Company / organization
DeepSeek
By and about DeepSeek · page 3
Apple's "The Illusion of Thinking" and the Rebuttal
Apple's research paper argues reasoning models like o1 and o3 aren't truly thinking but failing due to context limits, though a rebuttal shows o3 solves Tower of Hanoi when prompted to be concise.
newsDeepSeek Drops Best End-to-End Large Model Training Paper
DeepSeek released a comprehensive paper on end-to-end large language model training methodology.
newsBaidu ERNIE-4.5 — DeepSeek R1 level at half the price 👸
Baidu released ERNIE-4.5, matching DeepSeek R1's performance at half the cost while claiming benchmark wins over GPT-4o in multimodal tasks and GPT-4.5 in some text benchmarks.
newsManus: DeepResearch + Operator agent from China
Chinese startup Butterfly Effect released Manus, an autonomous AI agent powered by Claude 3.5 and Qwen that performs real-world tasks like building websites and analyzing stocks, outperforming OpenAI's Deep Research on the GAIA benchmark.
newsAI Dinner 7.0 — DeepSeek Workshop
AI Socratic hosts AI Dinner 7.0 on February 20th in Manhattan to discuss DeepSeek R1 research papers and install the model on attendees' laptops.
newsGemini 2.0 Flash — Cheapest Model Yet
Google released Gemini 2.0 Flash at $0.40 per million tokens with a 1M-token context window, enabling processing of thousands of PDFs for under $1.
newsCerebras — Fastest Inference Platform
Cerebras achieves 1,200 tokens/second inference speed on DeepSeek-R1-Distill-Llama-70B, claiming 10x faster performance than comparable models and 3x faster than Groq.
newsSFT Memorizes, RL Generalizes
Supervised fine-tuning memorizes training examples while reinforcement learning generalizes to new problems, a distinction DeepSeek's approach to model training highlights.