Podcast Atlas

Vol.114 AI的2025和DeepSeek们的未来 | 对谈复旦张奇教授

source Episode summary Updated 2026-08-06 Tags: Podcast, Ai, Deepseek, Agents

Summary

This [[QizhulouYanBinke|起朱楼宴宾客]] episode has [[DavidWeng|大胃翁]] interview [[ZhangQi|张奇]] of [[FudanUniversity|复旦大学]] for a 2024 AI review and 2025 outlook shaped by DeepSeek’s sudden public breakout. Zhang treats DeepSeek as an important engineering and cost-efficiency signal, but uses it to sharpen rather than abandon the LLM Statistical Boundary: current large models remain data-driven statistical systems with strong long-text, cross-language, multitask, and generation capabilities, but weak causal understanding. The source’s practical conclusion is that 2025 opportunity lies less in generic foundation-model startups than in Scenario-Specific AI, Model Post-Training Bottleneck execution, Contact Center AI, AI search, and Agentic Workflow systems that can reflect, recover, and act in bounded settings.

Key Claims

  • [[ZhangQi|张奇]] is introduced as a [[FudanUniversity|复旦大学]] computer-science professor, doctoral adviser, Shanghai intelligent-information-processing lab deputy director, and MOSS model lead, giving the episode a more academic AI perspective than many product or investment conversations.
  • The episode frames DeepSeek as exciting because of user experience, training-cost compression, and compute-efficiency implications, while warning that those achievements do not by themselves prove near-term AGI.
  • Zhang argues that the bottom layer of today’s large models is still statistical machine learning, so their visible generality should be understood through accumulated scene coverage rather than a human-like unified mind.
  • The source names four robust large-model capabilities: long-context handling, cross-language transfer, multitask handling, and generation.
  • LLM Statistical Boundary is illustrated through examples such as a model doing difficult math but failing to count letters in “strawberry”, or performing well on one region’s exam style but poorly on another’s.
  • Model Post-Training Bottleneck is presented as harder than pretraining publicity suggests: if knowledge is already in pretraining, small matched datasets may unlock behavior; if not, supervised fine-tuning can fail or disturb behavior.
  • DeepSeek’s MoE-style cost structure is treated as an engineering advantage, but the episode stresses that post-training, reinforcement learning, and high-quality expert labeling remain expensive and under-disclosed parts of frontier-model work.
  • Scenario-Specific AI is the product lesson: focused tools such as Cursor or Perplexity can outperform general chatbots because they optimize around a concrete task, data form, and user expectation.
  • Prompt engineering is described as less central than many 2023 narratives implied; differences often come from training, product scenario, and tool design rather than prompt wording alone.
  • Overreliance on AI tools may degrade language, writing, or programming muscles, so the episode keeps Human Judgment Under AI and AI Use Pacing beside productivity gains.
  • Zhang links o1/o3-style reasoning to Interleaved Thinking: models can improve by trying paths, reflecting, and revising rather than answering once.
  • The clearest 2025 direction is Agentic Workflow, but Zhang distinguishes real agents from ordinary workflow/RPA systems by reflection, self-correction, decision loops, and the ability to recover after an action fails.
  • Contact Center AI and phone-service automation are treated as strong incumbent-company opportunities because speech, large models, transfer interfaces, compliance, and management systems all matter.
  • AI search is presented as a new product species because large models can use long text, cross-language, multitask, and generation abilities together, while incumbent search companies risk cannibalizing traffic and ad revenue.
  • The source predicts that AI researchers and algorithm engineers maintaining many small models may be exposed earlier than system architects, because large models can compress multiple narrow model teams while large production systems still need engineering judgment.

Key Quotes

“AI 的核心是场景” — Zhang’s product-level correction to broad industry labels.

“长文本、跨语言、多任务和生成” — the episode’s compact list of current large-model strengths.

“预训练只是万里长征第一步” — the warning that the harder work continues after base-model training.

“不是真正有自我意识的智能体” — the episode’s restraint around the DeepSeek enthusiasm.

Connections

Contradictions

  • No direct contradiction with existing wiki pages found.
  • The source qualifies Vol.111 关于2025年的四个猜想: vol.111 treats DeepSeek as a speculative signal of China’s divergent AI route, while this episode, recorded after DeepSeek’s breakout, treats it as a real engineering shock but still not as proof that current large models have crossed into AGI.
  • The source adds a tension to optimistic Frontier Model Scaling narratives by arguing that current models’ four strengths are already visible alongside a causality ceiling; this is compatible with, but more skeptical than, pages that emphasize continuing scaling and agent progress.