1.
WSJ investigates Polymarket's fake-win marketing machine
Polymarket used offshore clippers, dummy websites, and fake-win videos to market an exchange that banned Americans from trading
Backlist — 21 Jun 2026 UTC
Kept the slate broad and mostly non-AI; skipped near-duplicate World Cup chatter and quote-tweet chains unless the underlying item had durable technical, market, or policy value.
2.
High-fidelity scans could make fast medical prediction possible
A slower, higher-quality scan can become the target for a faster scan, turning one hour of imaging into a 10-minute inference problem
3.
South Korea's June semiconductor exports keep surging
DRAM, NAND, SSD, and HBM exports are all ripping higher, showing how quickly memory is re-rating
4.
China tightens indium exports
The export data points to a familiar tightening playbook that can move downstream hardware markets
5.
JEDEC ratifies a new HBM standard
A new memory standard broadens HBM's addressable scope and can shift demand across the stack
6.
Citi sees a $5.5T tokenized-asset market by 2030
Citi's projection puts tokenized equities, Treasuries, and money-market funds in a category large enough to matter for real capital markets
7.
The US cannot absorb Vietnam-scale aircraft losses
If airpower losses were at Vietnam levels, the modern US fleet would not have the capacity to sustain them
8.
Plug-in solar could change rooftop power
A low-cost balcony-solar rule could expand clean energy access for renters and homeowners alike
9.
AI detection tools are fueling witch-hunts
AI detectors are being used to make pseudoscientific accusations feel authoritative
10.
America is getting a new trade school for the reindustrialization boom (x.com)
A Texas training factory aims to graduate 10,000 skilled workers a year and route them into employer-backed production lines
11.
Building value is not enough; it has to be legible
Creating something useful and explaining it in a way buyers can understand are two different skills
12.
The real skill is spotting which ambiguities matter
Success often comes from narrowing the few uncertainties that actually move outcomes while ignoring the rest
13.
Verification tech erodes honor
When performance becomes easier to measure, trust and honor stop carrying the same weight they once did
14.
LLMs learn the shadow of the world through language
Language models absorb an implicit world model from text, but humans bootstrap theirs from a multimodal world first
15.
Why PPO beats GRPO on long-horizon tasks
Group-level synchronization makes long-horizon training harder, which changes the variance and staleness profile of the method
16.
A personal agent runtime you can run in your own Cloudflare account
The stack includes durable conversations, MCP tools, recurring jobs, a sandbox workspace, browser runs, and voice in one deployable runtime
17.
Agent wallets need caps, approvals, and audit logs
A serious agent payment system becomes interesting only when policy limits, multi-step approvals, and full action logs are built in
18.
Voice agents get interesting when they can see the screen
Screen context turns voice agents from novelty chat into a real workflow interface
19.
Computer-use agents create a richer interaction channel
The most interesting part of computer-use agents is not automation alone but the much richer communication channel they open up
20.
A Linux RDMA bug worth blogging about (x.com)
The issue is a new cousin of CopyFail that abuses RDMA and deserves a full write-up
21.
Frozen image encoders are hiding physics in video benchmarks (x.com)
A straightening hypothesis pushed video benchmarks to new SOTA without video training or physics supervision
22.
NVFP4 GEMM on NVIDIA's latest GPUs
The post walks through converting a BF16 GEMM example into NVFP4 on Sm12x hardware
23.
An open-source extraction model pulls data out of messy contracts
The model can turn a 26-page contract into structured data, which is the kind of boring breakthrough that actually ships
24.
WWII paper drop tanks were a clever range hack
Allied fighters used glue-soaked kraft-paper tanks to extend range and then dropped them to save metal
25.
Medieval crucifixes with moving arms and bleeding mechanisms
A forgotten class of devotional objects built motion and even bleeding effects into crucifixes
26.
Solving an 8-year-old 0.5 ETH cryptographic puzzle
The answer was hidden in plain sight and opens the door to three more unsolved prizes
27.
A game where you draw Aomori Prefecture
The score depends on how your outline affects apples, mackerel, satisfaction, and access to Hokkaido
28.
A tiny wallpaper generator for new parents
A lock-screen idea turned into a web app where a child's name and birthday generate a wallpaper that changes during the day
29.
Adobe's crash loop is the market talking
The point is not just that Adobe is buggy; it's that the underlying design and video markets remain huge while the incumbent feels weaker by the month
30.
Temperature check on AI recruiting, here's an n of 1: a friend of mine just left OpenAI ($2M/year total comp) to …
Temperature check on AI recruiting, here's an n of 1: a friend of mine just left OpenAI ($2M/year total comp) to go to Google ($8M/year total comp, cash/RSUs). They joined OpenAI well past ChatGPT launch and are research adjacent, but mainl
31.
Damn, I lost my mind. This was the fucking hardest RE I've ever done in my life. First live result. Nokia 6600 bo…
Damn, I lost my mind. This was the fucking hardest RE I've ever done in my life. First live result. Nokia 6600 boots from a firmware image pulled from the internet. Bare metal emulation + patches and stubs. This is far from over. But fuck,
32.
here's how I got 80% increase in ARPU (or avg proceed per user)
here's how I got 80% increase in ARPU (or avg proceed per user) > remove your monthly sub option i was very hesitant because monthly was my workhorse with max share of MRR. claude + codex both suggested against removing monthly. but it
33.
Day 1: Opus 4.8 writes a careless code comment that says something cannot be done
Day 1: Opus 4.8 writes a careless code comment that says something cannot be done Day 2: Opus 4.8 reads the code comment and treats "the developer" as authority for what can and can't be done
34.
pancancer SAE atlas of pathology FM! the denoising result is very neat. zero the tissue-fold concepts at inferenc…
pancancer SAE atlas of pathology FM! the denoising result is very neat. zero the tissue-fold concepts at inference and classifier sensitivity jumps 0.61→0.95, no retraining!! subtractive SAE steering fixing a real artifact so clean
35.
This is the primary source of prediction market nihilism.
This is the primary source of prediction market nihilism. Two years ago, prediction market volume was > 90% Politics + Economics. Today, those same categories are under 6% of total volume. The onus is on venues to create novel products t
36.
on linkedIn recruiter rn. using ‘open to work’ filter and seeing lots and lots of people from glean :(
on linkedIn recruiter rn. using ‘open to work’ filter and seeing lots and lots of people from glean :(
37.
You can just rent Blackwell hours at 90% discount of inference API prices.
You can just rent Blackwell hours at 90% discount of inference API prices. Model providers are playing you all like fiddles.
38.
The community still hasn’t fully absorbed that you can train a 30B MoE on 10T tokens, from scratch, for just 500K
The community still hasn’t fully absorbed that you can train a 30B MoE on 10T tokens, from scratch, for just 500K The big caveat is you need to get data, infra etc right from day 1. Once this ecosystem of enablers is working in unison, own
39.
SOMEONE MADE AN APP THAT LETS YOU COLLECT CATS YOU SEE ON THE STREET LIKE POKÉMON, AND THE COOLEST PART IS MAKING…
SOMEONE MADE AN APP THAT LETS YOU COLLECT CATS YOU SEE ON THE STREET LIKE POKÉMON, AND THE COOLEST PART IS MAKING CHEATING IMPOSSIBLE. the logic is simple. you spot a cat on the street, you open the camera, you snap its photo. it gets adde
40.
Hey (x.com)
Hey @bcherny ! Claude Code Web does not support team-shared environment variables today. This means that it's really tough to securely share secrets across a team -- this feature is supported by other cloud agent systems like @cursor_ai
41.
My conversation with (x.com)
My conversation with @credistick touched upon this. Over-funded venture capital firms focus on raising more money from LPs. Thus, their portcos become ads pointed at LPs. The increase in tech marketing slop is the byproduct of distorted
42.
NYC creates so much value that a whole lot of little mafias skim off like 30% of it and it’s still a good deal to…
NYC creates so much value that a whole lot of little mafias skim off like 30% of it and it’s still a good deal to live here.
43.
the most important problem in ai safety, as well as the biggest unlock for letting RSI fucking rip, is formalizin…
the most important problem in ai safety, as well as the biggest unlock for letting RSI fucking rip, is formalizing and automatizing the science of robust model behavior evaluation
44.
Handsome male Linnet on a stone wall.
Handsome male Linnet on a stone wall. Taken last weekend at Martin's Haven in Pembrokeshire.
45.
On-chain trackers thought SpaceX held 8,000 Bitcoin.
On-chain trackers thought SpaceX held 8,000 Bitcoin. The SEC filing said 18,712. Nobody was watching. The largest IPO in history quietly doubled the known position.
46.
I've been doing MEV for nearly 4 years now, and I've never touched sandwich strategies. There are three reasons f…
I've been doing MEV for nearly 4 years now, and I've never touched sandwich strategies. There are three reasons for that: First, profit margins converge very quickly. Mature sandwich strategies rapidly converge to the point where only extr
47.
People talk about Deep Work, but for me it's about Deep Unwork.
People talk about Deep Work, but for me it's about Deep Unwork. I can work fine and context shift in 5-minute chunks, for most purposes, and interruptions are fine. But to properly relax, that's when I need hours of defended uninterrupted
48.
While many problems can be researched on a small scale, native long-context isn't one of them. Tackling it with l…
While many problems can be researched on a small scale, native long-context isn't one of them. Tackling it with limited compute and small datasets just gives you false signals about what actually works. You need scaling experiments and infr
49.
App Store search stopped reading your keyword field like a word list this June.
App Store search stopped reading your keyword field like a word list this June. It reads intent now. Natural language, not exact match. Half the indie apps I audit are still stuffing 100 characters of disconnected keywords and wondering wh
50.
A picture drawn just for the visible parts
A picture drawn just for the visible parts ←Rough sketch →Finished
51.
depth 10 achieved with a hand crafted policy.
depth 10 achieved with a hand crafted policy. codex's /goal is pretty good for these kinds of things. pretty overfit though (median depth 5)
52.
A few before and after shots from my run-time-absolutely-fake GI baker for fully random generated procedural worl…
A few before and after shots from my run-time-absolutely-fake GI baker for fully random generated procedural worlds for my game Drivers of the Apocalypse (wishlist on steam ) I made it for the cars, but it looks absolutely stunning in the
53.
Kurzweil mixed bold but solid predictions (given enough compute, a neocognitron like architecture can achieve AGI…
Kurzweil mixed bold but solid predictions (given enough compute, a neocognitron like architecture can achieve AGI before 2030) with dogshit (digitizing the connectome is a near term way to human immortality). But both looked like the same k
54.
Most orgs aren’t set up for the Builder PM archetype because their rituals demand doc sacrifice at the altar of p…
Most orgs aren’t set up for the Builder PM archetype because their rituals demand doc sacrifice at the altar of performance appraisals and exec reviews. They will start losing the PMs who value the freedom that AI design and SWE agents giv
55.
a fully interactive "recovery lab" where you simulate different kinds of agent failures/recovery
a fully interactive "recovery lab" where you simulate different kinds of agent failures/recovery built with opencode + glm 5.2 (176k tokens, 3.50$), 1 shot after I fed it my 4k word slop analysis and the agents sdk repo pretty fkin cool.
56.
Dana White says Trevor Wittman’s team wanted $100 million for his glove design:
Dana White says Trevor Wittman’s team wanted $100 million for his glove design: “I literally told my finance team ‘get this deal done, I don’t care what it takes’. They wanted like $100 million. You know how many f*cking gloves you gotta
57.
Polymarket created a clone website so creators could fake winning trades, according to WSJ.
Polymarket created a clone website so creators could fake winning trades, according to WSJ. It also used an agency to promote videos showing how easily insider trading could be carried out on the platform.
58.
South Korea Preliminary Exports (June 1–20)
South Korea Preliminary Exports (June 1–20) - MLCC: +24% YoY, +23% MoM - Semiconductor Manufacturing Equipment: +38% YoY, +43% MoM
59.
South Korea Preliminary Exports (June 1–20)
South Korea Preliminary Exports (June 1–20) - DRAM (including modules): +342% YoY, +3% MoM - Flash Memory (NAND): +336% YoY, +28% MoM - SSD: +405% YoY, +25% MoM - MCP (HBM): +209% YoY, +51% MoM
60.
Something a little different! I designed these sleeping bunks that you see throughout outpost
Something a little different! I designed these sleeping bunks that you see throughout outpost The big idea I wanted to get across is modularity and space efficiency. Were unused beds could be stowed, and that space could be used for somet
61.
This is a wild victory lap.
This is a wild victory lap. Step 1 - essentially legalize all drugs - overdose deaths skyrocket. Step 2 - Criminalize illicit drugs again - overdose deaths drop by half. Step 3 - take credit only for step 2.
62.
This post has antithesis, triplets, em-dashes, and single word sentences for emphasis (I. e., all the telltale si…
This post has antithesis, triplets, em-dashes, and single word sentences for emphasis (I. e., all the telltale signs of AI writing), but read it and you can tell this was written by a human. What are the features of this text that make it
63.
Hot take: The strong redundancy and the compositionality of images make them navigable and as such wonderful memo…
Hot take: The strong redundancy and the compositionality of images make them navigable and as such wonderful memory banks. Language developed on top of them and had no pressure to have such properties, which results in enormous challenges
64.
It’d just be me saying “how can we simplify this? Make it clearer?” over and over
It’d just be me saying “how can we simplify this? Make it clearer?” over and over This is also why it’s valuable to code and maintain things; you’re basically always just asking the question “How do I make future Josh appreciate his deci
65.
Korean media reported that CXMT employees who received equity incentives ahead of the company’s listing could end…
Korean media reported that CXMT employees who received equity incentives ahead of the company’s listing could end up holding shares worth hundreds of thousands of dollars per person. This is because CXMT granted large-scale equity awards i
66.
cliche, but big part of why i exited my memory longs from last year is bc the dumbest top buyers i know are sudde…
cliche, but big part of why i exited my memory longs from last year is bc the dumbest top buyers i know are suddenly bullish memory rn
67.
Asked a contractor pal about the $13M paint job peeling off the Lincoln Memorial Reflecting Pool.
Asked a contractor pal about the $13M paint job peeling off the Lincoln Memorial Reflecting Pool. Him: I knew that'd happen. Me: How? Him: You can't rush that kind of job. The pool bottom has to be really dry or it'll exude water. And you
68.
Apparently it is possible to go much faster than my 13.5 t/s, and this is a very good news. It means 4-bit GLM 5.…
Apparently it is possible to go much faster than my 13.5 t/s, and this is a very good news. It means 4-bit GLM 5.2 will be usable on Mac Studios M3 Ultras. The problem is that it is literally hardware you *can not buy* anymore.
69.
Saylor knows he's the only bid in the book, so he cannot stop buying or bitcoin will roll over and die, taking hi…
Saylor knows he's the only bid in the book, so he cannot stop buying or bitcoin will roll over and die, taking him with it But the market has sniffed out his weakness (see: STRC discount), i doubt he can raise long dated debt at rates whic
70.
you def need a cash cow attached or ecosystem enabling frequent and big rounds for frontier models…
you def need a cash cow attached or ecosystem enabling frequent and big rounds for frontier models… either of the two. openai, anthropic, xai, deepseek,zai all have one mistral, sarvam doesn’t seem to have one yet.
71.
I know robotics teams that have been told stop using Claude code because the bugs are impossible to find because …
I know robotics teams that have been told stop using Claude code because the bugs are impossible to find because the vibecoded slop is unreadable
72.
So far, any "understanding of the world" a LLM has is an implicit by-product of its modeling of language.
So far, any "understanding of the world" a LLM has is an implicit by-product of its modeling of language. The closest in a human mind is probably our model of abstract objects that have no reality but their textual description (mathematic
73.
The "California conferences" realize best conferences are like a good dinner party: you curate a mix of folks who…
The "California conferences" realize best conferences are like a good dinner party: you curate a mix of folks who know each other, famous outsiders with interesting views, and little known or younger folks doing cool things. Mostly small gr
74.
We don't use mood boards.
We don't use mood boards. Most studios start with a curated canvas full of things other people designed. We start with concept directions built from discovery, all tied back to what the business is trying to do. The concept is the source
75.
SNS 에서 잘못 알려진 결혼 사진
SNS 에서 잘못 알려진 결혼 사진 1. 인터넷에서 말없이 딱 이장면만 나가서 이슬람 조혼문화라고 악의적으로 나온 장면임 (첫번째 사진) 2. 근데 실제론 걍 잘 놀아주던 삼촌과 조카였고 삼촌이 결혼한다고 떠나는걸 서러워하며 삼촌에게 안기러 간 장면이라고 함 3. 실제 부인사진 (세번 째 사진)
76.
LASTEST PROJECTION: 85.7% chance Abelardo wins Colombia's runoff tomorrow.
LASTEST PROJECTION: 85.7% chance Abelardo wins Colombia's runoff tomorrow. That's the easy part. The interesting question is the margin, and the distribution doesn't look like the market thinks.
77.
Living in New York is like being Roger in Mad Men, carrying around too much cash and constantly being hit up for it
Living in New York is like being Roger in Mad Men, carrying around too much cash and constantly being hit up for it
78.
The pool of informed critics of artificial intelligence is very small in part because subscription costs mean tha…
The pool of informed critics of artificial intelligence is very small in part because subscription costs mean that most otherwise smart critics by nature get limited exposure to even widely available frontier models. Their worldview represe
79.
I have a less optimistic view of this. The core of philanthropic work is to *be liked by rich people*. It is abso…
I have a less optimistic view of this. The core of philanthropic work is to *be liked by rich people*. It is absolutely not *what to do*, which will be done by AI. This is bad news for those who want to do great things.
80.
after I saw (x.com)
after I saw @midjourney medical launch, I really wanted to replicate the visual they use on their website now live with cool modifications on http:// arlan.me/vault/midjourn ey …
81.
400k+ views in 2 days for a side project I almost didn't post. Here's the 3D map of the Bay Area that blew up ove…
400k+ views in 2 days for a side project I almost didn't post. Here's the 3D map of the Bay Area that blew up over the past couple of days. The fun really has been in all the details, the planes flying into and out of SFO, those boats and
82.
One way to think about meditation is as a way of increasing legibility into our perceptual architecture: the deep…
One way to think about meditation is as a way of increasing legibility into our perceptual architecture: the deeper we practice, the further back we go. If we go back far enough, it stops being "our" perceptual architecture at all.
83.
Added 5,000 more items to Wiki Spy! A lot more vintage cars, pinball machines, and puppies
Added 5,000 more items to Wiki Spy! A lot more vintage cars, pinball machines, and puppies
84.
"year of the horse" (x.com)
"year of the horse" i heard bojack horseman ( @floraai is free right now, so i mixed tarot card style with bojack and made a tiny deck )
85.
I hoard X bookmarks and never read them. So I built an extension that injects a bookmarked post into my main feed…
I hoard X bookmarks and never read them. So I built an extension that injects a bookmarked post into my main feed every time I open X (almost like an ad). Now I read my bookmarks. The trick was hijacking real estate I already check 50 tim
86.
Too much clickbait … let’s ground some nuance here
Too much clickbait … let’s ground some nuance here 1/ Start with the NSA part, because it’s backwards from what people think. The agency isn’t the target. It’s the buyer. Anthropic engineers reportedly embedded on site to run it 2/ The g
87.
Strategy is still showing
Strategy is still showing $STRC as 0.43 Sharpe despite the fact that it has negative sharpe
88.
The tentative plan: a command that integrates bun’s bundler with a manifest + plugin system to a deployable artif…
The tentative plan: a command that integrates bun’s bundler with a manifest + plugin system to a deployable artifact for multiple hosts (Vercel, Railway, Cloudflare, S3/Cloudfront, etc) so that you can do `bun deploy` and it just works whet
89.
Batch 8 graduated! (x.com)
Batch 8 graduated! But it was also @lossfunk last batch. We're pausing the residency to double down on our own research agenda.
90.
We all hoped for a unified agent each that could do all our work across various platforms. Instead, every product…
We all hoped for a unified agent each that could do all our work across various platforms. Instead, every product just has its own shit version of Siri.
How it ranks: Backlist reads my Twitter/X timeline, scores every tweet for substance with an LLM rubric (not engagement), and publishes the daily top picks with a one-line takeaway. Curated by Surya Dantuluri.