haara Metsa chandelier by Cameron Design House

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

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

GEOFF @geoffreywoo 2026-06-27 07:30Z

investing take that annoys people: the best deal memo is sometimes one ugly screenshot. queue before noon. queue after model. headcount line gone. put that next to your 47-page market map and watch the pdf start sweating.

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.

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

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

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

Dwarkesh Patel @dwarkesh_sp 2026-06-26 21:56Z

Suppose an AI model fully automates a $200k job. Its economic value - measured in contribution to GDP - will be both far more and far less than $200k. There are two ways of thinking about the model's economic value: (1) It'll be far more than $200k. After all, the human

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.

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

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

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

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

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 18:22Z

btw if u thought ur custom fanned-out data processing multi-agent orchestration harness was a particularly relevant long term moat, sorry, openai just added it as an api parameter

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)

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

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

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

Soren Larson @hypersoren 2026-06-27 00:20Z

Shopify doesn’t screw around

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

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.

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

altra @catboosted 2026-06-26 07:20Z

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

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

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

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

AI slop as straussian infrastructure

Zephyr @zephyr_z9 2026-06-27 01:56Z

Very important paradigm Rn all the computer use AI models are janky & slow Small models and reducing latency (use Groq like hardware) will be super important

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

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.

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

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.

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)

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

USD.AI @USDai_Official 2026-06-27 12:00Z

CMBS made commercial real estate loans investable by bundling mortgages into tradable instruments. USDAI is building a similar credit model for AI infra: bundling GPU-backed loans into a yield-bearing product – $sUSDai .

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

At Amp we all push to main.

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

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

showing stack traces gets you 80% of the way

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"

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.

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

Casio H—1 illustrated in @figma

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)

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

ar0cket1 @ar0cket1 2026-06-27 08:52Z

btw self teachers get noticeably better as you scale model size and pretraining/sft/midtraining. most of literature is at 4B params, I did some quick analysis and as you move towards 200B or even 30B (both MoEs), it gets quite a bit better in terms of supervision (you still

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

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

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/ …

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.

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 …

Chris Tate @ctatedev 2026-06-27 03:18Z

agent-browser 0.31 is out Browser agents now get durable session memory → stable worktree-scoped sessions → auto restore → validation checks → transactional saves and diagnostics

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

rohan anil @_arohan_ 2026-06-25 18:06Z

Very cool work with connections to the schedule free work and iterative averaging work. Interestingly, a natural step is to derive it for vanilla shampoo because it keeps the pre-conditioner around (b2>0) around. Essentially replace diagonal with the kronecker factors Could be

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

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

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

Elliot Lipnowski @ElliotLip 2026-06-27 11:37Z

Today's edition of "theory papers you should know" is Joel Sobel's 8-page gem, "How to Count to One Thousand". At a high level, it's about building in redundancy in completing a complex task. At a low level, it's about how to count a stack of $1 bills to see if you have $1000.

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

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

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

Noah Ryan @NoahRyanCo 2026-06-27 02:34Z

Disconnectivity will be a defining status signal. It already is. Handwritten letters, disposable cameras at weddings, physical business cards, offline maps, cash, board games, DVD collections, house with no Alexa, dumb phones. Technology is ubiquitous and cheap. Refusing tech

Yun-Ta Tsai @yunta_tsai 2026-06-27 06:16Z

It is easier to find a person to train a model than a person who is willing to do janitorial work, even though the former depends on the latter to scale.

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

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

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