World's Fair 2025
Latent Space Paper Club: AIEWF Special Edition (Test of Time, DeepSeek R1/V3) — VIbhu Sapra
Overview
This talk reviews recent advancements in AI research, focusing on the DeepSeek models and the evolution of training methodologies. It introduces a new "Test of Time" paper club initiative aimed at systematically covering foundational AI concepts. The discussion highlights how reinforcement learning and extended inference time are enabling models to develop advanced reasoning capabilities, leading to significant performance improvements.
Who should watch
- AI engineers interested in the latest open-source model developments.
- Researchers exploring advancements in model reasoning and training techniques.
- Builders looking to understand foundational AI papers and concepts.
- Product managers evaluating the capabilities of current LLMs.
Key takeaways
- The DeepSeek R1 model, an update to a previous version, now achieves performance comparable to OpenAI's GPT-4 and Google's Gemini 2.5, particularly in reasoning, math, and coding tasks.
- A significant improvement in DeepSeek R1 is its ability to perform reasoning over twice as many tokens compared to its predecessor, enabling more complex problem-solving.
- A distilled version of DeepSeek, the Quen 3 8B model, now matches the performance of much larger models like Quen's 235B reasoning model, demonstrating the effectiveness of distillation from advanced reasoning models.
- The "Test of Time" paper club is launching as a curriculum-based initiative to cover foundational AI papers, running from July to December, with both in-person and remote sessions.
- Advancements in training, particularly through reinforcement learning (RL) and extended inference time, are crucial for unlocking emergent reasoning abilities and "aha moments" in LLMs.
- The original DeepSeek paper introduced a four-stage training process for its R1 model, combining supervised fine-tuning (SFT), RL for reasoning, rejection sampling, and further RL polishing.
- The development of reasoning models has shifted focus from simply predicting the next token to enabling models to dynamically spend more compute on complex queries during inference.
- The DeepSeek team emphasizes the power of pure RL in developing reasoning capabilities without explicit supervised data, allowing models to autonomously discover problem-solving strategies.
Notable quotes
*The Deepseek team says our goal is to explore the potential of LLMs to develop reasoning capabilities without any supervised data focusing on their self-evolution through a pure RL process.*
*This moment is not only an aha moment for the model but also for the researchers observing its behavior. It underscores the power and beauty of reinforcement learning.*
Unofficial community note. Prefer the recording for nuance.