Robotics deep dive: HUMANOID CHOKEPOINTS Almost every supply-chain post about humanoids has one country in it. Trace five critical inputs one at a time and they land in four different places. Magnets first, because that is where the argument usually stops. Every high-torque joint runs on sintered NdFeB, and on the IEA's 2024 figures China holds 91% of refined output and 94% of sintered permanent magnets, against 60% of the mining. Mining is the distributed part, and the chokepoint is the chemistry. The precision reducer is the node almost nobody names. Nabtesco states on its own site that it holds approximately 60% of the global market for the precision reduction gears in industrial robot joints. Harmonic Drive Systems is the incumbent on the strain-wave side. Both Japanese, both sitting on decades of gear grinding, heat treatment and yield that nobody has replicated at price. There is no mine to open here. Optimus runs 14 linear actuators; per-robot screw counts get estimated anywhere from 14 to more than 40, at $1,350 to $2,700 each. A planetary roller screw is a micron-tolerance ground part, and only a handful of firms anywhere make them to that standard: Japanese, German, Swiss, with Chinese entrants scaling into it now. The constraint is grinding capacity and the people who can run it. Compute inverts the picture. Nvidia designs Jetson Thor in California and cannot manufacture it. Roughly 92% of the world's sub-10nm logic capacity sits on one island. Battery cells go back to China: over 80% of global cell capacity, about 99% of LFP. A McKinsey component map, reported by Forbes in June, puts it the same way: China is overwhelmingly dominant in exactly one category, magnets for motors, while bearings, driver boards and sensors have suppliers in several countries. So it is four dependencies, held by parties with different export regimes, different politics and no shared interest in coordinating. A supply chain with one chokepoint is one negotiation. Reshoring the magnets still leaves you asking Japan for the gearboxes and Taiwan for the brain.

2 hours 43 minutes. 442 steps. 5,406 lines of code. The page still does not load 💀 I gave Qwen3.8 Flash Next my standard bench prompt: a voxel Japanese pagoda garden in Three.js, 16,500 voxels, modular source, run it and screenshot it when it's done. Other models finish this one easy. It never opened a browser. Never started the dev server. Never counted a voxel. It spent the entire run fighting its own export statements. petals.js got rewritten 19 times. Not patched, rewritten from scratch every time. water.js 9 times, palette.js 11, lanterns.js 7. It was loop between exactly 2 errors: Export 'update' is not defined in module. 53 times. Duplicate export of 'createPetals'. 10 times. Add the alias and you get the duplicate. Remove it and the export goes missing. In the last 70 steps it flipped between those 2 states 9 times. Twice it deleted the file and wrote it back identical. 85 writes against 7 reads. It almost never looked at what was already on disk. It compacted its own context 4 times and came back to the same 2 errors every time. It stopped with 3 files still failing node --check, all on the same line, update as updatePetals. 2 modules that main.js imports do not exist at all, and one of them is the lanterns file it had written 7 times. Then I gave the same model the same prompt as a single HTML file. It shipped a complete working scene in 1 pass. So it is not a capability wall. The module code it wrote is decent. It cannot hold its own module graph together, and it has nothing that tells it that it is going in circles. 438,737 output tokens and it never once said it was stuck. I spent 3 days making this thing run on my rig. An FP8 KV cache path that vLLM rejects on Ampere in 4 separate places. The 51B n-gram table baked down to FP8 so it fits in host RAM. W4A16 weights, the full 262K context, on 4 gaming GPUs from 2020. All of that works. And then it cannot wire 16 files together. So I'm going back to Qwen3.8 27B as my daily driver until something changes. Could be something in my own build doing this, I'm rebuilding the quant to find out. Either way I'll post what comes back.

# OpenCode Features and Practical Usage 📜 Tired of re-explaining your project's conventions every single session? `AGENTS.md` fixes that: write it once, and the agent always operates with your rules in mind. 🏷️ Title: AGENTS.md 🔗 URL: 📘 Overview `AGENTS.md` is the file you use to give OpenCode custom instructions. Document your project's conventions, architecture, and build steps there, and that content is always included in the LLM's context so the agent behaves the way your team expects. ⚙️ How It Works Rules live at two levels: ・Project-level: an `AGENTS.md` at the repo root, applied only within that directory tree. ・Global-level: `~/.config/opencode/AGENTS.md`, shared across all sessions and best for personal preferences. At startup OpenCode searches in order: local `AGENTS.md` or `CLAUDE.md` (walking up from the current directory) → global `~/.config/opencode/AGENTS.md` → the Claude Code compatibility file `~/.claude/CLAUDE.md`. The first match in each category wins. The design is close to Cursor-style rules and makes migration easy. 🛠️ Practical Usage To pull in external docs as instructions, list them in the `instructions` field of `opencode.json`, e.g. `"instructions": ["CONTRIBUTING.md", "docs/guidelines.md", ".cursor/rules/*.md"]`. Globs are supported. If writing from scratch feels like a chore, run `/init`: it scans important files, may ask targeted questions, and generates or improves `AGENTS.md` for you. Commit the result to Git so the whole team shares it. 💡 Use Cases Capture tacit knowledge like "commit messages in Japanese," "tests use pytest," or "never import this layer directly" in `AGENTS.md`, and both new teammates and the agent share the same assumptions, cutting down on review churn. In monorepos, `instructions` globs let you bundle per-package conventions. ⚠️ Caveats File references hand-written inside `AGENTS.md` are not expanded automatically. When you need multiple files reliably loaded, the `instructions` field of `opencode.json` is the dependable choice. An existing `CLAUDE.md` is recognized for compatibility, but consolidating new content into `AGENTS.md` keeps things tidy. Remote URL references carry a 5-second timeout. #OpenCode# #AGENTSmd#

A TON OF THINGS HAPPENED IN THE STOCK MARKET TODAY. Here's a full recap: 1. Tesla $TSLA rose 5% today, marking its best day since early July. The move comes ahead of Tesla’s upcoming Cybercab Launch Event, where investors will be watching for concrete updates on robotaxi deployment, Cybercab production, autonomy progress, and whether Tesla can turn the vehicle from a concept into a real revenue-generating fleet. Tesla has already disclosed that Cybercab production has started, with public-road engineering tests and employee rides beginning at Giga Texas. The key question is whether the event delivers actual deployment details rather than another demo, including timing, fleet size, city expansion, and cost-per-mile economics. 2. Treasury Secretary Bessent today at the G20 summit said he believes the Japanese government and Bank of Japan will take actions that lead to a stronger yen. On the Fed, he said he would not speculate on what policymakers may or may not do, but added that traditionally “you don’t raise into a supply shock.” On bonds, Bessent said he and Warsh are “on the same page,” while noting he does not believe he can change the market’s equilibrium price. He also said Trump believes Iran is not ready to make a deal yet, and that Operation Outcast is aimed at creating conditions that bring Iran back to the table. 3. Anthropic has reportedly signed a $35B cloud deal with Lambda, an Nvidia-backed cloud provider, according to WSJ. The twist: Nvidia $NVDA is effectively the landlord. The data center is being built by Hut 8 in Nueces County, Texas, and Nvidia signed its own agreement with Hut 8 a few weeks ago to lock down the capacity. Lambda will use that space to run chips it bought from Nvidia, which is also one of its investors. The deal highlights how Nvidia is extending its role beyond selling GPUs and deeper into the AI infrastructure stack with financing, capacity access, cloud partners, and the physical data center footprint behind frontier AI demand. 4. The FTC and 22 states are reportedly preparing to sue Amazon $AMZN over alleged secret ad price hikes. The FTC is expected to claim Amazon manipulated its ad auctions by inserting its own “soft reserve” bid above the runner-up bid, effectively raising the minimum price advertisers had to pay. The practice allegedly began in 2018 and generated tens of billions of dollars over 7 years. In recent years, Amazon allegedly used the mechanism to raise minimum bids 70%–80% of the time, with the FTC claiming it increased pay-per-click costs by as much as 50% on major shopping days. Amazon generated $68B in advertising revenue in 2025, and this would mark the FTC’s third major case against the company after its Prime case and ongoing monopoly lawsuit. 5. Retail investors bought roughly $250M of Nvidia $NVDA shares on Wednesday, their 3rd-largest daily purchase since mid-May. That extended the retail buying streak to 15 consecutive trading sessions, with investors purchasing more than $2.5B of Nvidia stock over that period. For context, the largest daily retail purchase came in late February at about $1.0B. Over the last 12 months, retail investors have bought roughly $30.0B of Nvidia shares, the most of any Magnificent 7 company. 6. The White House published details of the U.S.–Venezuela oil deal, saying Venezuela has granted North American Blue Energy Partners 100-year concessions for 17 oil fields. The fields reportedly contain roughly 65B barrels of proven oil reserves, while NABEP plans to invest up to $100B into Venezuelan oil infrastructure to rapidly increase production. As part of the deal, the U.S. government receives a 35% equity stake in NABEP’s parent company at “zero cost to taxpayers,” the right to purchase 20% of all current and future production at production cost, and the right of first refusal on the remaining 80%. The U.S. also gets veto power over board appointments, with a majority of NABEP’s board required to be U.S. citizens. Venezuela is expected to receive roughly $200B in royalties and tax payments over the first 25 years, while the White House is calling it the “biggest oil deal in world history.” 7. U.S. annual interest expense has climbed to a record 18.5% of federal government revenue, officially surpassing the prior record of 18.4% set in 1991. This share has more than quadrupled over the last 4 years as interest costs on public debt have surged. Annual interest expense now stands at a record $1.25T, more than 4x the level seen in 1991. Meanwhile, the 30-year Treasury yield is trading around 5.21%, just 13 bps below its highest level since 2007. For context, the 30-year yield was around 8.00% in 1991, highlighting how much larger the U.S. debt burden has become. 8. BofA reiterated its Buy rating on Meta $META with an $810 price target, highlighting the potential launch of Meta’s consumer AI agent Hatch in early September inside Instagram and WhatsApp. Hatch can reportedly browse websites and navigate user interfaces to complete autonomous tasks like purchases, restaurant bookings, form filling, and communications. Meta is also reportedly targeting an October launch for its new AI model, Watermelon, which has shown frontier-level performance on certain internal benchmarks, with Hatch expected to integrate Watermelon’s automation capabilities after launch. BofA sees a major long-term opportunity in consumer AI agents, with Meta’s global user base giving it a significant adoption advantage, but said success will depend on ease of use, real utility, ecosystem integration, user trust around automation and security, and subscription pricing that creates consumer value while earning a return on Meta’s compute investment. 9. The top 10 most active options today by contracts traded were $TSLA with 4.4M contracts, $NVDA with 3.3M contracts, $AAPL with 1.3M contracts, $AMZN with 1.0M contracts, $MU with 952K contracts, $META with 606K contracts, $SPCX with 561K contracts, $INTC with 470K contracts, $PCG with 460K contracts, and $GOOG with 453K contracts. 10. Trump said interest rates are “too high” and said he has “a lot of respect” for Fed’s Warsh, adding that Warsh will “do what he has to do.” On Iran, Trump said any strikes would be limited, while noting that a lot of oil is still coming out of the Strait of Hormuz. He said ships came through Hormuz last night with Navy assistance and that the U.S. is averaging about 30 ships per night out of the Strait. Trump also said Iran “doesn’t know who the leader is,” adding, “we’ll see what happens.” 11. Take-Two $TTWO fell 7% as GTA 6 leaks continued ahead of the November 19 release. Recent leaked footage has reportedly shown combat, vehicle theft, the return of the six-star wanted system, a biplane flight over Vice City, and a 4.5-minute Lucia story clip. The pressure reflects growing concern around how much of the game is being revealed before launch, even as GTA 6 remains one of the most anticipated releases in gaming history. 12. China’s CXMT has reportedly started small-scale production of HBM3E, the advanced memory used in AI processors including Nvidia’s H200 and Blackwell GPUs, according to The Information. Alibaba’s T-Head and Cambricon are already testing CXMT’s HBM with their AI chips and could begin using it in products as early as 2027, with CXMT planning to expand production next year. The development could ease a key bottleneck for China’s domestic AI-chip industry, which remains restricted from buying the most advanced foreign HBM. The caveat is that CXMT is still facing low yields and remains roughly 3–5 years behind the leading HBM players, while $SKHY SK Hynix, Samsung, and Micron $MU are already mass-producing HBM4 and sampling HBM4E. WALL STREET IS THE GREATEST SHOW ON EARTH.

i think personalisation is about to become the biggest differentiator in llms, and not the kind most people think of. most people mean an assistant that remembers your name and what youve typed in the box previously. i mean one that already knows what you did and what you like before you ask it anything. @myrad_hq's already built that part. it skips the box entirely, it just reads what you did. > ask it what to eat tonight and it skips past japanese, since you've already ordered it four times this week, and puts mexican in front of you instead > ask for a gym routine and it already has your strava, three runs after work tuesday wednesday friday, two rides the other days, so it builds strength around the days you're already spent instead turning your legs to jelly before tomorrows run. > ask for a spending audit and it already knows your uber rides spiked the same week your food app orders doubled, so it flags the actual pattern instead of making you dig through two apps most chatbots guess from what you say. the myrad approach is reading what you did and never hallucinating based on inputs 3 context windows ago. your data mode is walled off, wallet gated, nobody else's data touches yours. the verification underneath runs on zero knowledge proofs, it confirms you actually watched something, rode somewhere, ordered something, without myrad ever holding your netflix password or logging into the account itself. proof of the activity, never custody of the account. this is what personalisation actually evolves into, not a memory feature bolted onto a chatbot, the actual next layer underneath every ai product built after this one. now picture that same verified context sitting inside a humanoid assistant instead of a chat window, one that already knows your routine and your habits before it opens its mouth to ask you anything. myrad is building for the future base:0x693bad964f815f32fabe0b9d4911865bffc30172

I run a fund the United States built in 1934 with $2 billion it made in an afternoon. It did not sell anything for the $2 billion. It raised the official price of gold from $20.67 an ounce to $35 an ounce and kept the difference on the gold it already held. We wrote a larger number next to a thing we were not touching, and the larger number was real enough to open a fund. That fund is mine now. On Friday I used it to buy the yen. I did not print the yen; the yen is Japan's to print. To buy yen you sell something else, and on Friday I sold euros. I did not sell them personally. The fund has no trading floor. It has the Federal Reserve Bank of New York, at 33 Liberty Street, and 2 primary dealers, which on Friday were Goldman Sachs and Morgan Stanley. There is, in a room on Liberty Street I have visited twice, a terminal with a key on it marked EXECUTE. I have never pressed the key. On Friday someone pressed it before 9:00 AM, several times, and I learned that the yen had been supported the way I learn most things, from the confirmation log. I did not attend. Attending is not the mandate. I keep one of the confirmations framed. Not Friday's. An older one, from a prior operation, because it is the most honest document the fund owns. It sits on a bank's letterhead and it reads, in full: Sold. EUR/JPY. Your account. The words United States do not appear in it anywhere. A nation reached across the ocean and overruled the price of money, and the paper that records the act is a form letter that does not know which country it served. I find that clean. A trade should know the account and nothing further. People treat the 1934 story as history. I keep it as the mandate, unedited. The fund was born from a single discovery, which is that a price is a number, and a number is correctable by whoever holds the pen. Everything I do is that discovery, kept current. The market spends all day arriving at a number for the yen. My work begins the moment the number it arrives at is one we would have preferred it did not. I do not argue with the market. I keep buying at the number we prefer until the screen shows the number we prefer, and then the number we prefer is the market's, retroactively, and everyone can go home. A reporter asked me last year whether any of this changes prices inside Japan. It was a generous question and I gave him the clean answer. My mandate is the exchange rate. The grocery store belongs to someone else. The rice at the Ito-Yokado in a Tokyo suburb costs about the same whether the yen prints 152 or 145; a 7-yen move on my screen is not a grain of rice on that shelf. The work stops at the exchange rate. It was built to stop there. The fund is not a rice program. I have filed nothing with the rice program. My filings go to the Federal Reserve Bank of New York, Settlement Division, and in 30 years they have never once asked me about rice. We are not careless about it. Japan spent record amounts of its own reserves defending the yen last year, and the yen kept falling, and we studied that closely. I built the deck myself. It was titled Lessons From Prior Interventions. Slide 14 read, Coordinated action amplifies signal. The slide after 14 recommended coordination. I reviewed the deck again on Saturday, and the deck was correct. None of this waits on Congress. Nobody forgot to close that door; the open door is the design. A fund that had to be voted on could not move before the market moves, and a fund that cannot move before the market is only a slower participant in it. So I was placed to the side of the vote, with money that was never appropriated, for the exact mornings when a vote would have slowed me down. I am the part of the government that does not wait to be asked. The yen closed Friday where I needed it to close. The rice at the Ito-Yokado cost the same on Saturday as it had on Thursday. My mandate covers the number on the screen. The rice is a different number, at a different desk, and I have never met the person who sits at it. But the number is holding. And I would like the record to show that the fund supports their work.

BREAKING: Japan's central bank just pulled money out of the economy faster than at any point since 2007. Japan's monetary base shrank 15.7% year over year in August, worse than the 13.5% decline economists expected. This is the Bank of Japan actively draining liquidity, mainly by letting bond holdings run off instead of reinvesting, with the average outstanding base falling from ¥554.9 trillion in July to ¥543.0 trillion in August. This is the same force behind Japan's bond market chaos. Fewer yen in the system means fewer buyers for government bonds, which is a big reason JGB yields have been hitting 31-year highs across the board. The BOJ is trying to normalize policy after over a decade of money printing, but doing it while inflation is already hot and the yen is already weak is what's turning this into a full blown crisis.

Japan’s fund for retired racehorses has raised more than 1.8 million US dollars over the past 10 years, mostly thanks to fans of the mobile game Umamusume Pretty Derby. This fund, called the Nice Nature Memorial Donation, is run by the Retired Horse Association of Japan. It uses the money to pay for vet care, safe homes, and other needs so old racehorses can live well after they stop racing. The fund started small in 2017, but donations jumped more than 30 times higher after the game came out in February 2021. People gave about 35.8 million Japanese yen, or around 225,000 US dollars, in that first year alone because fans fell in love with the characters based on real horses and learned about the hard life of retired horses from the game and its TV show. Even after the real horse Nice Nature died in May 2023 at age 35, people still donate a lot to the fund named in her honor. Fans have also helped other horses by raising money to fix stables for Meisho Doto and by giving support to Haru Urara.

🇯🇵 Some of you who are interested in Japan might have noticed recently of this new "trend" of "sad salaryman videos" coming out of Japan. The majority of these videos are fake, based on exaggerations and have been debunked. While overworking is a real problem in Japan, many of these videos are often fake. They often get millions of views, thousands in donations, etc. Some of them are not even run by Japanese people as the videos are just subtitles (in English) but contain no actual audio and are likely using AI. Chinese people have also started making these to make money from YouTube, TikTok, etc. It's gotten to the point where foreigners have actually started donating and foreign YouTubers are even making videos debunking them. I'm not suggesting everything on the internet is fake but unfortunately a lot of things are so we should all be careful.

Build a humanoid robot yourself! 🪚 OpenArm is an open-source humanoid robot.  It comes with full CAD, control code, firmware, and simulation tools, everything needed to build, modify, and operate it. The arms are designed to be compliant and backdrivable. Teleoperation is supported, with force feedback and real-time gravity compensation so operators can guide the arm naturally. Super important, on the simulation side, OpenArm works with platforms like MuJoCo and Isaac Sim, letting developers test policies in virtual environments before running on hardware.  Assemble it yourself from a kit or get it prebuilt, the goal is accessibility for research labs, small teams, and enthusiasts. The project is run by @enactic_ai in Tokyo, Japan, and aims to lower the barrier for experimenting with dexterous manipulation. Let's put robotics in mainstream! 🔥🔥 🔗 Find the link here: ~~ ♻️ Join the weekly robotics newsletter, and never miss any news →