TB: The Beauty of Slow Research Assistants (Fletcher)

Core Thesis

The slowness of research assistants was not purely a bug — it was a structural feature that enforced reflection, reprioritization, and deep thinking. AI removes this friction, inverting the bottleneck: execution is now near-instant and the scarce resource is the researcher's own attention and strategic judgment. Without deliberately reintroducing friction, researchers risk a firehose dynamic where ideas are consumed faster than they can be evaluated.

Key Takeaways

The RA Model as Pacing Infrastructure

  • Weekly lab meetings with 10+ students created a natural rhythm: progress reported, feedback given, then a week to act
  • The waiting period forced a shift in gears — you couldn't chase every tangent or act on every impulse
  • Over a year, the portfolio naturally resolved: some projects finishing, some dying, new ideas emerging from ashes
  • The system was inefficient in the human-systems sense but productive in a deeper sense: it created space for thinking

The Bottleneck Inversion

  • In the RA model: your time was rarely the binding constraint; the pipeline moved at the pace of your slowest collaborator
  • With AI: the constraint is unmistakably you — your attention, your ability to decide what matters, your capacity to say no
  • This shift doesn't scale easily; attention is not a parallelizable resource

The Firehose Problem

  • Before AI: ideas accumulated because execution was slow
  • With AI: execution is fast — ideas are now the thing being consumed (old dormant ideas, new ideas, speculative ideas now cheap to try)
  • The AI system is always ready and waiting; unlike a student who might be unavailable, the system continuously asks you for input

What Is Lost

  • Enforced pauses that created space for reflection
  • Natural pacing from human collaboration rhythms
  • The apprenticeship model: students learned by doing, slowly, imperfectly
  • The buffer that kept the researcher from becoming the bottleneck

Mental Models

  • Slow is Smooth and Smooth is Fast — Fletcher's argument inverts this: AI makes the execution smooth but may make the thinking fast and shallow; deliberately keeping some things slow is the meta-level application
  • Slack — the RA waiting period was institutional slack; AI collapses slack and concentrates all pressure on the researcher
  • Second-Order Thinking — the first-order effect of AI is faster execution; the second-order effect is that attention becomes the scarce resource and may degrade under firehose pressure
  • Incentives Matter — AI's incentive structure (always ready, always willing) creates implicit pressure to generate and execute rather than reflect and prune

See also

  • Tyler Cowen on AI and Academia — Cowen diagnosing the failure mode; Fletcher identifying a specific mechanism (pacing loss) most academics haven't noticed yet
  • Andy Hall: Five Pieces of Advice on AI for Polisci/PE Academics — Hall says raise your ambition and do more; Fletcher says be careful, the bottleneck shift is real; these are complementary, not contradictory
  • Attention and Deep Work — Newport on protecting deep work; Fletcher identifies AI as a new category of attention threat that also removes the structural slack of human collaboration
  • Lifelong Learning — the apprenticeship model is at risk; student learning by doing slowly may be short-circuited if AI does the work first
  • Agentic AI Systems — the multi-worker orchestration model is structurally the AI-accelerated lab meeting; Fletcher's "bottlenecked by yourself" problem is the master-session attention problem

Source

Jason Fletcher, Mentorless Apprentice (Substack), 2026-04-10 https://jasonmfletcher.substack.com/p/the-beauty-of-slow-research-assistants

Backlinks