Does gemini 4 even exist or are we hallucinating
I asked 4 models across 4 different harnesses to recreate themselves as a WALL-E character 4 models, the exact same prompt, and such different results: - muse spark: built such a cute, polished robot - deepseek: way more
The slightly broken output of opencode deepseek v4.1 flash used via Codex-router has been fixed, and it…
The slightly broken output of opencode deepseek v4.1 flash used via Codex-router has been fixed, and it is now fully upward compatible with gpt-5.6 luna, which is exciting. This speed is addictive. 200-300 tokens/sec.
Andrej karpathy could have charged $2,000 for this course.
He put it on YouTube. The full training stack. Tokenization. Neural network internals. Hallucinations. Tool use. Reinforcement learning. RLHF. DeepSeek. AlphaGo. 3 hours of the most comprehensive LLM education that
284B-class DeepSeek-V4.
Two 24GB 3090s. 18.26 tokens/s decode. The routed experts live in system RAM. The GPUs keep attention. That is the author’s own `llama-sweep-bench` on a hybrid `--cpu-moe` box — not an H100 rack, not a Discord screenshot. 🆕 ik_llama.cpp
You have reached the end of the archive
All of deepseek