All tags
Person: "yitayml"
not much happened today
gpt-5-pro gemini-2.5 vllm deepseek-v3.1 openai google-deepmind microsoft epoch-ai-research togethercompute nvidia mila reasoning reinforcement-learning inference speculative-decoding sparse-attention kv-cache-management throughput-optimization compute-efficiency tokenization epochairesearch yitayml _philschmid jiqizhixin cvenhoff00 neelnanda5 lateinteraction mgoin_ blackhc teortaxestex
FrontierMath Tier 4 results show GPT-5 Pro narrowly outperforming Gemini 2.5 Deep Think in reasoning accuracy, with concerns about problem leakage clarified by Epoch AI Research. Mila and Microsoft propose Markovian Thinking to improve reasoning efficiency, enabling models to reason over 24K tokens with less compute. New research suggests base models inherently contain reasoning mechanisms, with "thinking models" learning to invoke them effectively. In systems, NVIDIA Blackwell combined with vLLM wins InferenceMAX with significant throughput gains, while Together AI's ATLAS adaptive speculative decoding achieves 4× speed improvements and reduces RL training time by over 60%. SparseServe introduces dynamic sparse attention with KV tiering, drastically improving throughput and latency in GPU memory management.
not much happened today
kimi-k2 qwen3-235b-a22b qwen3-coder-480b-a35b gemini-2.5-flash-lite mistral-7b deepseek-v3 moonshot-ai alibaba google google-deepmind openai hugging-face vllm-project mixture-of-experts agentic-ai model-optimization model-training benchmarking code-generation long-context multimodality math reinforcement-learning model-architecture model-performance open-source alignment demishassabis rasbt alexwei_ yitayml
Moonshot AI released the Kimi K2, a 1-trillion parameter ultra-sparse Mixture-of-Experts (MoE) model with the MuonClip optimizer and a large-scale agentic data pipeline using over 20,000 tools. Shortly after, Alibaba updated its Qwen3 model with the Qwen3-235B-A22B variant, which outperforms Kimi K2 and other top models on benchmarks like GPQA and AIME despite being 4.25x smaller. Alibaba also released Qwen3-Coder-480B-A35B, a MoE model specialized for coding with a 1 million token context window. Google DeepMind launched Gemini 2.5 Flash-Lite, a faster and more cost-efficient model outperforming previous versions in coding, math, and multimodal tasks. The MoE architecture is becoming mainstream, with models like Mistral, DeepSeek, and Kimi K2 leading the trend. In mathematics, an advanced Gemini model achieved a gold medal level score at the International Mathematical Olympiad (IMO), marking a first for AI. An OpenAI researcher noted their IMO model "knew" when it did not have a correct solution, highlighting advances in model reasoning and self-awareness.