Agents Over Bubbles | Stratechery by Ben Thompson

节目时间轴

1. 0:02

   Ben Thompson introduces the paradox of AI prognostication and argues that despite bubble fears, the rise of agents means we are not in a bubble.

   • Thompson notes a weird paradox in AI prognostication: pressure to take doomsday scenarios seriously on one hand, and pressure to validate bubble fears on the other, though he has long argued bubbles can be good.
   • Sitting in March 2026 on the morning of Nvidia's GTC, he has come to a different conclusion: he doesn't think we're in a bubble, which paradoxically may be the truest evidence that we are.
   • He frames the episode around three LLM inflection points—ChatGPT, o1, and Opus 4.5—to explain why the industry is compute-constrained and why hyperscaler capex is justified.
2. 1:51

   The first two LLM paradigms: ChatGPT's flaws and o1's reasoning breakthrough.

   • ChatGPT's November 2022 launch opened the world's eyes to LLM capabilities, but its two flaws—frequent errors and hallucinations, plus the need for users to proactively manage mistakes—made LLMs feel like a parlor trick.
   • OpenAI's o1 model in September 2024 introduced reasoning over answers before delivery, making LLMs internally proactive about managing mistakes and shifting them from merely useful to reliable and essential.
   • Traditional autoregressive LLMs are path dependent: once they commit to a guess they are locked in and doomed to failure, which reasoning models overcome by self-evaluating and considering alternatives.
3. 4:01

   The third paradigm: agentic workloads with harnesses like Claude Code and Codex.

   • Anthropic's Opus 4.5, released November 24, 2025, and OpenAI's GPT-5.2 Codex around December 18 suddenly enabled agents to accomplish tasks taking hours and do them correctly.
   • The critical component of agentic workloads is the harness—software that controls the model—which abstracts the user away from the model and lets agents use deterministic tools to verify results.
   • In this third paradigm, an agent directs a model to generate code, checks if it works, and retries without user involvement, substantially mitigating the original ChatGPT's flaws for verifiable use cases like coding.

关键概念

• agents— The central thesis of the episode: functional AI agents are the third paradigm shift that justifies massive compute investment.
• AI bubble debate— Thompson argues we are not in a bubble, reversing his earlier position that bubbles can be good.
• three LLM inflection points— ChatGPT, o1 reasoning models, and Opus 4.5 agentic workloads form the framework for the entire analysis.

精选金句

• I don't think we're in a bubble, which paradoxically maybe is the truest evidence we are.

  💡

  — A counterintuitive admission that the very act of declaring no bubble could be the strongest sign one exists.

  跳到 0:37

• The first flaw is that LM frequently got things wrong and worse would hallucinate when it didn't know the answer. This made LM feel like something of a parlor trick.

  🤯

  — Captures the early perception of LLMs as unreliable novelties, which many bubble proponents still cling to.

  跳到 1:30

可执行的洞察