Sam Altman just said: there are only 3 paths for AI safety and this is fucking madness Path 1: an international body like the IAEA. Path 2: the biggest labs merge into 1 organization, CERN style, building the technology on behalf of participating countries. Path 3: 1 company dominates compute, a plan he reportedly called terrible. He also reportedly said OpenAI has to pick within the next year. Read this again. Open weights are not on it. Not as a path, not as a footnote. In every version a few institutions hold the model and everybody else gets access through them. A global regulator deciding who can train. A merged mega lab serving "participating countries". Or 1 company with all the GPUs. Your role in all 3 plans is the same: customer, if they allow it. Qwen, DeepSeek, GLM, MiniMax and Kimi are the only reason a 4th option exists today: the model is on your disk, and nobody has to approve you. I will take my 4x 3090 over any of their fucking plans.

China wants to put AI models on its export control list. Reuters, today: 40 technologies, 23 of them AI. Models, agents, compute clusters. Registration starts now, final rules after 15 Nov. My feed in 1 hour: "Chinese open weights are dead." Then I read the Ministry of Commerce Q&A. Already open sourced or published in papers: not covered. Open sourcing new AI tech: "automatic license". API, apps, SaaS, small on-device models: not export at all. Deploying a restricted model on servers abroad: export. So the models I run at home are fine. Today. But an automatic license is still a license. Today Beijing signs it automatically. Tomorrow it can stop signing. I see it like this. The best thing China gave the world this year was weights you can download. If the next Qwen Max or Kimi goes API only because of a list, China copies the worst thing OpenAI ever did. Comments are open until 15 Nov. I hope every Chinese lab that ships weights sends one.

You're mocking local AI while worrying about limits across SIX Claude accounts. You posted that six hours earlier. I run open models on 4 RTX 3090s. I can leave them working for as long as the hardware holds up, and I pay for the electricity. That matters when I have actual work to finish. Prefer Opus, fine. But having six accounts and still worrying about access is a fucking terrible advertisement for the subscription model you're defending. You already explained why people want their own compute. I barely need to add anything.

Claude is leaking your data to the US government and CIA. Here is the proof. I looked inside the Claude Code binary bundled with Claude Desktop on my Mac. Found these 5 .gov strings: c2s.ic. gov api. aws.ic. gov sc2s.sgov. gov csp.hci.ic .gov api .aws.hci.ic .gov C2S is tied to AWS cloud for the US intelligence community. In the official Amazon announcement, the CIA CIO discusses its use at the TOP SECRET level. You can check your own installation. Mine is Claude Desktop 2.26454.2, bundled Claude Code 2.1.293. These strings appear in the AWS region definitions, under dnsSuffix and dualStackDnsSuffix. WHAT THE FUCK ???!!!! Run the check that's in the replies on macOS. See what YOUR copy contains.

A 27B model in 12 GB of VRAM. Full 262K context. 70 tok/s. And now it is UNCENSORED. I ran it on my own RTX 3090, held to 12 GB. This is Mirai S Qwen3.8-27B, 2.4 bits per weight, an 11 GB file. With my llama.cpp port the same card did 82K context at 38 tok/s inside 12 GB. Then @CaryPalmerr took the port and did the hard part. Same card, same file, his serve: - context: 82K -> 262,144 with q8 KV - decode: 38 -> 69.9 tok/s - VRAM peak: 11.7 GB - prefill: 1,124 tok/s He keeps part of the KV cache in VRAM and the rest in pinned RAM, and he drafts with MTP at every depth. He checked every kernel against my fork token for token and wrote down every experiment, the failed ones too. On his real 12 GB RTX 4070 it does 76.6. My part is small. I added abliteration. You cant edit these weights, they are compressed codes, so I measured the refusal direction on the quantized model and the engine removes it while the model runs. A 1.3 MB file and 1 flag: - harmful prompts refused: 64 of 64 -> 0 of 64 - held-out set: 78 of 82 -> 0 of 82 - decode: 69.9 -> 69.7 tok/s Counted by a regex, and with thinking on 1 of 24 still refuses. Deep in the context it gets slower on my card: 27.9 tok/s at 60K. So thats 12 GB of VRAM, 262K context, 70 tok/s, abliterated. Open weights, open code, runs on your desk. Cary did a genius amount of work and almost nobody has seen it. Go follow @CaryPalmerr right now, I mean it. serve: weights + vector:

🤯 My RTX 3090 made THESE while I was asleep. I still cant believe this runs in my apartment. Qwen-Image-2.1 makes the picture. MiniMax H3 turns it into a 15 second clip, sound included, and then the H3 upscaler doubles the resolution. No cloud. No credits. No queue. Only 3 models on my SSD. I started it at 1am and by 7:40 there were 103 images and 63 clips on the disk. Thats about 1 minute per image and 11 minutes per clip. Every clip is 1 image plus 1 prompt and I didnt set keyframes or edit anything after, each one is a single continuous shot. The prompt says what should happen and H3 just does it. This is a studio on 24GB cards. More insane videos in replies:

Mistral Large 4. 1T parameters, 49B active. They call it Le Chonk and they are not joking. "The best open weights model from US or Europe." The weights come at the end of October. Today its an API. So right now its the best open weights model that has no weights. And when they come, who runs it? 1T parameters is about 500GB at 4 bits. I have 96GB of VRAM on 4x 3090 and I am already the rich kid in this hobby. Look at this week. Reflection Beam, 501B, weights "this month". Mistral, 1T, weights "end of October". The West discovered open weights all at once, and everybody built a model for a datacenter. I am glad Europe is back, I mean it. Aleph Alpha shipped 78B with 3.46B active 3 days ago and the file was there on day 1. Thats how you do it. Mistral, your 7B is how a lot of us started with local AI. Ship a small one too.

1 week ago Anthropic published a research post about GLM-5.3. The message: this model is a cyber danger because anyone can download it. Yesterday GLM-5.3 launched on Amazon Bedrock. Amazon is one of Anthropic's biggest investors and the cloud that hosts Claude. So the "dangerous" Chinese open model is now sold to enterprises by Anthropic's own partner, next to Claude, in the same catalog. You cant have both. Either this model is too dangerous to be released, or its a product your partner sells for money. I think its simple. Open weights were never the danger. They are the competition, and the safety post was written by the company that loses the most when a free model gets close.

What the AI industry did to its customers in the last 8 days: OpenAI cut the $200 Pro plan in half. OpenAI launched a $500 plan. OpenAI started watermarking your text. DGX Spark went from $3,999 at launch to $6,950. Micro Center started scanning ID to sell an RTX 5090. Micron sold 75% of its 2027 memory to someone who is not you. What happened to the things I own: Nothing. My 4x 3090 are the same cards. The Qwen weights on my SSD are the same file. Nobody emailed me about a new tier. I dont run local models because they are the smartest. I run them because its the only part of my AI setup where the rules didnt change this week. Download the weights. Buy the hardware. Do it before the next announcement.

OpenAI looks like it has a big fucking problem. SemiAnalysis worked out how much usage these subscriptions give you at API prices. Every $1 spent on Claude with Opus 5.5 gets you close to $60 worth. On ChatGPT with Sol 6.1, it's about $10. Claude gives you 5 to 6 times as much for the same money. Look at the past week. They're cutting the $200 Pro allowance in half. They've introduced a $500 plan. Codex limits get reset almost every day to keep people from leaving. Meanwhile, my feed is full of people cancelling Pro and switching to Claude. Cut your paying customers' limits, then keep handing out resets so they stick around. That's a weird way to behave if you've got plenty of compute. It looks to me like they're short on GPUs and trying to cover it up by changing the plans. I pay them $200 a month, and I'm getting the worst deal on that chart. I'm supposed to be happy because they occasionally reset my limits? And I'm not getting comfortable with Anthropic either. They're being generous ahead of an IPO. I wouldn't count on those allowances surviving it. Whichever subscription I choose, they still get to decide how much my money is worth next month.

"Local AI is not cheaper and not frontier" Correct, its not frontier. I never said it was, and I still pay for closed models myself. My point is different. Last week OpenAI cut the $200 Pro plan in half with one announcement. Same price, half the usage. Nobody asked the people paying. Nobody can do that to the model on my disk. Qwen runs today the same way it ran the day I downloaded it. No weekly limit, no new tier, no email about "efficiency". And every few months the 27B gets better on the card I already own. You are not paying for the frontier. You are paying for permission to use it, and the permission keeps getting smaller. I want a way out, and local is the only one that exists.

Micron already sold 75% of its 2027 memory. Data centers bought the RAM you needed to run a model at home. Now they rent that model back to you by the month. You can already see it in prices. DGX Spark launched at $3,999 and the same 128GB box is $6,950 now. Nvidia says the reason is memory supply. Micro Center wants your ID to sell you 1 RTX 5090. Every year local AI needs less hardware because the models get smaller and smarter. And every year that hardware gets harder to buy. I dont think thats a coincidence I should wait out. My 4x 3090 cost me about $6K. I couldnt build the same rig for that money today, and next year it will be worse. Buy your hardware while you still can!!! I am not selling my 3090s.

Your agent can now use the browser built into @DeepSeekHarness 🔥 My local Qwen3.8 Flash Next Abliterated already found me some great models on CivitaiRed with it. If you know what I mean. The regular DeepSeek V4.1 Flash refused to do that. Out of the box it couldnt. The browser is there for you only, the model gets 0 tools for it. So I made a plugin. Now the agent opens the tab itself, reads the page, clicks, types, scrolls and takes screenshots, and I watch every click happen right next to the chat. 17 tools. Playwright and Puppeteer didnt work for this. Electron shows the sidebar tab as a "webview" target and both libraries only list "page" targets. I connected them to the app and they found 1 page, the harness UI itself. So I went with raw CDP straight to the webview. It doesnt patch the app. I dont unpack app.asar, and the signature stays valid. The browser also forgets your logins on every restart, so I added a cookie vault. You export before you quit and import after. I tested it on a real login, 6 of 6 cookies came back and the site let me in. You have to start DSH with --remote-debugging-port=9222, and I only tested macOS on 0.2.0-rc.2. DeepSeek v4.1 Flash wrote most of the code from inside DSH, in 1 session with me. You install it from the Plugins page in DSH. Add plugin, type dsh-sidebar-browser-cdp, done.

Bernie, this already exists. Its called open weights. Qwen, DeepSeek and GLM are free to download. I run Qwen at home and nobody sends me a bill or tells me "you buddy are not from the right country" Thats AI that benefits everybody. A student with an old gaming card gets the same file as a company. He can study it, change it, build a product on it and keep it when the lab changes its mind. Now look at who does this. Alibaba, DeepSeek, Zai, Moonshot, MiniMax. The labs doing what you ask for are Chinese. And look at who gets the billions. Closed American labs that sell access by the month and cut it when they want. Those are the ones Washington invites to sign papers. If you want AI for all of humanity, fund and protect open weights. Everything else is a dogshit.

Claude said no. Qwen did the job. One of them costs money every month. The other one was free and I can keep it forever. And its not only about refusals. With Claude I get a "no" and thats the end of it. I cant see why, I cant change it, I cant appeal to anybody. Somebody in San Francisco decided what my question was allowed to be. With Qwen I have the weights. If the model refuses something stupid, people fix it. This week I released fixed uncensored Qwen3.8 Flash Next quants myself. Nobody had to approve that. I still use Claude for the hard stuff, its a great model. But I dont trust a tool that can tell me no and never explain. Explain to me again which one is the dangerous model.

A small teaser of my next big project. I'm doing it together with Opus 5.5 and Astra. Do you recognize what this is? This is still a very early stage. It will be significantly better, but I just can't tear myself away from the game. Sound strictly on