The last six months in LLMs in five minutes

TEXT ANALYSIS: "The Last Six Months in LLMs in Five Minutes"

The Dissection

This is a tech insider's breezy retrospective on six months of LLM developments, presented as a PyCon lightning talk. The framing is deliberately casual—pelican drawings, holiday coding sprees, "LLM psychosis," Claws as "digital pets." The content covers: (1) the November 2025 "inflection point" when coding agents crossed a quality threshold, (2) the rapid cycling of "best model" between Anthropic, OpenAI, and Google, (3) the emergence of autonomous agent frameworks ("Claws"), and (4) the dramatic improvement of open-weight models running on commodity hardware.

The Core Fallacy

The entire piece is written as if the reader's primary relationship to these developments is delighted spectatorship. The framing—"look how far we've come!"—treats the automation of cognitive labor as a fun show rather than an extinction event for the economic order described in the DT framework.

The most critical failure: the section "the real news from November was that the coding agents got good" is presented without a single word about what "coding agents as daily drivers" means for software developers. It crossed a quality barrier. The barrier was human employment. This is not mentioned. It is celebrated.

Hidden Assumptions

• *"Good" is defined purely in terms of capability, never in terms of who loses
• Democratization of AI access is inherently positive—ignoring that open-weight models on laptops mean the autonomous agent wave hits everywhere simultaneously with no lag defense
• The pelican benchmark—presented as clever humor—is actually a perfect synecdoche for the industry's priorities: benchmarks that measure nothing economically relevant while the real benchmark (job displacement) goes unmeasured
• "LLM psychosis"—the author's throwaway joke about spinning up "wildly ambitious projects"—is actually a diagnostic signal. When even insiders are in a manic state chasing the technology, you have crossed into speculative frenzy. This is not treated as relevant data about the system's trajectory.

Social Function

This is transition management theater disguised as a lightning talk. Specifically:

• Prestige signaling: PyCon speaker, insider, someone who "was there" for the inflection
• Lullaby: The breezy tone and humor ("digital pets," "Alfred Molina's Doc Ock") lull the audience into viewing mass cognitive automation as entertainment
• Transition management: The implicit message is "adapt and learn these tools" without acknowledging that adaptation is irrelevant at the velocity being described
• Elite self-exoneration: The author acknowledges his own "LLM psychosis" and retired projects—a small act of self-critique that paradoxively demonstrates how thoroughly the insiders are themselves swept up in the acceleration without ethical grounding

The Verdict

This is a competent, engaging summary of the past six months that treats the destruction of cognitive labor markets as a series of fun facts. The piece is a near-perfect artifact of the DT's "transition management" phase: sophisticated people who understand the technology perfectly while remaining blind to its structural implications.

The November 2025 inflection was not "when the coding agents got good." It was when P1 (Cognitive Automation Dominance) crossed the threshold from aspirational to operational. Simon Willison is describing the execution of a capitalism-killing mechanism while drawing pelicans and making jokes about digital pets.

The piece ends on the note of open-weight models running on laptops outperforming expectations. That sentence, in DT terms, reads: "the diffusion of productive AI capital into decentralized ownership is accelerating, which accelerates the collapse of the mass employment-wages-consumption circuit."

The pelican test has not "exceeded its limits." The pelican test has always measured nothing that matters. The benchmark that matters is the employment statistics that won't come for another 18-24 months, when the lag defense of "we need human oversight" finally exhausts itself.

Autopsy, not debate. This post is a lovely retrospective of how the post-WWII order died, written by someone who enjoyed the ride.