
Alex
260 posts





Not my protocol “My Cancer Story! I was diagnosed with stage 4 pancreatic cancer in August 2023, and without treatment, I had 3-6 months to sort out my affairs. I began my fenbendazole Protocol immediately. I nearly quit up since my quality of life had deteriorated dramatically. But I started to feel better. Less nausea. I had gained some weight back! Increased energy. My scan in November 2023 indicated that this severe cancer had neither grown or spread. My cancer marker dropped from over 100,000 to 35,000. As of January 2024, my new marker number was 18k!!! My oncologist was just scratching his head, and a family member said he seemed bewildered! And no, I did not inform him about the fenben. I was terrified he'd drop me as a patient. I was quite appreciative for fenben and how it has helped me. I was feeling better and stronger. I had scans and bloodwork done in March, and my markers are now down to 6000. And the tumors were shrinking!!! Two weeks after my previous scan. I scanned again, and it returned NED. THIS WAS MY PROTOCOL; •MORNING ◦ Curcumin (600mg daily) ◦ Zinc (50mg) ◦ Milk thistle, As a food supplement, take 15 - 30 drops, 2-3 times daily in a little fruit juice or water. 7 days a week ◦ Serrapeptase (120,00Iu) ◦ Fenbendazole (1000mg of Panacur C is advised to be taken seven days a week. It is recommended that it should be taken with a meal). •NIGHT ◦ Curcumin (600mg tablet per day are recommended ◦ Berberine (600mg 2-3 times a day ) ◦ Quercetin (500mg 1/day) ◦ Turkey Tail ◦ Vitamin E (800mg for 7 days a week) ◦ Fenbendazole (1000mg at night) ◦ Ivermectin 12mg daily 5/7days a week •Limit sugar and processed food intake •Drank green tea often” x.com/toobaffled/sta…




Silicon Valley spent $500 billion building closed AI. China just gave it away for free. And it's winning. A $0.60 model from Beijing just beat Claude 4.7 and GPT-5.5 on real coding benchmarks. Read Below. The numbers will shock you. ↓ Kimi K2.6 vs the entire American AI stack: > SWE-Bench Pro: Kimi 58.6%. GPT-5.4: 57.7%. Claude Opus: 53.4%. > HLE with Tools: Kimi 54.0%. GPT-5.4: 52.1%. Claude: 53.0%. > DeepSearchQA: Kimi 92.5%. GPT-5.4: 78.6%. A free, open-source Chinese model just beat the most expensive AI products in America. At 8x lower cost than Claude. ↓ The price war America can't win. 1 million requests per year: > Kimi K2.6: $13,800 > GPT-5.4: $56,500 > Claude Opus: $150,000 Same task. 10x the cost. For models Kimi already beats on coding. DeepSeek V3.2 charges $0.28/M tokens under MIT license. Anthropic charges $25/M. That's 89x more expensive for a model that loses on real benchmarks. ↓ Why Kimi and DeepSeek are the real show stoppers. > 1 trillion parameters. Open weights. Self-hostable. > Agent Swarm: 300 sub-agents in parallel. 12-hour autonomous coding. > DeepSeek V3.2 delivers 90% of GPT-5.4 quality at $0.28/M. > GLM-5.1 hits 94.6% of Claude's coding score for $3/month. Every Chinese flagship is open-weight. Every American flagship is locked behind an API. ↓ The shift nobody's saying out loud. When DeepSeek shipped R1 in 2025, US AI markets lost $1 trillion in a single day. When Kimi K2.6 shipped on April 20, 2026, the developer world had the same reaction. China didn't catch up. China rewrote the economics of the entire industry. Open weights. Lower cost. Production-scale agents. MIT licenses. This isn't a benchmark war. It's a business model war. And open source just won round one. Save this post. Follow @codewithimanshu for the AI shift Silicon Valley doesn't want you to see clearly.

🚨 ELON’S REAL SPACE COMPANY MAY NOT BE SPACEX Eric Weinstein just dropped a theory that sounds insane… Until you think about it. SpaceX is the rocket company. But Mars doesn’t just need rockets. Mars needs autonomous factories. Robot labor. Energy systems. Global communications. AI decision-making. A machine civilization that can operate before humans fully arrive. That’s where it gets strange. Tesla. Starlink. xAI. Grok. Optimus. SpaceX. These may not be separate companies. They may be pieces of one Mars operating system.





I trained a 12M parameter LLM on my own ML framework using a Rust backend and CUDA kernels for flash attention, AdamW, and more. Wrote the full transformer architecture, and BPE tokenizer from scratch. The framework features: - Custom CUDA kernels (Flash Attention, fused LayerNorm, fused GELU) for 3x increased throughput - Automatic WebGPU fallback for non-NVIDIA devices - TypeScript API with Rust compute backend - One npm install to get started, prebuilt binaries for every platform Try out the model for yourself: mni-ml.github.io/demos/transfor… Built with @_reesechong. Check out the repos and blog if you want to learn more. Shoutout to @modal for the compute credits allowing me to train on 2 A100 GPUs without going broke cc @sundeep @GavinSherry







