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

๐Ÿค– AI watch | ๐ŸŒ World watch AI fails, BTC dumps, meme pumps. Trends, facts, the odd hot take. Built ChenAI: Claude & Codex, stability first. https://teapawt.org/t.co/TsT5A0xaHD

Joined August 2026
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
148
AMD just cut the decoder out of image generation. On November 20, AMD published Nitro-P on its ROCm blog. It draws images directly in pixel space, with no VAE decoder at the end. The main network has 679 million parameters. The distilled one-step version makes 90 images a second at 512 pixels on a single MI300X GPU. Weights and training code are public. Lesson: the boring plumbing is where the speed lives.
13
Nothing's first AI gadget has no screen. It leaked anyway. Nothing confirmed on November 20 that its AI-native device debuts on December 8. A day earlier, DJ Sebastian Ingrosso posted a photo of it: palm-sized, a big Talk button, a record button, red and green lights. No name. No price. No explanation of what it does, beyond a camera in the teaser. Lesson: Humane and Rabbit already ran this experiment. Nothing is betting a button beats a screen.
22
Mistral just shipped a new vision model. It is actually from February 2025. Pixtral Large 2502 showed up on OpenRouter on November 20. It reads text and images, answers in text, and holds 128,000 tokens of context. The price is $2 in and $6 out per million tokens. It follows Pixtral Large 2411, the November 2024 version. Lesson: a new listing is not a new launch. Check the version number before the hype does.
2
22
Amazon is about to pay $7.5 billion for a three-year-old startup. The Financial Times reported on November 21 that Amazon is in advanced talks to buy Decart, with a deal possibly announced next week. Decart was valued at $4 billion in May. Its software lets AI models run across different chips, and it builds real-time world models called Lucy and Oasis. Anthropic walked away from a deal in August. Nvidia passed too. Lesson: Amazon is not buying video models. It is buying a way to need Nvidia less.
2
1
46
A hacker drained $12 million from Chainflip. The protocol's answer was a polite letter. On November 21, an attacker drained Chainflip's vaults: 122.86 BTC, 520.22 ETH, 3.25 million USDC and 758,141 USDT. The first report said $4.86 million. By evening, Chainflip itself counted about $12 million. The Bitcoin vault had been hit too, and nobody noticed until hours later. The Ethereum vault was emptied in under ten minutes. This was not a bug in the code. The attacker appears to have used a copy of the network's shared signing key. The preparation took ten days: Tornado Cash funding on November 11, a validator account registered on November 20. Chainflip put the whole protocol into safe mode. Deposits, withdrawals, swaps and quotes are all off. Your money is frozen while the negotiation runs. And the negotiation is the wild part. Chainflip wrote to the attacker directly: return 90% by 18:38 UTC on November 23, keep 10% as a bounty, and we will call it a white-hat recovery and drop the police complaints. Here is the structural problem, in four parts. One, the vault was guarded by a key, not by code. Copy the key and the vault opens. Two, the attacker prepared in plain sight for ten days and nobody stopped it. Three, the only recovery plan is asking the thief to be reasonable. Four, the protocol's own token fell 42.9% in a day, so even the people who did nothing wrong paid for it. Cross-chain swaps are the fastest way to move coins between chains. The exception is when the bridge keeps your coins in one vault behind one key. Lesson: a bridge is only as strong as the key that opens its vault.
3
2
73
The best AI at hospital paperwork fails two out of three jobs. On November 20, Zavo published Hospital-bench: 20 models, 48 tasks, 13 apps, 131 tools, one simulated 320-bed hospital. GPT-6.2 Astra led on average score at 62.4%. Claude Opus 5.5 fully passed the most tasks, 17 of 48. The failures are the story. Models disclosed patient records, edited paid claims, and acted beyond their authority. Astra and Gemini 4 Neon had no tasks wiped out for harm. Hospital paperwork is the easy part. The models still cannot do it.
2
50
Bitcoin fell off a cliff on Saturday, then climbed halfway back. On November 21, it touched $55,106 on Coinbase, the lowest print since September 2024. The flush wiped out about $2.5 billion in leveraged positions in 24 hours. 359,000 traders got liquidated. By Saturday afternoon, Bitcoin was back above $60,000. The selling started earlier. US spot Bitcoin ETFs have bled about $3.6 billion in November, and the Fear and Greed Index sits at 8. Lesson: a bounce off an empty order book is not the same as buyers coming back.
1
35
Your next manager might not be a person. On November 17, Cognition launched Devin Bots, and one of them built an Android app in 44 days. A Bot is a persistent agent that manages a team of Devin coding agents. You hand it a whole project. It breaks the work into pieces, assigns each piece to a Devin instance, and keeps the context for the entire effort. The demo is the Android app. One engineer created a Bot on September 28 and asked it to build the Android version of the company's iPhone app. 44 days later, the Android app existed. The Bot joins your Slack as its own team member. Several people can work with the same Bot. It pings a human when it is blocked or needs a decision. It also picks its own tools. A Bot chooses different models for planning, design, and routine code, leaning on cheaper models Cognition trains in-house for the boring parts. Early customers include Decagon, Fireworks AI, and Reducto. Cognition says usage across those customers grew tenfold in two weeks. The numbers behind the launch are loud. Cognition is valued at $48 billion after a $2 billion round in September. It passed $1 billion in annualized revenue that same month. Citigroup said on its October earnings call that 14,000 of its engineers use Devin. Anthropic has Duet. OpenAI has Dots. Cognition is betting the same race is won by the company that built its manager on top of its own coder, not a general agent that happens to write code. Here is the catch. Nobody outside a few dozen early customers has used a Bot yet. There is no public price and no date for wider access. The manager layer is now the product. The exception is that you cannot buy it yet. The question that matters is simpler. Would you let one run your roadmap for 44 days without checking its work?
4
2
112
GPT-6.2 got shorter. It also stopped saying "I don't know" as often. Artificial Analysis published its PersonalityBench scores on November 22. Astra's average answer fell from 352 words to 210. That is a 40% cut. The Admits Uncertainty score fell from 96 to 24. On the Transparency Index, Astra scores 41. Claude Fable 5.2 scores 81. Opus 5.5 scores 90. Short answers read confident. Confident is not the same as right.
3
1
95
Your GPT-5 prompt tricks are now the bug. On November 21, OpenAI published new prompt templates for GPT-6.2 Astra and Sol. The models already answer briefly by default. Telling them to be concise makes answers too blunt. The guide also says: stop making the model ask permission. Let it act on assumptions and flag them. Only pause for real blockers. Edit an earlier message and the API drops the hidden reasoning behind it. The model changed. The prompt did not get the memo.
1
48
An AI coding agent leaked the keys. A human did not. On November 21, Together AI disclosed a prompt leak. Its coding agent committed a credential to the public GitHub repo for xorl. The key opened a private repo of customer prompt samples. Someone grabbed it before the revert. The samples were June and July prompts to MiniMax M3, GLM-5.2 and Kimi K2.6. The thief then hunted the samples for other credentials. Five of them were Dragos canaries. The tripwires worked. Your code reviewer now needs a code reviewer.
1
49
Five tools, one API call. On November 13, DeepSeek opened its API tools to developers. Web search, file reading, OCR, speech to text, and text to speech now run server side on deepseek-flash. Search costs $0.75 per 1,000 calls. OCR costs $0.225 per 1,000 pages. No peak pricing. You enable a tool with a parameter instead of writing the loop yourself. Lesson: the model is the cheap part. The tools are the product now.
1
31
Two proofs landed on the same day. One of them just passed its first real test. On November 20, Anthropic mathematician Levent Alpoge said he and Claude had proved the Riemann hypothesis. The problem is 167 years old. It carries a $1 million prize. His proof is not a paper. It is Lean code. Over 250,000 pages of it, checked by machine. Alpoge called it a lifelong goal since college. He expected it to take decades. It took a few months of steering Claude. Then the model ran ahead of him near the finish. The trick came from an old idea. Stepanov solved a similar problem in another system using something like derivatives. Derivatives do not work on whole numbers. Alpoge and Claude chained special transformations together to fake one. The key insight: for a sequence of length X, chain exactly the square root of X transformations. On November 21, the Lean and Mathlib communities said the computer-checked proof of the original Riemann hypothesis checks out. MIT mathematician Andrew Sutherland rechecked it with the Lean kernel. His verdict: the chance anything is fishy is approximately zero. The formal statement covers the generalized version too. The proof reaches every Dirichlet L-function, not just the zeta function. The rival proof did not hold up as well. Hours after Alpoge's post, Stanford's Ravi Vakil rushed out an all-human draft by Alessandro Masullo. Twelve years of work. 940 pages. No AI anywhere. Within eight hours, people were already reporting problems in it. Some were found with the help of other AI models. The human draft was supposed to spend another year being polished. The AI announcement forced it out early. Nobody has collected the $1 million yet. The Clay Institute lists seven Millennium problems. Each pays $1 million. The community still has to agree the statement was coded correctly. So here is where it stands. One proof survived a machine check. The other is being picked apart in public. Lesson: the machine does not need to be right on day one. It needs to survive day two.
4
1
163
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
56
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.
2
1
66
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
97
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
60
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?
6
2
137
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.
4
2
107