Where AI Value Lands

The Autonomy Slider

6 min read

The Autonomy Slider

One-line

Karpathy’s framing of human-AI interaction as a continuous spectrum — from tap-completion (human writes, AI suggests) to full agentic autonomy (AI rewrites entire repositories with no human approval) — with the design choice being where to fix the slider for a given product or task.

The spectrum (as Karpathy sketches it)

Anchored at the two ends:

  • Manual / tap-completion — human writes code, AI completes the next token or line. Human approves every keystroke.
  • Inline assistant — human in the IDE, AI proposes diffs, human accepts/rejects.
  • Goal-directed agent — human gives a goal, AI loops on prompt → tool → observe → repeat with intermittent checkpoints.
  • Fully autonomous — AI plans, executes, modifies its own code, and ships, with human reviewing only after the fact (or never).

The slider is not destiny: it’s a product-design knob set differently per use case.

The “Iron Man suit” — partial autonomy as default

Karpathy’s prescriptive claim: today’s LLMs are too fallible, gullible, and prone to hallucinations to be trusted across the full slider. The right default is partial autonomy — build products like an Iron Man suit, not an autonomous robot:

  • AI generates output fast.
  • Specialized GUIs designed to exploit the human visual cortex (e.g. side-by-side diffs in Cursor, source-cited answers in Perplexity) let humans verify rapidly.
  • The bottleneck is no longer AI generation; it’s human verification speed, so the UI is optimized for that.

Cursor, Perplexity, and similar tools succeed because they do not expose the raw LLM “OS” — they orchestrate multiple LLM calls and present a custom verification surface.

Where the claim comes from

karpathy-software-3. Corroborated by Peter Steinberger inside the same source: Steinberger frames the “agentic trap” as novice developers trying to skip the slider entirely and automate everything with monolithic prompts.

Vertical-industrial corroboration. deloitte-ai-dossier-eri restates the Iron Man Suit imperative in physical-infrastructure terms: “AI models simulating reservoir drilling or field safety must have escalation paths… for manual review.” In ER&I, the slider cannot reach full autonomy not because the AI cannot act, but because someone must be on the hook when the AI acts wrongly on a pipeline, a grid, or a drilling decision. The constraint is liability allocation, not just engineering reliability.

Relationship to H1_L0-L7-ladder

The autonomy slider and the L0–L7 ladder are distinct but correlated axes:

  • L0–L7 = who owns the substrate (chat → IDE → orchestrator → sandbox → fleet → model → datacenter).
  • Autonomy slider = how much human-in-the-loop oversight there is at each step.

A product can be high on the ladder, low on the slider (own your own GPUs but require human review on every action) or low on the ladder, high on the slider (run autonomous agents on someone else’s cloud). The two should likely be treated as orthogonal axes — cf. H3_orthogonal-axes-under-priced.

Tensions inside the framing

  • Karpathy vs. Steinberger inside the same source. Karpathy says the optimal product is the verification GUI (Iron Man suit). Steinberger predicts agents will eliminate 80% of apps by talking directly to endpoints, making human-facing GUIs redundant. These two are not reconciled.
  • The slider is asymmetric over time. Where the slider lands is an engineering choice today, but Huang’s agentic-scaling-law argues test-time compute will push the slider rightward automatically as reasoning improves. Karpathy’s “Iron Man” phase may be a transitional design pattern, not a permanent equilibrium.
  • The slider treats reliability as a smooth knob, but reliability is bimodal in safety-critical domains. A 99% autonomous agent is not 99% of a 100% autonomous one — it’s catastrophic. The slider framing under-weights threshold effects.
  • Karpathy vs. Sutskever — Iron Man vs straight-shot. sutskever-age-of-research articulates the opposite deployment posture: build superintelligence in a silo (the “straight-shot” method) and release only when ready. Karpathy’s slider treats partial autonomy as the durable design pattern; Sutskever treats partial autonomy as the slow, market-iterative path he is willing to abandon when the underlying capability is sufficient. The interviewer’s pushback — complex systems (aviation, Linux) became robust because they were deployed, failed, and were iteratively patched — is essentially the Iron Man argument applied to civilisational deployment. Sutskever appears to be updating; the tension is unresolved. The slider is also implicitly anchored by the eval-real-world-gap: as long as benchmarks and real performance diverge, the slider cannot be pushed far right safely.

Why this matters for Where Value Lands

  • Suggests that the defensibility of “top of the U” depends on whether verification GUIs remain valuable. If Steinberger is right, the application layer collapses; if Karpathy is right, the application layer is exactly where moats live.
  • Reframes “agents” not as a single category but as a slider position. A lot of confused capital flows (circular-ai-economy) come from treating “agents” as one thing instead of as a design choice.

Open questions

  • Empirically, which slider positions have shipped durably profitable products as of 2026? Mostly toward the manual end (Cursor, Copilot) — does this validate Karpathy’s claim or just say we’re early?
  • Are there domains where the slider naturally lives at full autonomy (well-bounded data pipelines, log triage, code formatting) vs. domains where it must stay low (legal, medical, finance)?

Graph View

Backlinks