<?xml version="1.0" encoding="utf-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
    <channel>
        <title>MLST Archive</title>
        <link>https://archive.mlst.ai</link>
        <description>Technical papers from Machine Learning Street Talk</description>
        <lastBuildDate>Thu, 12 Feb 2026 03:15:24 GMT</lastBuildDate>
        <docs>https://validator.w3.org/feed/docs/rss2.html</docs>
        <generator>https://github.com/jpmonette/feed</generator>
        <language>en</language>
        <image>
            <title>MLST Archive</title>
            <url>https://archive.mlst.ai/images/og-default.png</url>
            <link>https://archive.mlst.ai</link>
        </image>
        <copyright>2026 Machine Learning Street Talk</copyright>
        <atom:link href="https://archive.mlst.ai/feed.xml" rel="self" type="application/rss+xml"/>
        <item>
            <title><![CDATA[Why Creativity Cannot Be Interpolated]]></title>
            <link>https://archive.mlst.ai/paper/why-creativity-cannot-be-interpolated/</link>
            <guid>https://archive.mlst.ai/paper/why-creativity-cannot-be-interpolated/</guid>
            <pubDate>Tue, 10 Feb 2026 00:00:00 GMT</pubDate>
            <description><![CDATA[AI systems can produce individually novel outputs, but novelty alone is not creativity. We argue that genuine creativity requires *respect for constraints* -- the accumulated structure of prior discoveries -- and that current AI systems lack this capacity because their training takes greedy paths that preclude the right kind of representations.

### Key claims

- **Creativity requires deep understanding and constraint-respecting extension.** "Slop" is the conjugate failure mode -- what you get when novelty is unconstrained by understanding.
- **More intelligence and more agency can *harm* transformational creativity.** As a corollary of Kenneth Stanley's *Why Greatness Cannot Be Planned*, objective-driven optimisation helps with exploratory creativity but inhibits -- perhaps completely -- the transformational kind. This matters for AI efforts trying to square knowledge synthesis with genuine "unknown unknown" discovery.
- **Constraints operate at three levels:** *physical* (baked into matter), *concrete* (instantiated in a fixed substrate), and *modelled* (represented so they can be manipulated, transferred, and counterfactually varied). Understanding is the cognitive form of this third level -- the capacity to navigate between constrained perspectives and integrate across them.
- **Current AI is only meaningfully creative under human supervision.** LLMs are coherent within any single frame, but they possess no trajectory of their own; their aggregated voice resolves into a coherent perspective only when a competent human supplies the grounding.
- **Human-AI co-creativity is the most promising path forward**, though the risk of underplaying human supervision is real -- as evidenced by wild extrapolations about near-term labour-market disruption.
- **We leave the door open for better architectures.** Any system -- biological or artificial -- whose learned, factored, path-dependent representations let it extend its own phylogeny could be genuinely creative, regardless of substrate.]]></description>
        </item>
    </channel>
</rss>