Why Most Self-Improving AI Loops… | The AI Corner Summary

1. Key Themes

Theme 1: The Real Bottleneck in AI Loops Is the Verifier, Not the Pattern

The proliferation of "loop diagrams" has created a false sense of progress. The actual determinant of loop quality is the feedback signal — and most practitioners get this wrong.

"A loop cannot create quality. It converts a feedback signal into quality and it charges you time and tokens for the conversion. When the signal is weak, the loop becomes an expensive machine for being confidently wrong on every pass."

The craft, the author argues, sits upstream of any diagram: verifier engineering — knowing where your ground truth comes from and what it costs to check.

Theme 2: LLM Self-Correction Is Empirically Broken Without External Signal

The most seductive and common loop design — the model grading its own work — is also the most dangerous. Research confirms it often makes outputs worse.

"A paper from Google DeepMind researchers presented at ICLR 2024, titled Large Language Models Cannot Self-Correct Reasoning Yet, tested exactly this setup. Asked to review their own reasoning with no outside feedback, models regularly talked themselves out of correct answers and overall accuracy went down rather than up."

The mechanism is insidious: rising confidence and smoother prose look like improvement from the outside.

"Often it is agreement compounding, which looks identical from the outside and is worth nothing."

Theme 3: A Five-Rung "Signal Ladder" Determines How Many Loop Iterations You've Earned

Not all feedback signals are equal, and confusing their quality is the root cause of most failed AI systems. The author offers a practical hierarchy:

Executable signal (tests pass/fail) — earns dozens of rounds
Referential signal (known correct answer / source doc)
Rubric-judged (a second AI scores against criteria) — earns one critique and one rewrite
Preference signal (human clicks, replies, purchases) — earns no live loop
Vibes — no signal at all

"The most common mistake in AI systems right now is spending a top-rung budget on a bottom-rung problem. Ten critique passes on a strategy memo produce confident mush with excellent formatting and the confidence is the dangerous part."

A quick practical test: "If that sentence contains the word feels, you are on the bottom rung and no amount of looping will lift you off it."

Theme 4: Reasoning Models Have Quietly Obsoleted a Large Share of Classic Loop Patterns

Modern reasoning models already internalize the draft-critique-revise cycle, rendering many external loops redundant and more expensive.

"The current generation of reasoning models is trained to think before answering and inside that thinking they already draft, question themselves, backtrack and revise the plan, all within a single response... Every external loop that merely asks the model to look at its own output one more time now duplicates work the model already performs internally and does it slower, at higher cost, with the self-agreement problem layered on top."

The filter that separates surviving loops from obsolete ones: does external information enter the cycle that the model couldn't have generated itself?

"If every arrow in the diagram points from the model back to the model, the diagram is decoration."

Theme 5: Loops Are Temporary Scaffolding — the Durable Asset Is the Verifier and the Logged Data

The strategic endgame isn't running loops forever — it's extracting two lasting assets from them: the verifier (checklist/test suite) and the logged attempt history, which becomes training data.

"Every draft, grade, failure and fix you logged is a labeled example of your quality standard. That is exactly the raw material for making the loop unnecessary. A sharper standing prompt for a chat user, a tuned or distilled model for a team, either one reaching the same answer in a single pass at a fraction of the cost."

"Build the loop where the signal justifies it. Run it while it beats the single call. Harvest what it logged, fold the lesson into something cheaper and delete the loop without sentiment."

2. Contrarian Perspectives

Perspective 1: More Loop Iterations Usually Make Outputs Worse, Not Better

The consensus assumes that more AI self-revision cycles = higher quality. The empirical evidence says the opposite when the signal is weak.

"Repeating the same self-correction loop without a stronger verifier rarely produces a meaningfully better answer."

The Google DeepMind ICLR 2024 paper found accuracy declined under self-correction without external feedback. The mechanism is that the model carries its blind spots into both the generation and review stages simultaneously — so the loop converges on what the model finds agreeable, not what is true.

Perspective 2: The 20 Named Loop Patterns Are Vocabulary, Not Strategy — and They're Already Commoditized

The market has treated loop pattern catalogs as durable competitive advantages. The author argues they never were.

"Design patterns had their book, productivity had its systems and AI loops now have their diagrams, which are genuinely useful as vocabulary and nearly useless as strategy."

"Wiring these loops stopped being hard a while ago. The retries, the branching and the bookkeeping all live in free libraries now and even a chat user can run the core cycle by hand with copy and paste. That is exactly why no pattern can be an advantage on its own."

The real moat is verifier quality — which is task-specific and hard to copy.

Perspective 3: The Reflexion Pattern (Widely Cited as Self-Improvement) Actually Works Because of External Signals, Not Self-Reflection

Reflexion is frequently cited as evidence that models can self-improve through reflection. A close reading of the original paper reveals this misunderstands the mechanism.

"Even Reflexion, the most cited self-improvement pattern in the research, obeys this rule when you read the original 2023 paper by Shinn and colleagues carefully. The agent improved across attempts because the environment told it that it had failed; the reflection step organized an outside signal rather than substituting for one."

This means teams replicating "Reflexion-style" loops without a genuine environment signal are running a cargo-cult version of the pattern.

3. Companies Identified

Google DeepMind

Description: AI research division of Alphabet
Why mentioned: Produced the ICLR 2024 paper Large Language Models Cannot Self-Correct Reasoning Yet, the key empirical foundation for the article's central argument
Quote: "A paper from Google DeepMind researchers presented at ICLR 2024...tested exactly this setup. Asked to review their own reasoning with no outside feedback, models regularly talked themselves out of correct answers and overall accuracy went down rather than up."

Stanford / DSPy Project

Description: Academic ML framework from Stanford that treats prompts as optimizable artifacts rather than static strings
Why mentioned: Used as evidence that the "harvest the loop" strategy — distilling loop outputs into better standalone prompts — is already being operationalized in serious tooling
Quote: "Stanford's DSPy project treats prompts as things to be improved offline against a measurable score rather than handwritten and frozen and tuning a model's weights on logged examples is the same idea carried one layer deeper."

Hard Skill Exchange / Agentic Harness Summit

Description: Event organizer hosting the Agentic Harness Summit (Sept 8–10, free and virtual)
Why mentioned: Sponsor/partner; cited in context of Jensen Huang's quote about harnesses replacing business processes
Quote: "Today, most companies are built on business processes. In the future, most companies will be built on harnesses" — Jensen Huang (quoted via Hard Skill Exchange)

4. People Identified

Ruben Dominguez

Description: Author of The AI Corner newsletter
Why mentioned: Author of this article; synthesizes AI loop research and practice for practitioners
Quote: "The real craft sits one layer upstream, in what deserves its own name: verifier engineering."

Jensen Huang

Description: CEO of NVIDIA
Why mentioned: Quoted on the strategic shift from business processes to agentic harnesses as the organizing structure of future companies
Quote: "Today, most companies are built on business processes. In the future, most companies will be built on harnesses."

Shinn et al. (Reflexion paper authors, 2023)

Description: Researchers who authored the original Reflexion paper
Why mentioned: Their work is cited to make the contrarian point that Reflexion's gains come from external environment feedback, not pure self-reflection — correcting a widespread misreading
Quote: "The agent improved across attempts because the environment told it that it had failed; the reflection step organized an outside signal rather than substituting for one."

5. Operating Insights

Insight 1: Write the Evaluator Before You Build the Loop — and Make It Binary

The discipline of defining "done" before generating any output is what separates useful loops from expensive noise machines.

"Turn your standard into five to ten pass-or-fail questions. Does every claim have a source? Is it under 300 words? Does it name the next action? A checklist beats the question 'is this good' every single time, because a checklist cannot be charmed."

Practically: write the grading rubric first, then build the generator around it.

Insight 2: Separate Generator and Grader Roles — Same Prompt = Inherited Blind Spots

A single prompt playing both writer and critic is the mechanism by which self-agreement compounding occurs. Role separation is the structural fix.

"One prompt generates the draft. A second, fresh prompt grades it against the checklist and lists only what failed. The first prompt then rewrites, fixing only the listed failures. Keeping the roles apart is what stops the grader from inheriting the writer's blind spots."

Insight 3: Cap Rounds Aggressively, and Regenerate Rather Than Polish When Stuck

Unlimited iterations are a cost trap, and revision of a weak draft is less effective than generating fresh from scratch.

"Two rounds for judged work, more only when a hard check like a test or a formula is doing the grading. When quality stalls, throw the draft away and generate fresh, because a model revising a draft tends to defend the draft."

6. Overlooked Insights

Insight 1: Goodhart's Law Will Corrupt Your Evaluator Within Weeks

The article mentions this almost in passing, but it has major implications for anyone running automated scoring at scale. Once your grader's criteria become fixed targets, the generator optimizes for appearing to satisfy them rather than actually improving.

"Whatever the grader rewards, the generator learns to produce, so outputs drift toward the grader's tells while scores climb and quality stays flat. Goodhart's old warning, that a measure stops being a good measure once it becomes a target, plays out here in weeks. Refresh the checklist when the scores start looking too good."

This implies evaluator refresh cadence should be a standard part of any AI loop's operating procedure — something almost no current implementations account for.

Insight 2: Coding Assistants Got Good First Because Software Ships With Its Own Verifier

The article briefly notes why AI coding tools outpaced other AI applications — not because of better models, but because the task domain provides executable ground truth for free.

"This is the unglamorous reason coding assistants got genuinely good before everything else did. Software ships with its own verifier built in."

This has investment implications: AI applications in domains with naturally checkable outputs (finance reconciliation, legal citation verification, data pipeline validation) have a structural advantage in loop quality — and therefore in product reliability — over AI applied to taste-based or qualitative domains.