<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Manuel Alejandro Rios on Medium]]></title>
        <description><![CDATA[Stories by Manuel Alejandro Rios on Medium]]></description>
        <link>https://medium.com/@manuel.rios_11557?source=rss-98ae34f4b6a3------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*yYW2CbEbTFJePPr7</url>
            <title>Stories by Manuel Alejandro Rios on Medium</title>
            <link>https://medium.com/@manuel.rios_11557?source=rss-98ae34f4b6a3------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 06:05:11 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@manuel.rios_11557/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[What AI Safety Can Learn from Civil/Agricultural Engineering Tolerances]]></title>
            <link>https://medium.com/@manuel.rios_11557/what-ai-safety-can-learn-from-civil-agricultural-engineering-tolerances-472533be81f9?source=rss-98ae34f4b6a3------2</link>
            <guid isPermaLink="false">https://medium.com/p/472533be81f9</guid>
            <category><![CDATA[engineering]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[hydraulics]]></category>
            <category><![CDATA[ai-safety]]></category>
            <dc:creator><![CDATA[Manuel Alejandro Rios]]></dc:creator>
            <pubDate>Fri, 16 Oct 2026 22:09:45 GMT</pubDate>
            <atom:updated>2026-10-16T22:09:45.919Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6Jd1XZUqtwfskkmqimLvTQ.jpeg" /><figcaption>Irrigation system in a Colombian flower farm</figcaption></figure><p>In the high-capacity irrigation and hydraulics projects I was involved in, a 1% failure rate was a disaster; since there were thousands of joints, valves and fittings, a 1% failure rate meant that something was failing somewhere every day and a single burst pipe could cause the pressure to drop throughout all the equipment downstream.</p><p>Water and civil engineers don’t design on the basis of the expected load; instead, we include factors of safety, build the components well above the level of working conditions, and check the limits before the system is put into use. Bridges and water networks are not safe simply because they are usually in good condition; rather, it is because we understand how they can fail and how far they are from failing.</p><p><strong>The problem: models without tolerances</strong></p><p>Most machine learning approaches work in the opposite manner. A model is released based on its average performance score, that is, the average load. Nothing short of a civil engineer would regard the average load as the basis for a design. Failure occurs in the tail: it is the peak demand, the unusual combination of conditions, and the input that lies just outside the training data distribution.</p><p>The first thing I notice when I go from designing networks to building models is that the tail is seldom examined. Since the benchmark data is generally taken from the same distribution as the training data, it cannot show you how performance deteriorates. Rare events are poorly represented, so the model can learn to ignore them, just as a hydraulic model calibrated on normal flow appears to work very well until the day when peak demand occurs. Moreover, a single average score masks which inputs cause failures and whether those failures are gradual or sudden.</p><p>In any pipe network, a failure which occurs at peak demand results in a loss of service. Frontier models are used by millions of people and have the same structure, being generally strong but unpredictable at the extremes, with no stated margin between normal operation and failure.</p><p><strong>The solution: stress-test to failure before release</strong></p><p>Two tests are used by engineers since they correspond well to this situation. In the lab, samples of components are tested until they are destroyed in order to determine the pressure level at which they actually fail. For installed networks, a proof test is carried out at a pressure higher than the working pressure in order to identify any weak joints before the system supplies water to anyone. Neither of these tests can guarantee that the pipe will never fail. Both of them do show how much margin is available.</p><p><strong>Red teaming should do the same three jobs for frontier models:</strong></p><p>Work out the breaking point by exploring the input space with people who are separate from the development team, and provide a failure envelope rather than a simple pass/fail test. Establish the margin; decide on the failure rate or the dangerous-capability threshold that you are unwilling to accept, and then insist on having demonstrated headroom below that level. Testers tend to underestimate the extent to which motivated users will find something, just as engineers assume that the loads and the strength of the materials are uncertain.<br>Look at the ways in which it can fail; list the different ways in which the system can fail and what effect each one has. After each change, re-test it just as you would when repairing a main.</p><p>Models are not acting as bridges since the input space is open-ended and the adversary adjusts, something that a pipe network never has to deal with; this means that wider margins are needed and that testing should continue after the release, not that the discipline should be abandoned.</p><p><strong>What an engineer brings</strong></p><p>For an engineer who is getting into the field of AI, the contribution does not consist in coming up with a new architecture but rather in getting the habit of asking what the design load is, where the system breaks, how much margin is still available, and what occurs downstream when it fails. My aim is to introduce rigorous, systems-level risk management into machine learning architecture, and frontier models should have to give answers to those questions before being released, just as every pressurised network has to.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=472533be81f9" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>