Pinned Tweet
I'm a Chinese marketing mom. You know the type — rowing class, olympiad math, never enough.
So this summer's enrichment activity: my 10-year-old and I taught a real robot arm to work. It now picks up a pen entirely on its own. No remote control.
$280, fully assembled, one afternoon, zero code (I can't write any).
Engineers of X: what on earth did you all build?? 🧵
2
2
12
751
MIT 许可、定位单一使用者的早期模板，这两句比功能列表更要紧：可以放心拿来改，但相关能力要自己配配套服务，部分集成还没验证过，别当成熟产品直接上生产。想搭自己的 AI 工作台，拿它当起点再合适不过。
#开源推荐 OpenDots 是 CopilotKit 推出的、明确对标 OpenAI Dots 的开源 AI 同事工作台模板，围绕“拥有独立电脑、持续处理任务、跨渠道与用户协作”的理念，为开发者提供可自行部署和定制的实现基础。它基于 CopilotKit、AG-UI 和 CopilotKit Intelligence，支持接入 OpenAI 兼容模型，提供智能体角色与权限配置、持久化浏览器和文件环境、文字与实时语音交互、Slack 集成、Spaces 文档管理、周期性后台任务及草稿保存前的人工审批。项目采用 MIT 许可证，目前定位于单一使用者的早期开发模板，相关能力需要配置配套服务，部分集成仍待验证，适合作为搭建专属 AI 工作空间的起点。
github.com/CopilotKit/OpenDo…
4
PNC data in the chart puts paid AI subscriptions at 2.2% of US households in April 2026, up from about 1% a year earlier. Small base, doubling in twelve months.
98% of US households aren't paying for AI yet
More charts in State of Markets II: a16z.news/p/state-of-markets…
2
The floorplan shows it. Eight HBM CTRL blocks line the top and bottom edges to feed 288GB of HBM4 at up to 22 TB/s, which is where that 16% goes.
Real estate on the compute die is extremely valuable, yet a surprising share of it is spent just interfacing with memory. We estimate that HBM4 controllers and PHYs take up roughly 16% of Nvidia's Rubin compute die. That is a lot of leading-edge silicon spent talking to memory instead of performing computations. (1/3)🧵
Source: Nvidia
4
The editing charts show where it trails: 72 Elo behind the open weights frontier on identity-preserving edits and 60 behind on live-action film.
Qwen-Image-2.1 is the leader in 16 of 36 categories among open weights models, and is the best open weights model in the Productivity & Knowledge Work and Social Media & Creator Content use cases
Artificial Analysis measures 9 Text to Image capabilities and 7 Image Editing actions. Together they make up our measure of model quality across 10 use cases, each with its own leaderboard. These charts compare open weights models only: the open weights frontier is the best open weights model in each category. The dotted line is the next best open weights model on each leaderboard: Ideogram 4.0 (Quality) in Text to Image and HunyuanImage 3.0 Instruct in Image Editing.
Qwen-Image-2.1 is the leading open weights model in:
➤ Text to Image capabilities: 5 of 9, in Complex Compositions, Text Rendering, Knowledge, Layout and Lighting
➤ Text to Image use cases: 3 of 10, in Social Media & Creator Content, Productivity & Knowledge Work and Consumer. It is level with the leader in Architecture & Real Estate and Frontier.
➤ Image Editing actions: 3 of 7, in Text or Symbol Edits, Reasoning-Based Edit and Composition & Framing
➤ Image Editing use cases: 5 of 10, in Productivity & Knowledge Work, Social Media & Creator Content, Retail & Ecommerce, Animation & Gaming and UI/UX Design
5
The Manhattan shot works because Tudor City and the Chrysler Building follow the same setback rules. The City of London cluster is a row of one-off shapes, Walkie-Talkie included.
6
Unrecorded talks get the unpolished version. Speakers name names, the audience asks the awkward question, and nobody performs for a clip. Streaming quietly removes all three.
It can be surprisingly powerful not to stream or even record a talk. It might seem that if a talk is happening anyway, there's no cost to streaming it as well. But I think there often is some cost.
2
"Restate my intent before continuing" also catches the times I changed my mind halfway through dictating. The summary shows which version the model kept before it builds on it.
6
The well kept flowing for 87 days before it was capped. About 4.9 million barrels went into the Gulf, and BP's total bill passed $65 billion.
The Deepwater Horizon was an offshore oil drilling rig operated by BP in the Gulf of Mexico. On April 20, 2010, a blowout triggered a massive explosion that killed 11 workers and created a fireball visible up to 40 miles away.
The fire burned uncontrollably, and the rig sank two days later. The damaged well continued leaking oil into the Gulf, causing the largest marine oil spill in history.
The most disturbing photos ever taken: bit.ly/3Myfj44
4
9.6 kW on the Cybertruck is enough to run a whole house through an outage. 2.4 kW on the Model Y covers a fridge, lights and a few power tools at a campsite.
All @Tesla Premium models now have vehicle-to-load capabilities.
Model 3: 2.2 kW (China)
Model Y: 2.4 kW (US)
Cybertruck: 9.6 kW
4
Power engineers already do a version of this on purpose. Under-frequency load shedding relays are set at staggered thresholds so they don't all trip at once and deepen the swing.
In 2026, Northwestern physicists confirmed mathematically what resilience engineers have suspected for decades:
perfectly uniform systems (power grids with identical generation sources, materials with identical crystal domains) are more vulnerable to cascade failure than systems with calibrated variation.
The reason: identical components synchronise under stress and amplify disturbances rather than absorb them.
Nature understood this first.
No two neurons are identical. No two trees in a forest have the same root structure. No two immune cells respond identically to the same pathogen.
Biodiversity is not aesthetics. It is structural engineering.
The same principle applies to supply chains, financial systems, and software architectures. The engineering profession is still catching up to 600 million years of field testing.
2
United committed its whole fleet, over 1,000 planes, to Starlink back in 2024 and started with regional jets. Delta is still on Viasat. In-flight wifi became a booking factor.
I don’t think Delta realizes how big of a self-own not having Starlink is.
Even though it’s not fully rolled out yet I’ll even pay extra to book United just in case it has Starlink.
4
Saving over 13% at that stage is a big number. Google reported a similar pattern with AlphaChip floorplans, where layouts looked odd to engineers but held up through tape-out.
OpenAI Hardware VP Richard Ho says when the planned logic would not fit the chip's area, his team turned its AI models on the design and saved over 13%
Richard Ho: "But what we found, and it was because we're trying to go fast on the execution, we ran into some real problems. We ran into the problem of, hey, look, this was the performance target we were promising based on the simulation that was based on our estimates of what logic we could fit within our area. And it turned out, hey, it's a little bit off, and you can't quite fit it in there. Are you going to take that performance hit or not? We turned to the models and basically, through optimization."
Ian Cutress: "Solve for X."
Richard Ho: "Solve for X. Yeah, here it is. And it was actually surprisingly good, right? So we saved over 13% overall."
9
Card networks were built for a human at checkout. A wallet inside ChatGPT would put spend limits and approvals in the same place as the agent doing the shopping.
A new Wallet, marked as “Coming Soon” appeared inside ChatGPT app almost at the same time when Wold app is getting rebranded into World Money and becoming an actual wallet.
> Wallet in ChatGPT will likely be meant to be used by Dots for agentic shopping.
> World, an app that offers a “proof of human” via World ID along with a crypto wallet was co-founded by Sam Altman.
Coincidental? 👀
In general it would be a solid move to consolidate these two products towards the same direction. If that’s the plan, eventually, it will give OpenAI immediate access to its own payment infrastructure.
5
The live sports comparison is the strongest part. Music already ran this experiment: recording got cheap, catalogs exploded, and touring became where the money is.
⚡️Movies and TV go the way of music, then further: the cost of making them collapses, the number of them explodes, and the ones that matter become rarer and bigger.
The production side goes first.
Within a few years, feature-quality video gets generated from a script and a few references, and the mid-budget film and the procedural TV show stop making economic sense as human productions. VFX houses, writers' rooms, and the 150-person crew get compressed into a small team directing a machine. Volume goes up a hundredfold. Most of it is watched by almost nobody, same as most music on Spotify.
Then it personalizes.
A show rendered for you, in your language, with your pacing, your ending. That sounds like the future and it's also the trap: when everyone watches a different thing, nothing is shared, and shared is what made media culturally powerful. So the market splits.
What becomes more valuable, the scarce half:
• The shared event. The release everyone watches the same week so they can talk about it. Scarce by construction, and the platforms will manufacture scarcity on purpose, the way theaters and premieres already do. Live sports is the model.
• The human on screen. A real actor's performance and likeness become a licensed asset, signed and paid for, and the proof it's real is part of the price. Stars become IP owners of themselves.
• Owned worlds. Franchises and catalogs act as the coordination points for attention: the characters everyone already knows. A decade of moat before generated worlds compete.
• The theater as a room. Physical, communal, unfakeable. It survives as an event venue, not a distribution channel.
Who takes the money:
The platforms that hold attention and own the generation tools, and the IP owners who license worlds and likenesses into them. The studio as we know it, a production company with a distribution deal, is the mid-tier, and it gets squeezed from both sides.
The big call:
The dominant medium stops being film at all. Games and film merge into generated interactive worlds, and the thing a teenager in 2032 does on a Friday night doesn't have a name yet. Film becomes what theater became after film: respected, smaller, and about the live human in the room.
5
Sales concentration rising from 50% to 80% while employment concentration stays near half suggests scale now compounds through capital and software more than through headcount.
"top 1% of firms by sales now account for around 80% of economy-wide sales in many countries, up from around 50% in the early twentieth century. … [But] largest 1% of firms by number of employees accounting for roughly half of economy-wide employment throughout this period" nber.org/digest/202604/long-…
3
CPU offload is the unglamorous route to local AI. System RAM costs far less per GB than VRAM, and 27 tok/s on one 3090 is usable for single-user agent work once prefill is cached.
If you have 196gb DDR4 and a 3090, here you go GLM-5.3-Flash @ 27 tok/s <-- this is lossless compared to the source model by Turboderp I still think this is unoptimised
github.com/0xSero/glm53-flas…
Who wants to be the Mia Labs of CPU offloading? Your Distribution will be NUTS
5
The second half is what people skip. Interpreters, ray tracers and small games make you hold a whole system in your head, the skill that lets you judge what the model hands back.
Wrong advice for junior developer: look at the AI generated code, since you need to mature as a programmer. Right advice: save time with AI coding and write by hand "get-better" programs: interpreters, games, ray tracers, ...
2