John Schulman on scaling RL and research institutions

The following is a summary of the key points from the interview with Sam Altman (assumed to be the AI researcher mentioned in the document):

Computing Power vs. Technical Methods:
Sam believes that AI progress is not solely dependent on computing power; technical method optimization (e.g., post-training techniques, fine-tuning data construction) is equally critical. For example, the birth of ChatGPT may have occurred earlier than the current computing power level, indicating that technological breakthroughs require appropriate prerequisites.
AGI Timeline Prediction:
Researchers generally underestimate the time required to achieve AGI; predictions should be extended by 2-3 times (e.g., the case of autonomous driving). However, the self-accelerating effect of AI may shorten the timeline, leading to uncertainty in predictions.

Tinker’s Core Positioning:
As a foundational fine-tuning API, Tinker provides small primitives (e.g., training and sampling tasks), allowing users to directly call them via Python scripts without handling the complexity of GPUs/distributed systems, thereby lowering the threshold for AI model fine-tuning.
Future Plans:
- Support multi-modal training (text, image, audio, video).
- Expand into a full-stack solution, enabling non-experts to rapidly build customized models.
- Lower the threshold for AI entrepreneurship, allowing entrepreneurs to focus on model innovation rather than底层 computing power issues.

AI Tools for Research:
- Use tools like Cursor, Claude Code to program, generating code snippets or supplementing research content.
- Efficiency in literature retrieval and open-source library searches is significantly improved.
Work Routine:
- Prefers thinking in cafes, away from distractions, to record inspirations.
- During project execution, deep reading of code/documentation is required to guide others in research.

Surge in Researchers but Stable Breakthroughs:
- The rigor of experiments in 70-90s papers was lower, whereas today’s standards are higher (e.g., multi-task benchmark tests).
- Scaling and engineering drive domain maturity, eliminating the need to write code from scratch, relying instead on existing tools and codebases.
Shift in Skill Requirements:
- Engineering skills have become more critical, while research taste has relatively declined (simple ideas can achieve significant results through scaling).

Technical Methods Take Center Stage:
Sam emphasizes that the future of AI depends more on optimizing technical methods and research taste rather than purely relying on computing power.
Revival of Classic Concepts:
Forgotten theories (e.g., early reinforcement learning methods) may regain attention due to new demands or technological maturity.

Sam’s interview highlights the multidimensional development path of AI technology:

Tool Innovation (e.g., Tinker) lowers technical barriers, driving application implementation;
Research Method Optimization (e.g., experimental rigor, engineering) enhances domain maturity;
Balancing Computing Power and Technology, avoiding over-reliance on a single factor.
His insights provide practical directions and thinking frameworks for AI researchers, developers, and entrepreneurs.

以下是关于Sam Altman（假设为文档中提到的AI研究者）的访谈内容核心要点总结：

算力 vs 技术方法：
Sam认为AI的进步并非单纯依赖算力，技术方法优化（如后训练技术、微调数据构建）同样关键。例如，ChatGPT的诞生可能早于当前算力水平，说明技术突破需要合适的前提条件。
AGI时间线预测：
研究人员普遍低估了实现AGI的时间，需将预测时间延长2-3倍（如自动驾驶的案例）。但AI的自我加速效应可能缩短时间线，导致预测存在不确定性。

Tinker的核心定位：
作为底层微调API，Tinker提供小型原语（如训练和采样任务），用户可通过Python脚本直接调用，无需处理GPU/分布式系统复杂性，降低AI模型微调门槛。
未来规划：
- 支持多模态训练（文本、图像、音频、视频）。
- 扩展为全栈解决方案，使非专家用户也能快速构建定制模型。
- 降低AI创业门槛，让创业者专注于模型创新而非底层算力问题。

AI工具辅助研究：
- 使用Cursor、Claude Code等工具编程，通过AI生成代码片段或补充研究内容。
- 文献检索和开源库查找效率大幅提升。
工作日常：
- 喜欢在咖啡馆思考，远离干扰，记录灵感。
- 项目执行阶段需深入阅读代码/文档，指导他人研究。

研究者数量激增但突破速度稳定：
- 70-90年代的论文实验严谨性较低，而如今标准更高（如多任务基准测试）。
- 规模化和工程化推动领域成熟，无需从零编写代码，依赖现有工具和代码库。
技能需求变化：
- 工程技能重要性上升，研究品味相对下降（简单想法通过规模化可取得显著成果）。

Sam的访谈强调了AI技术的多维发展路径：

https://www.youtube.com/watch?v=29BYxvvF1iM