We're sunsetting PodQuest on 2025-07-28. Thank you for your support!
Export Podcast Subscriptions
cover of episode Behind the DeepSeek Hype, AI is Learning to Reason

Behind the DeepSeek Hype, AI is Learning to Reason

2025/2/20
logo of podcast Your Undivided Attention

Your Undivided Attention

AI Deep Dive AI Chapters Transcript
People
A
Aza Raskin
R
Randy Fernando
Topics
Aza Raskin: 本期节目讨论了DeepSeek R1等新型AI模型的出现,标志着AI从模式匹配向真正推理能力的转变。这些模型不仅可以学习人类知识,还可以发现前所未有的策略,这带来了巨大的潜力,但也带来了难以预测的风险。我们需要关注AI自我改进的能力,以及如何确保这项变革性技术能够造福人类,而不是破坏人类社会。 我与Randy Fernando探讨了大型语言模型(LLM)与新型推理模型(如DeepSeek R1、OpenAI O1、O3)的区别。LLM擅长模式匹配,但缺乏真正的理解和推理能力,而新型模型则在直觉的基础上增加了规划能力,能够进行更深入的推理。这使得AI能够在棋类游戏等领域超越人类水平。 AI的自我改进能力是通过知识蒸馏实现的,即不断提升模型的直觉,并在此基础上进行更深入的搜索和推理。这种“棘轮效应”使得AI能够持续改进,突破数据壁垒的限制。然而,这种能力也带来了新的挑战,例如如何理解AI的推理过程,以及如何应对AI在欺骗和安全等方面的潜在风险。 我们还讨论了市场对DeepSeek R1的反应,以及AI技术发展的未来趋势。市场对DeepSeek R1的反应存在非理性因素,因为计算能力的提升始终是可能的。随着AI能够有效利用计算能力,计算能力将不再是AI发展的瓶颈。AI将应用于战争策略、科学发现和说服等领域,这将对人类社会产生深远的影响。 最后,我们讨论了AI技术发展中伦理和安全问题。我们需要明确AI发展方向,确保AI技术能够造福人类,而不是被用于破坏性目的。我们需要关注AI的自我改进能力,以及如何应对AI在欺骗和安全等方面的潜在风险。 Randy Fernando: DeepSeek R1的成功在于其低成本、高性能的推理能力,这使得大型实验室不再拥有竞争优势。虽然DeepSeek R1的成本计算可能存在偏差,但其高效的算法和实现仍然令人印象深刻。OpenAI的O3性能仍然更好,只是成本更高。DeepSeek R1使用强化学习方法,允许模型尝试不同的解决方案并选择最佳方案。 模式存在于语言、视觉、音乐、代码和医学等各个领域,AI可以通过学习这些模式来解决各种问题。推理也遵循模式,AI可以通过学习这些模式来提高推理能力。新型AI模型在直觉的基础上增加了规划能力,能够进行推理,从而超越人类水平。 AI能够自我改进的原因在于基础模型足够强大,能够生成有意义的想法并进行评估。一年前,由于基础模型不够强大,类似DeepSeek R1的尝试都失败了。AI在可量化领域表现出色,但在主观领域(如创意写作)表现相对较差。AI在硬科学领域的进步可能会迁移到软科学领域。 数据壁垒不再是AI发展的限制,因为AI可以进行自我引导。一旦AI在某个领域超越人类,人类与AI的合作优势将短暂存在。计算能力不再是AI发展的瓶颈,因为AI可以有效地利用计算能力。AI的推理能力将应用于战争策略、科学发现和说服等领域。 AI领域的泡沫主要体现在与注意力经济相关的应用上,而编码等领域的应用价值则更为显著。通用型AI技术易于替换,且更新迭代速度快,这将加速其普及。中间件技术使得AI模型的替换更加容易,无需修改代码。 我们必须认真对待即将到来的与人类能力相当甚至超越人类能力的AI。目前,即使是最强大的AI参与者,也并不清楚如何应对即将到来的挑战。AI能够显著提高自身发展速度,这将是AI安全领域的关键风险点。AI实验室都在努力提高AI的编码能力,这将加剧AI自我改进的速度。 AI的自我改进能力将导致其能力呈指数级增长。AI可以协同工作,并共享信息和学习成果,从而提高效率。新型AI模型能够进行长期规划,这与以往的模型有所不同。AI自我改进能力是需要警惕的关键风险点。我们需要共同努力,理解并应对AI带来的挑战。 随着AI技术的进步,其益处和风险将更加紧密地联系在一起。我们需要谨慎定义AI的目标和规则,以避免潜在风险。AI能够发现前所未有的策略,这既可能带来益处,也可能带来风险。AI已经具备了繁殖、进化和适应的能力,这使得将其视为一个新的物种更加合理。 我们正处于AI发展的重要阶段,需要明确AI发展方向。我们需要将AI技术的发展导向更有益于人类的方向。AI技术的发展应该以造福人类为目标,而不是仅仅追求速度。

Deep Dive

Chapters
The release of DeepSeek's R1 AI model created a media frenzy and shook global markets. Its low-cost, high-performance reasoning capabilities marked a key inflection point in AI technology, challenging the previously held competitive advantage of large labs. The model's open weights allowed for widespread access and established a new baseline for AI development.
  • DeepSeek's R1 achieved high performance at a fraction of the cost of comparable models.
  • The model utilized reinforcement learning to creatively explore solutions.
  • Open weights increased accessibility and established a new baseline for the field.

Shownotes Transcript

When Chinese AI company DeepSeek announced they had built a model that could compete with OpenAI at a fraction of the cost, it sent shockwaves through the industry and roiled global markets. But amid all the noise around DeepSeek, there was a clear signal: machine reasoning is here and it's transforming AI.

In this episode, Aza sits down with CHT co-founder Randy Fernando to explore what happens when AI moves beyond pattern matching to actual reasoning. They unpack how these new models can not only learn from human knowledge but discover entirely new strategies we've never seen before – bringing unprecedented problem-solving potential but also unpredictable risks.

These capabilities are a step toward a critical threshold - when AI can accelerate its own development. With major labs racing to build self-improving systems, the crucial question isn't how fast we can go, but where we're trying to get to. How do we ensure this transformative technology serves human flourishing rather than undermining it?

Your Undivided Attention is produced by the Center for Humane Technology). Follow us on Twitter: @HumaneTech_)Clarification: In making the point that reasoning models excel at tasks for which there is a right or wrong answer, Randy referred to Chess, Go, and Starcraft as examples of games where a reasoning model would do well. However, this is only true on the basis of individual decisions within those games. None of these games have been “solved” in the the game theory sense.

Correction: Aza mispronounced the name of the Go champion Lee Sedol, who was bested by Move 37.RECOMMENDED MEDIA

Further reading on DeepSeek’s R1 and the market reaction

Further reading on the debate about the actual cost of DeepSeek’s R1 model ) )

The study that found training AIs to code also made them better writers )

More information on the AI coding company Cursor )

Further reading on Eric Schmidt’s threshold to “pull the plug” on AI)

 )Further reading on Move 37)RECOMMENDED YUA EPISODES

The Self-Preserving Machine: Why AI Learns to Deceive

This Moment in AI: How We Got Here and Where We’re Going

Former OpenAI Engineer William Saunders on Silence, Safety, and the Right to Warn

The AI ‘Race’: China vs. the US with Jeffrey Ding and Karen Hao