Unverified50% confidenceEventExact time
A developer following Karpathy's nanochat project trained 9 checkpoints across 3 different versions of LLMs from scratch with a budget of approximately $750, with parameter counts ranging from 353M to 672M.
1
Sources
50%
Confidence
Long-term
Relevance
8/29/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedAndrej Karpathy open-sourced the nano-chat project, demonstrating the complete process of building a ChatGPT-like system from scratch at a GPU cost of approximately $100.68% similarVerifiedTraining a 7B model for 5 epochs on 4.7 million tokens of training data requires 4 A100 GPUs, with a direct compute cost of approximately 2,000–4,000 yuan67% similarUnverified训练GPT-4级别模型的算力成本估计超过5000万美元66% similarUnverified通义零码的MoE架构激活比例约为7.3%（350亿/4800亿），推理成本接近一个350亿参数的密集模型65% similarUnverifiedThe entire FiveM server-building experiment consumed approximately 150,000 tokens, and the creator purchased the Claude Max plan for this purpose at £75 per month (approximately 680 RMB).64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/819738API
curl https://kongchang.com/api/v1/knowledge/claims/819738MCP
get_claim(id=819738)