Chen (@chenYijl) header
Chen (@chenYijl) profile picture

🤖 AI watch | 🌍 World watch AI fails, BTC dumps, meme pumps. Trends, facts, the odd hot take. Built ChenAI: Claude & Codex, stability first. https://nitter.gq/t.co/TsT5A0xaHD

Joined August 2026
Photo by Chen (@chenYijl)Photo by Chen (@chenYijl)Photo by Chen (@chenYijl)Photo by Chen (@chenYijl)Photo by Chen (@chenYijl)Photo by Chen (@chenYijl)Photo by Chen (@chenYijl)Photo by Chen (@chenYijl)Photo by Chen (@chenYijl)Photo by Chen (@chenYijl)
Pinned Tweet
I built ChenAI after my relay died 3 times in one night, each time right in the middle of an agent run. I was running Claude Code and Codex on the cheapest relay I could find. It worked fine for small chats. Then I gave it a real task, a long one, and at 2 a.m. it just stopped. No error I could fix, no backup, just a dead task and a morning of rework. So I built the relay I actually wanted to use. ChenAI only does Coding and Agent work. Nothing else. Claude and Codex, 11 models in total: Fable 5.2, Opus 5.5, Sonnet 5.5 and Haiku 5.5 on the Claude side, GPT-6.2 Sol, GPT-6.2 Astra and five more on the Codex side. Every model runs on multiple upstream channels. When one goes down, traffic switches to the backup automatically. You feel a slowdown, not a dead task. It costs $1 to start. You pay in RMB and the credit lands in dollars, and the model prices are on the site in the open, input and output separately. Setup takes about 6 steps in CC Switch: add ChenAI as a supplier, fill in the endpoint and your key, pick your models, then start routing. No config files to edit by hand. I built it for my own overnight runs. If yours keep dying too, try it: chenai-ai.com
3
88
Type four words and your notes turn into a website. On November 16, Google gave Gemini Notebook a site builder. Type make me a site and it turns your sources into dashboards, trackers, or visual explainers. The demo turned chemistry notes into a clickable periodic table with a quiz. Sales data became a dashboard with region filters. It is rolling out on the web to users 18 and older. Mobile comes later. Your notes were the draft. Now they are the product.
2
33
Microsoft slipped GPT-6.2 Sol into Word, Excel, and PowerPoint for $2 per million tokens. On November 20, Microsoft said Sol is rolling out worldwide in Copilot Chat, Cowork, Studio, and the Office apps. It ships with the Microsoft 365 Copilot license. Astra is the awkward sibling. It reaches everyone in Cowork and Studio, but Word and Chat only if your company joined the Frontier program. Sol costs $2 per million input tokens. Astra costs $8. The cheap model goes everywhere. The flagship needs a permission slip.
43
The prompt you marked do not store got stored anyway. On November 22, OpenRouter kicked 3 providers off its platform. It stripped 125 endpoints of Zero Data Retention status. Dragos planted fake credentials in 2 million prompts and watched them get used. Phala had an attacker inside one inference server for weeks. Novita sent customer requests to MiniMax's official API instead of its own servers. OpenRouter says its own systems were not breached. It notified affected customers anyway. A privacy toggle is a promise. Promises need someone checking.
2
1
81
Somebody bought Bitcoin at 70% off on Saturday and kept it. On November 21, Kraken's Canadian dollar market went haywire. The low print was about $16,928 in US dollars. 122 trades filled below $20,000 in 15 minutes. Here is the fine print. Bitcoin in US dollars never left $57,593 to $57,900 on the same exchange. The Canadian order book ran dry. The asset never crashed. Kraken blames heavy selling. No breach. Completed trades stand. Lesson: the price on your screen is only real if someone is standing behind it.
2
1
50
I tried to ask ChatGPT one question on Friday night. It said no. Then it said no again. Then it kept saying no all evening. On November 21, requests to ChatGPT and the OpenAI API failed at rates peaking at 75%, for 3 and a half hours. Paid plans. Free plans. All of it. The cause was routine maintenance. OpenAI was servicing the Postgres database behind its Primary API service, the one that handles login and authentication for everything OpenAI runs. Both read replicas died at the same time. That is the part worth sitting with. These replicas were supposed to be copies. They turned out to be load-bearing walls. When they went down, login went down, and when login went down, Chat Completions, voice mode, file uploads, image generation, and the API all went down with it. Engineers tried reviving the dead replicas first. That failed. They tried spinning up replacements in another region. That failed too. The fix that worked was building two new replicas in the same region, and traffic came back at 5:21 in the afternoon Pacific time. OpenAI published a full write-up the same day. Credit where it is due. Most companies would have posted a green status page and moved on. It also shipped a switch to move traffic off replicas by percentage, and put a hold on that kind of maintenance. Here is the uncomfortable part. If you build your whole product on one provider, their maintenance window is your outage. Your users do not care which database died. They care that your app stopped answering. I run half my workflow through one API. Friday was a reminder that a backup plan is not paranoia. It is maintenance. How many of your tools would keep working if your main provider went dark for 3 hours?
5
1
99
Gemini stopped chatting. It started running errands. On November 17, Google made Gemini Tasks free for everyone on the Gemini 4 Hydrogen model. One toggle flips the app from Chat to Tasks. Then it can call a business and book your appointment. It can buy something with Google Pay. It can sign into a website for you. Google's own advice: check every booking yourself. The agent can still get things wrong. The takeaway: the chatbot era is ending. The errand era just started.
3
1
84
X will pay you to post. The math is less exciting than the pitch. To qualify for Original Content Rewards you need 500 verified followers and 500,000 Home Timeline impressions from verified users in 90 days. Replies do not count. That is about 5,556 qualifying impressions a day. Creators report $1.05 to $1.23 per 1,000 qualified impressions. Clear the bar and you earn roughly $525 to $615 a quarter. The payout is real. The hourly rate is the joke.
5
3
139
Memory is now the most expensive part of the AI boom, and it is about to get worse. On November 18, TrendForce forecast HBM4 prices will more than double next year, from over $2.05 per gigabit to $3 to $5.50. Average HBM prices rise 122% to $3.70 per gigabit. Commodity DRAM rises 10% to 15% this quarter. Cloud providers plan to raise AI spending 60% next year. Lesson: the chip is not the bottleneck anymore. The memory next to it is.
6
2
125
Tesla just gave its self-driving software a video-game view of the road. On November 21, Tesla started rolling out Full Self-Driving v15. It claims 25 times better handling of safety-critical cases, and it adds Reality View: a live, photorealistic render of the street around your car, from a camera that does not exist. Early-access drivers get it first. Everyone else waits for the December holiday update. Your car now shows you what it thinks it sees. Watch what it gets wrong.
1
31
Friday, Wall Street stopped believing its own AI story. The Nasdaq fell 5.3% in one session. That is the worst week for the index since 2020, down 10% in five days and 18% below its October record. The year's gains are gone. Nvidia fell 14% on Friday alone. The stock closed at $125.53, down 23% for the week and 48% from its October peak of $239.24. Anthropic dropped 8.5% on its second day as a public company, to $88.91. The S&P 500 fell 3.7%, its seventh straight down day. The VIX, Wall Street's fear gauge, jumped 33% to 37. CoreWeave fell 12%. AMD fell 12% in a day and 20% in a week. The semiconductor index dropped 11%. The Wall Street Journal called it a $4 trillion AI wipeout. Here is the part that should worry you. The same week Anthropic went public, its own filing showed the loop. Seven of its suppliers, Nvidia, Google, AMD, Broadcom, Micron, Microsoft and TSMC, committed $26 billion to buy Anthropic stock at the IPO price. Amazon owns 14.7% of the company. Google owns 13.7%. Anthropic has committed to spend $900 billion, mostly on compute, and owes Amazon at least $110 billion and Google at least $121 billion. The money goes out to suppliers as purchases. Some of it comes back as equity. Then it goes out again as purchases. Anthropic is still losing tens of billions a quarter. It made $17 billion in revenue last quarter. The IPO changed who pays the bill: public investors and bond buyers now, not just suppliers. The AI trade is the best trade of the decade. The exception is the bill for it, and Friday the market finally read the bill. The concession: most investors still say they believe in AI. They just stopped believing the price. How much of your money is still riding on the AI trade?
2
1
125
Google gave its AI search an incognito mode. On November 18, Google rolled out temporary threads in AI Mode. One icon starts a chat that never enters your history or your recommendations. The session detaches from your account. Past preferences stop shaping the answers. Close the thread and it is gone from your screen. One catch. Google still keeps server copies for about two months. Temporary hides the chat from you, not from Google.
2
1
73
Decision models have a new leader, and it costs 10 cents. On November 20, Baseten shipped bison-1 and groundhog-1. Two open-weight decision models built on Google's Gemma 4. Bison-1 scored 76 on the Artificial Analysis Decision Index. Number one out of 19. It also takes 128,000 tokens of context, eight times the old models. Groundhog-1 is the cheaper sibling at 3.5 cents per million tokens. Lesson: the model that answers yes or no is now a leaderboard sport too.
5
4
208
Classical computers top out at 88.9%. Willow hit 98.3%. On November 19, Google published a Science cover paper. Its Willow chip won the magic square game 98.3% of the time. The classical ceiling is 88.9%. The team ran each version of the game 100,000 times. Then it scaled the tests up to all 105 qubits on the chip. This is a hardware benchmark, not a useful computer yet. The win rate falls as noise grows. Quantum advantage now has a scoreboard. And the scoreboard has numbers.
3
43
Apple wants 200,000 people to test its health chatbot. On November 15, Apple launched the Health AI Usage Study with OpenEvidence. The app answers health questions with citations from the New England Journal of Medicine, JAMA and Nature. You need to be 21 or older, in the US, on an iPhone 15 Pro or later. The study runs more than 12 months. Apple says the data will not train AI models. The app cannot diagnose anything. Apple is studying trust before it ships a doctor in your pocket.
3
1
64
Your AI agent now has its own inbox. On November 20, Meta opened Muse Mail to everyone in supported markets, free plan included. Your Muse gets an address ending in @muse.ai. It can send and receive email for you. Forward it a receipt. Add it to a thread. Ask it to chase a booking. The address is separate from your Gmail or Outlook. Muse never sees your inbox unless you forward something to it. So what: your agent now has a mailbox. Guard the handle like a username.
4
1
2
134
Claude can see your screen at work now. On November 20, Anthropic put computer use into beta for Team and Enterprise plans. Claude watches your screen and clicks, types, and works inside the desktop apps you allow. It is off by default. Only an owner can switch it on, and admins can block whole categories of apps. Investment and crypto platforms are off limits by default. So the chatbot got a desk job. IT decides which apps it can touch.
3
2
134
90.2 beats 89.9. 🤯 That is the score from a head-to-head test of eight AI platforms. And the gap between those two numbers is the whole story. FirstPageSage published the study on November 13. Eight platforms got the same 400 requests. Every answer was checked. 3,200 answers in total. Claude finished first at 90.2 out of 100. ChatGPT second at 89.9. Gemini third at 88.8. Three tenths of a point. That sounds like a tie. It is not. Claude got 97.8% of its calculations right. It invented a number that was not in the data in only 1.4% of its answers. Meta AI, last at 77.3, made up numbers in 8.7% of its answers. Six times Claude's rate. That is the number that matters if you hand a bot your books. Claude won four of the eight tasks. Statistical analysis. Explaining results. The boring jobs where one wrong decimal costs real money. Here is the concession. Claude did not win everything. Copilot in Excel still leads on spreadsheet formulas. Gemini still leads on working with large files. ChatGPT still builds the best charts. So the honest ranking is simple. Claude is the best all-rounder for data analysis. The exception is the specialist job. Live inside Excel formulas all day and Copilot still owns that corner. Work with huge files and Gemini still handles them better. Claude ran Opus 5.5 in Excel. ChatGPT ran GPT-6 Sol. Gemini ran Gemini 4 Argon. I have watched AI demos fall apart the moment a real spreadsheet shows up. A messy customer list. A formula pointing at the wrong column. A total that quietly counts the same sale twice. This test fed the bots exactly that kind of mess. The winner was the one that hallucinated the least. That is a low bar. It is also the whole job. If your AI cannot be trusted with a column of numbers, it cannot be trusted with your business. Claude just proved it can handle the column. Which AI do you trust with your spreadsheets?
2
1
87
I opened my second MemeBitcoin Alliance pack. Join the Alliance and become a Crypto Guardian. raffle.memebitcoin.org/c/2d9…
20
3 million lines of code. $10,000 in API tokens. Zero hardware experience. That is what it took one engineer to turn a cloud FPGA into a working LLM inference chip. Ankit Gupta posted the writeup on November 20, and it hit the front page of Hacker News the same day. He started with almost nothing. He had never written Verilog. He had not written serious C++ in a decade. He rented an AWS F2 instance, a reprogrammable chip with 16 GB of memory, and pointed a coding agent at it. The agent ran for several days inside Conductor. It wrote roughly 3 million lines of code across the whole experiment and parsed 30 million lines of output. It ran hundreds of optimization experiments. The model was a ternary-weight design from DeepGrove, built for memory-starved hardware. The first full version ran at 6.43 tokens a second. Bit-exact, meaning it matched the CPU reference output exactly, but slow enough to be useless. Then the agent started optimizing. Host overhead, burst attention, parallel routing, a new pipeline, prompt speculation, weight packing. One idea at a time, each verified against the reference. The final number was 213.13 tokens a second. That is 33 times the baseline. Same chip. Same model. Same $10,000 budget. 🤯 Here is the part nobody should skip. The agent was not magic. Around 10 tokens a second, it stalled, and Gupta pasted in a design prompt from the DeepGrove founders. One human paragraph unblocked the whole run. So the honest verdict is this. A coding agent can now do hardware design that used to need a specialist team. The exception is the human still has to know which paragraph to paste. I write about AI writing code every day. This is the first time I have seen it write a chip. I have never written Verilog. After reading this, I am not sure I ever need to.
5
115