Technology · Travel · Ideas

Andrew Koski

Writing about technology, travel, and the questions that connect them.

Recent writing

Evidence Before Action: When AI Systems Should Decide, Ask, or Abstain

November 6, 20267 min read

An AI system can produce a plausible answer before it has enough evidence to justify the action attached to that answer. I argue that acceptance should depend on the specific claim, the evidence supporting it, and the consequence of being wrong. This working paper proposes an evidence-to-action contract that separates candidate generation, claim validation, and action authorization. The framework applies to visual recognition, document processing, equipment identification, and information retrieval. It builds on selective prediction and confidence calibration without claiming to invent abstention. Its proposed contribution is an operational evaluation framework for claim-level evidence requirements, targeted requests for additional information, and proportionate escalation. A controlled evaluation is specified, including domain shift, incomplete inputs, and false acceptance costs. No new cross-domain experiment is reported.

Read paper

Working paperArtificial intelligence

When Corrected AI Answers Return: A Consistency Contract for Human-Reviewed Systems

November 6, 20267 min read

A human correction can be acknowledged while an older AI result remains capable of reappearing through a cache, delayed worker, or secondary read path. This is a consistency problem with direct implications for user trust. I propose a revision-based contract in which a result's eligibility depends on the current correction state of the subject to which it applies. The paper distinguishes cache expiry, invalidation, and atomic rejection of stale writes, then describes a bounded protocol and its assumptions. An illustrative execution shows why deletion alone is insufficient. A proposed evaluation measures stale-result exposure, correction visibility, availability, and cost under controlled interleavings and failures. The proposal applies to reviewed extraction, catalog identity, classification, and recommendation records. It does not claim that human correction retrains a model, removes all copies of data, or makes the corrected answer true.

Read paper

Working paperArtificial intelligence

The Quantum Necessity Conjecture for Artificial General Intelligence

November 5, 202616 min read

I conjecture that artificial systems achieving human-level general intelligence within realistic physical resource limits will require a functional quantum component integrated with classical computation. The motivation is that intelligence may depend on physical processes whose computational role is lost when the brain is treated only as a network of classical information-processing units. That possibility does not follow merely from the fact that brains are biological or that digital computers use binary representations. A credible necessity claim must connect a specific physical mechanism to an indispensable cognitive capability and show why classical alternatives cannot supply that capability within stated constraints. This paper separates quantum advantage, practical quantum necessity, and impossibility of classical AGI in principle. It examines relevant work on neural dynamics, decoherence, nuclear-spin proposals, and collective optical effects in biological structures, then develops a framework for testing the conjecture. The framework requires independent evidence of biological relevance, comparisons against strong classical alternatives, and explicit conditions for rejecting the claim. The result is a research position and a proposed standard of evidence, not an experimental demonstration or an impossibility proof.

Read paper

ArticleBlockchain & digital assets