https://shulai.ru/https://shulai.ru/about/https://shulai.ru/contacts/https://shulai.ru/legal/https://shulai.ru/posts/01-tokenization-1-words-and-tokens/https://shulai.ru/posts/01-tokenization-2-bpe/https://shulai.ru/posts/01-tokenization-3-bpe-in-translation/https://shulai.ru/posts/02-embeddings-1-what-is-embedding/https://shulai.ru/posts/02-embeddings-2-cooccurrence-cosine/https://shulai.ru/posts/02-embeddings-3-word2vec/https://shulai.ru/posts/03-self-attention-1-why-attention/https://shulai.ru/posts/03-self-attention-2-qkv/https://shulai.ru/posts/03-self-attention-3-causal-mask/https://shulai.ru/posts/04-multihead-1-heads/https://shulai.ru/posts/04-multihead-2-positional/https://shulai.ru/posts/04-multihead-3-input-matrix/https://shulai.ru/posts/05-block-1-residual-stream/https://shulai.ru/posts/05-block-2-ffn-layernorm/https://shulai.ru/posts/05-block-3-stacking/https://shulai.ru/posts/06-architectures-1-three-types/https://shulai.ru/posts/06-architectures-2-encoder-bert/https://shulai.ru/posts/06-architectures-3-encoder-decoder/https://shulai.ru/posts/07-pretraining-1-self-supervised/https://shulai.ru/posts/07-pretraining-2-cross-entropy/https://shulai.ru/posts/07-pretraining-3-perplexity/https://shulai.ru/posts/08-scaling-1-power-laws/https://shulai.ru/posts/08-scaling-2-compute-allocation/https://shulai.ru/posts/08-scaling-3-chinchilla/https://shulai.ru/posts/09-moe-1-idea/https://shulai.ru/posts/09-moe-2-switch-routing/https://shulai.ru/posts/09-moe-3-scaling-flops/https://shulai.ru/posts/10-inference-1-kv-cache/https://shulai.ru/posts/10-inference-2-flashattention/https://shulai.ru/posts/10-inference-3-mamba-ssm/https://shulai.ru/posts/11-sft-1-instruction-tuning/https://shulai.ru/posts/11-sft-2-loss-masking/https://shulai.ru/posts/11-sft-3-data-and-template/https://shulai.ru/posts/12-lora-1-idea/https://shulai.ru/posts/12-lora-2-low-rank/https://shulai.ru/posts/12-lora-3-qlora/https://shulai.ru/posts/13-rlhf-1-three-stages/https://shulai.ru/posts/13-rlhf-2-reward-model/https://shulai.ru/posts/13-rlhf-3-results/https://shulai.ru/posts/261018773-jeva-system-one-models-from-first-principles/https://shulai.ru/posts/261018829-omnicompass-a-holistic-benchmark-for-omni-modal-video-grounding-and-understanding/https://shulai.ru/posts/261020424-bizbench-unveiling-hidden-challenges-in-deploying-long-horizon-agents-for-real-world-business/https://shulai.ru/posts/261021114-spectral-spatial-interpretation/https://shulai.ru/posts/261022276-mjev-spatial-system-one-spatial-intelligence-via-native-attention-matching/https://shulai.ru/posts/261028402-streameval-evaluating-interactive-streaming-video-generation-models/https://shulai.ru/posts/all-5-tencent-hy-simulmt-30b-a3b-fp8/https://shulai.ru/posts/all-9-pika-labs-pika-h3-flex/https://shulai.ru/posts/all-10-ccore-qwen3-5-4b-3d-coder/https://shulai.ru/posts/all-10-speridlabs-idol/https://shulai.ru/posts/anthropic-expanding-partnership-gsa/https://shulai.ru/posts/anthropic-s1-sec/https://shulai.ru/posts/asr-1-nvidia-nvidia-nemotron-3-5-transcribe-30b-a4b/https://shulai.ru/posts/asr-7-third-intelligence-haas-ja-v1-0/https://shulai.ru/posts/asr-8-ibm-granite-granite-speech-5-0-2b/https://shulai.ru/posts/asr-9-conwayresearch-underdog-bark-3-0-pro/https://shulai.ru/posts/asr-9-nvidia-nemotron-labs-unified-asr-0-6b/https://shulai.ru/posts/asr-9-openai-whisper-large-v3/https://shulai.ru/posts/asr-10-atomicchat-nvidia-nemotron-3-5-transcribe-30b-a4b-gguf/https://shulai.ru/posts/asr-10-ibm-granite-granite-speech-5-0-470m-turboctc/https://shulai.ru/posts/gemini-3-8-flash/https://shulai.ru/posts/gemini-4-carbon/https://shulai.ru/posts/google-ai-introducing-gemini-4-carbon/https://shulai.ru/posts/google-ai-introducing-genie-4/https://shulai.ru/posts/google-ai-project-suncatcher-satellite-response/https://shulai.ru/posts/google-deepmind-gemini-4-carbon-our-next-frontier-model/https://shulai.ru/posts/google-deepmind-genie-4-a-new-generation-of-world-simulation/https://shulai.ru/posts/gpt-6-1-terra/https://shulai.ru/posts/hf-all/https://shulai.ru/posts/hf-asr/https://shulai.ru/posts/hf-image-text-to-text/https://shulai.ru/posts/hf-t2i/https://shulai.ru/posts/hf-text-generation/https://shulai.ru/posts/hf-tts/https://shulai.ru/posts/image-text-to-text-4-mistralai-mistral-large-4-1t-a52b/https://shulai.ru/posts/image-text-to-text-6-baidu-unlimited-ocr-2/https://shulai.ru/posts/image-text-to-text-8-liquidai-lfm3-vl-2b/https://shulai.ru/posts/image-text-to-text-9-sc117-qwen3-8-flash-next-gsq-rco-abliterated-gguf/https://shulai.ru/posts/image-text-to-text-10-ista-daslab-qwen3-8-27b-gsq-rco-gguf-2/https://shulai.ru/posts/image-text-to-text-10-ista-daslab-qwen3-8-27b-gsq-rco-gguf/https://shulai.ru/posts/image-text-to-text-10-liquidai-d1-2b/https://shulai.ru/posts/image-text-to-text-10-openbmb-minicpm-ocr-1-0/https://shulai.ru/posts/minimax-m3-1-flash/https://shulai.ru/posts/mistral-mistral-large-4-oss/https://shulai.ru/posts/nvidia-local-ai-rtx-spark-laptops-updates-models-agents-games/https://shulai.ru/posts/openai-building-scientific-understanding-with-academic-communities/https://shulai.ru/posts/openai-community-safety-report-and-new-protections-for-teens/https://shulai.ru/posts/openai-helping-websites-express-ai-training-preferences/https://shulai.ru/posts/openai-how-we-detect-frontier-model-misalignment/https://shulai.ru/posts/openai-introducing-gpt-6-1-terra/https://shulai.ru/posts/openai-openai-servicenow-partnership-expansion/https://shulai.ru/posts/openai-oracle-apps-in-chatgpt/https://shulai.ru/posts/openai-updated-model-spec/https://shulai.ru/posts/t2i-8-krea-krea-2-turbo/https://shulai.ru/posts/text-generation-7-liquidai-lfm3-1-7b/https://shulai.ru/posts/tts-9-maya-research-maya2-native/https://shulai.ru/posts/wed-artificial-analysis/