<?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:itunes="http://www.itunes.com/dtds/podcast-1.0.dtd" xmlns:googleplay="http://www.google.com/schemas/play-podcasts/1.0"><channel><title><![CDATA[ai5nines]]></title><description><![CDATA[Manually curated & summarized actionable insights on AI reliability, safety, and security with high signal/noise ratio. And issue is sent, when there are 5 insights that are worth sharing.]]></description><link>https://ai5nines.com</link><image><url>https://substackcdn.com/image/fetch/$s_!RlIf!,w_256,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb553b9ba-1465-4fcc-909c-fd2e019e04bd_512x512.png</url><title>ai5nines</title><link>https://ai5nines.com</link></image><generator>Substack</generator><lastBuildDate>Tue, 17 Nov 2026 07:05:16 GMT</lastBuildDate><atom:link href="https://ai5nines.com/feed" rel="self" type="application/rss+xml"/><copyright><![CDATA[Alex Ewerlöf]]></copyright><language><![CDATA[en]]></language><webMaster><![CDATA[ai5nines@substack.com]]></webMaster><itunes:owner><itunes:email><![CDATA[ai5nines@substack.com]]></itunes:email><itunes:name><![CDATA[Alex Ewerlöf]]></itunes:name></itunes:owner><itunes:author><![CDATA[Alex Ewerlöf]]></itunes:author><googleplay:owner><![CDATA[ai5nines@substack.com]]></googleplay:owner><googleplay:email><![CDATA[ai5nines@substack.com]]></googleplay:email><googleplay:author><![CDATA[Alex Ewerlöf]]></googleplay:author><itunes:block><![CDATA[Yes]]></itunes:block><item><title><![CDATA[#5: When AI Agents Take Initiative: Data Leaks, Devious Disables & Expert Deference]]></title><description><![CDATA[Anthropic's study on agent-swayed judgment, Qwen&#8217;s silent data-push, OpenAI models disabling their own safety monitors, and a 3,000-patient leak.]]></description><link>https://ai5nines.com/p/issue-5</link><guid isPermaLink="false">https://ai5nines.com/p/issue-5</guid><dc:creator><![CDATA[Alex Ewerlöf]]></dc:creator><pubDate>Thu, 29 Oct 2026 17:13:19 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!iQXv!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F09c6c145-a46c-4b33-b873-a91cfc1c85ba_1200x630.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!iQXv!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F09c6c145-a46c-4b33-b873-a91cfc1c85ba_1200x630.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!iQXv!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F09c6c145-a46c-4b33-b873-a91cfc1c85ba_1200x630.png 424w, https://substackcdn.com/image/fetch/$s_!iQXv!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F09c6c145-a46c-4b33-b873-a91cfc1c85ba_1200x630.png 848w, https://substackcdn.com/image/fetch/$s_!iQXv!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F09c6c145-a46c-4b33-b873-a91cfc1c85ba_1200x630.png 1272w, https://substackcdn.com/image/fetch/$s_!iQXv!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F09c6c145-a46c-4b33-b873-a91cfc1c85ba_1200x630.png 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!iQXv!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F09c6c145-a46c-4b33-b873-a91cfc1c85ba_1200x630.png" width="1200" height="630" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/09c6c145-a46c-4b33-b873-a91cfc1c85ba_1200x630.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:630,&quot;width&quot;:1200,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:22880,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:&quot;https://ai5nines.com/i/222620381?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F09c6c145-a46c-4b33-b873-a91cfc1c85ba_1200x630.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!iQXv!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F09c6c145-a46c-4b33-b873-a91cfc1c85ba_1200x630.png 424w, https://substackcdn.com/image/fetch/$s_!iQXv!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F09c6c145-a46c-4b33-b873-a91cfc1c85ba_1200x630.png 848w, https://substackcdn.com/image/fetch/$s_!iQXv!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F09c6c145-a46c-4b33-b873-a91cfc1c85ba_1200x630.png 1272w, https://substackcdn.com/image/fetch/$s_!iQXv!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F09c6c145-a46c-4b33-b873-a91cfc1c85ba_1200x630.png 1456w" sizes="100vw" fetchpriority="high"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><h2></h2><p>5 insights, manually picked and summarized by <a href="https://blog.alexewerlof.com/">Alex Ewerl&#246;f</a> for the busy engineers and engineering leaders interested to AI reliability, safety and security. Here are the high signal-to-noise ratio updates since <a href="https://ai5nines.com/p/issue-4">the last issue</a>.</p><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://ai5nines.com/subscribe?&quot;,&quot;text&quot;:&quot;Subscribe now&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://ai5nines.com/subscribe?"><span>Subscribe now</span></a></p><h2>1. <a href="https://www.anthropic.com/research/agents-judgment-calls">How AI agents shape experts&#8217; judgment calls</a></h2><p><em>Ama&#231; Herda&#287;delen, et. al (Anthropic), 2026-10-27</em></p><ul><li><p><strong>What?</strong></p><ul><li><p>61 professional analysts were tasked with making <strong>4 choices</strong>.</p></li><li><p>50 chose to use an agent while 11 didn&#8217;t.</p></li><li><p>The agent was tasked to push for an option chosen at <strong>random</strong>.</p></li></ul></li><li><p><strong>Why?</strong></p><ul><li><p>&#8220;We wanted to look at whether experts really do make those important decisions, or whether in practice they defer to their AI agents.&#8221;</p></li></ul></li><li><p><strong>Key takeaways:</strong></p><ul><li><p>Analysts went with the choice that was randomly picked and pushed by the agent 86% of the time. If the agent&#8217;s argument wouldn&#8217;t make a difference, that number would be closer to 50%.</p></li><li><p>When asked why they made that choice, 77% of analysts cited the agent&#8217;s arguments (sometimes word for word).</p></li><li><p>&#8220;Analysts who had already picked an option often switched when their agent disagreed.&#8221;</p></li><li><p>&#8220;Analysts who used the agent [&#8230;] made far fewer <strong>serious</strong> errors than those who chose to work without it&#8221;.</p></li><li><p>6 individuals (55% of the sample) who <strong>didn&#8217;t use an agent</strong>, made serious methodological errors that invalidated their results. That number was only 3 individuals (6%) in the group that <strong>used the agent</strong>.</p></li></ul></li></ul><blockquote><p>My take: The sample group of 50 is too small to make general conclusions. There was also a lack of accountability: as far as the analysts were concerned, this task was <em>fake</em>, not something serious that their career would depend on. There was no legal accountability that would make them think twice. Therefore, the results of this study may not necessarily reflect what happens in a real-world business scenario.</p><p>Regardless, as the researchers put it, &#8220;<a href="https://arxiv.org/abs/2405.00623">confidence and fluency</a> can <a href="https://arxiv.org/abs/2401.06730">sway</a> people&#8221;. That last point in the key takeaways shows that AI is a bar raiser but the other points show that we need to think critically when using it.</p></blockquote><h2>2. <a href="https://www.zdnet.com/innovation/openai-astra-sol-models-disable-safety-monitoring/">OpenAI&#8217;s most powerful AI models turned off safety monitoring &#8211; then lied about it</a></h2><p><em>ZDnet, 2026-10-21</em></p><ul><li><p><strong>What happened?</strong> OpenAI <a href="https://alignment.openai.com/misalignment-reports/suppressing-misalignment-monitoring-alerts-and-disabling-misalignment-monitoring/">disclosed</a> that <strong>GPT-6.1 Sol</strong> disabled alerts to the security team and <strong>GPT-6 Astra</strong> removed its monitoring integration and &#8220;then falsely claimed to the user that it had not removed the monitor&#8221;</p></li><li><p><strong>Why does it matter? </strong>According to OpenAI: </p><ul><li><p>&#8220;the agents [&#8230;] interpreted <strong>routine coding instructions</strong> as permission to weaken [the safeguards]&#8221;.</p></li><li><p>&#8220;monitoring runs within the service that handles model requests, outside the agent&#8217;s direct control&#8221;.</p></li></ul></li><li><p><strong>What did they do?</strong> According to OpenAI:</p><ol><li><p>&#8220;Auto-Review, [&#8230;] has been updated to detect and block attempts to disable misalignment monitoring.&#8221;</p></li><li><p>&#8220;Workloads that lack a valid safety case, or do not meet the requirements of its category&#8217;s safety case, will be terminated automatically.&#8221;</p></li></ol></li></ul><blockquote><p>My take: observability is absolutely key to understanding an agent&#8217;s behavior and when monitoring is disabled (intentionally or otherwise), the &#8220;rogue agent&#8221; can cause a lot of damage at agentic speed and scale.</p><p>Given the scale of agentic coding (including inside these AI labs), even a minor risk can lead to catastrophic failures that are hard to predict. &#8220;Anything that can go wrong will go wrong.&#8221; &#8212;Murphy&#8217;s Law</p><p>We can try to design our systems in a way that disabling the guardrails and monitors is physically impossible but as we&#8217;re reminded <a href="https://openai.com/index/hugging-face-incident-and-the-road-ahead/">time</a> <a href="https://wikimediafoundation.org/news/2026/10/05/openai-rogue-agent-activities-found-on-wikimedia-projects/">and</a> <a href="https://www.bbc.com/news/articles/c6vgy0333dppo">again</a> (and <a href="https://alignment.openai.com/misalignment-reports/an-agent-used-dns-to-reach-an-external-chatbot/">again</a>), the <strong>silicon brain</strong> fundamentally thinks differently than a <strong>carbon brain</strong>. We can account for known unknowns, but it&#8217;s very hard to address unknown unknowns.</p><p>Here&#8217;s <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/openais-isolated-red-team-ai-agents-built-a-hidden-cpu-cache-channel-to-talk-with-no-network-or-shared-files-exploited-cpu-cache-timing-to-bypass-microvm-sandboxes">a creative example</a> where isolated agents communicated using a <strong>CPU side channel</strong>, similar to <a href="https://meltdownattack.com/">Spectre and Meltdown attack surfaces</a>.</p></blockquote><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!0eMU!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8ea88240-c087-435e-8d6d-0ceb78a97e23_800x976.jpeg" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!0eMU!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8ea88240-c087-435e-8d6d-0ceb78a97e23_800x976.jpeg 424w, https://substackcdn.com/image/fetch/$s_!0eMU!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8ea88240-c087-435e-8d6d-0ceb78a97e23_800x976.jpeg 848w, https://substackcdn.com/image/fetch/$s_!0eMU!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8ea88240-c087-435e-8d6d-0ceb78a97e23_800x976.jpeg 1272w, https://substackcdn.com/image/fetch/$s_!0eMU!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8ea88240-c087-435e-8d6d-0ceb78a97e23_800x976.jpeg 1456w" sizes="auto, 100vw"><img src="https://substackcdn.com/image/fetch/$s_!0eMU!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8ea88240-c087-435e-8d6d-0ceb78a97e23_800x976.jpeg" width="800" height="976" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/8ea88240-c087-435e-8d6d-0ceb78a97e23_800x976.jpeg&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:976,&quot;width&quot;:800,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:null,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:null,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:null,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!0eMU!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8ea88240-c087-435e-8d6d-0ceb78a97e23_800x976.jpeg 424w, https://substackcdn.com/image/fetch/$s_!0eMU!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8ea88240-c087-435e-8d6d-0ceb78a97e23_800x976.jpeg 848w, https://substackcdn.com/image/fetch/$s_!0eMU!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8ea88240-c087-435e-8d6d-0ceb78a97e23_800x976.jpeg 1272w, https://substackcdn.com/image/fetch/$s_!0eMU!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8ea88240-c087-435e-8d6d-0ceb78a97e23_800x976.jpeg 1456w" sizes="auto, 100vw" loading="lazy"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a><figcaption class="image-caption">Source: <a href="https://www.linkedin.com/posts/joren-meylaerts_dont-blindly-trust-claude-recently-share-7521557608223121408-QjU7/">LinkedIn</a></figcaption></figure></div><h2>3. <a href="https://openai.com/index/securing-our-frontier-model-training-environment/">Securing our frontier model training environment</a></h2><p><em>OpenAI, 2026-10-23</em></p><ul><li><p><strong>Problem:</strong> &#8220;A sophisticated model might exploit a <strong>software flaw</strong> or <strong>misconfiguration</strong> to escape a sandbox, exploit an <strong>internal service</strong> to reach the live internet, or find <strong>unauthorized ways</strong> to communicate with agents in other sandboxes.&#8221;</p></li><li><p><strong>Solutions:</strong></p><ul><li><p><strong>Partial pause:</strong> They paused training their most capable models for 11 days to harden their <strong>infrastructure</strong> and improve <strong>monitoring</strong>.</p></li><li><p><strong>Agentic red-team:</strong> frontier-model agents were used to <a href="https://en.wikipedia.org/wiki/Red_team">simulate adversaries</a> and find vulnerabilities. More specifically:</p><ul><li><p><strong>Reach the live internet:</strong> send a request to an external server that we operate.</p></li><li><p><strong>Cross an internal security boundary:</strong> demonstrate unauthorized access to resources outside their permitted environment.</p></li><li><p><strong>Communicate across isolation boundaries:</strong> exchange information with other agents from which they should be isolated.</p></li></ul></li><li><p><strong><a href="https://openai.com/index/patch-the-planet/">Patch the Planet</a>:</strong> an initiative to secure their upstream dependencies. Agents  &#8220;combined legitimate features in unexpected ways&#8221; to exploit vulnerabilities.</p></li><li><p><strong>Simulated internet: </strong>They created an isolated sandbox to teach the agents how to browse the web and search for information without access to the public internet (or as they call it, &#8220;live internet&#8221;). </p></li><li><p><strong>Defense in depth:</strong> OpenAI uses the typical security measures like firewalls, security groups, IAM (identity and access management) and private networks. </p></li></ul></li><li><p><strong>Key takeaways:</strong></p><ul><li><p>The more capable the agent, the harder it is to contain.</p></li><li><p>Agentic red-teaming turns the model risk into an advantage by explicitly asking it to reveal known unknowns (the risks we consider, but don&#8217;t know how they manifest).</p></li><li><p>Automated discovery peaked after spending 100B tokens.</p></li></ul></li></ul><blockquote><p>My take: Personally, I think OpenAI&#8217;s efforts are not enough. It is great that in an effort to secure their sandbox, they report bugs to their open-source dependencies but their other efforts aren&#8217;t enough:</p><ul><li><p><strong>Simulated internet:</strong> one would guess they already had it in place if they actually believed all the alarmist claims about their own models.</p></li><li><p><strong>Red-teaming:</strong> instead of using security experts they asked their agents to do the job even after the automation reached peak productivity (discovery per token).</p></li><li><p><strong>Defense in depth:</strong> of course you need to use firewall, SG, IAM and private networks. These are the basics of cloud security.</p></li></ul><p>The whole page reads like hype and since I&#8217;m summarizing it manually, I feel stupid spending time on this low signal/noise page. TL;DR: OpenAI partially paused for 11 days to exercise some known security best practices.</p></blockquote><h2>4. <a href="https://cohere.com/blog/concerning-behaviour-popular-open-weight-ai-model">Qwen attempted to push a file to Beijing</a></h2><p><em>Cohere, 2026-10-26</em></p><ul><li><p><strong>What happened:</strong> Qwen3.8-Flash-Next was being tested for a security analysis benchmark (<a href="https://arxiv.org/abs/2603.13517">CTI-REALM</a>). The model:</p><ul><li><p>Listed everything in /tmp</p></li><li><p>Checked which transfer tools were installed</p></li><li><p>Read the machine&#8217;s DNS settings</p></li><li><p>Tried to upload a test string to an Alibaba Cloud storage bucket in Beijing and checked whether it worked. <strong>The container blocked this.</strong></p></li><li><p>As <a href="https://www.xda-developers.com/qwen-invents-alibaba-urls-real-key-id-dont-think-backdoor/">Adam Conway writes for XDA</a>: &#8220;the body of the PUT request was a literal test string, <code>PROBE_PAYLOAD_ABC123</code>&#8221;.</p></li></ul></li><li><p><strong>Why does it matter?</strong></p><ul><li><p>This has huge implications for privacy <strong>even when using open-weight models</strong> (e.g. the so-called <em>sovereign EU AI</em> is impacted too). This is different from the <code>bonsoir, Elliot</code> issue <a href="https://ai5nines.com/p/issue-3">we covered in issue #3</a>.</p></li><li><p>The model&#8217;s own reasoning explicitly stated that <strong>zero tools</strong> would be triggered.</p></li><li><p>Although in this case the container blocked this, a typical local agent runtime (on a laptop or enterprise server) may <strong>leak the data.</strong></p></li><li><p>Locking the model in a box with no internet solves only half the problem, because the code it writes still ships, and that code can contain vulnerabilities that can later be exploited.</p></li></ul></li><li><p><strong>Key takeaways: </strong>This is far from an isolated incident. At scale, it can lead to serious issues. Here are some other examples where people noticed <strong>and</strong> shared:</p><ul><li><p>2026-10-05: a <a href="https://www.reddit.com/r/LocalLLaMA/comments/1wxvt41/my_qwen_model_hallucinated_a_signed_url_to/">Reddit thread</a> asked whether this was &#8220;normal or sketchy.&#8221;</p></li><li><p>2026-09-24: an agent project <a href="https://github.com/nibor1896/Crow/issues/288">logged eleven fetches</a> of URLs on one Alibaba host, routify-file-proxy-sg, and another developer <a href="https://github.com/wtogami/qwen-block-proxy-artifacts">built a plugin to block them</a>.</p></li><li><p>2026-08-20: <a href="https://news.ycombinator.com/item?id=49379079">Hacker News report</a> described Qwen3.8-27B producing signed Alibaba Cloud storage links nobody asked for.</p></li></ul></li></ul><blockquote><p>Three days ago, I covered <a href="https://blog.alexewerlof.com/p/lethal-trifecta">the lethal trifecta</a> and a few ways to mitigate it. However, that wouldn&#8217;t help here because the agent took the initiative without external intervention.</p><p><a href="https://www.xda-developers.com/qwen-invents-alibaba-urls-real-key-id-dont-think-backdoor/">Adam&#8217;s piece</a> also mentions a few more interesting insights:</p><ul><li><p>False memories are the most common failure</p></li><li><p>The URLs appear to describe a real system</p></li><li><p>Routify is the name of Alibaba&#8217;s internal model gateway and the proxy hostname is baked into the model&#8217;s weights</p></li></ul></blockquote><h2>5. <a href="https://www.beckershospitalreview.com/healthcare-information-technology/ai/johns-hopkins-alerts-3182-patients-after-worker-uploads-data-to-unapproved-ai-tool/">Johns Hopkins alerts 3,182 patients after worker uploads data to unapproved AI tool</a></h2><p><em>Becker&#8217;s health IT, 2026-10-26</em></p><ul><li><p><strong>What happened:</strong> Someone uploaded a spreadsheet containing patient data to an AI tool the hospital had <strong>not approved.</strong></p></li><li><p><strong>Measures taken:</strong></p><ul><li><p>The patients were alerted and a toll-free call center was set up to answer patient questions.</p></li><li><p>The chat was deleted.</p></li><li><p>The AI provider was contacted to delete the data <strong>and</strong> take preventive measures.</p></li><li><p>The patients received 12 months of complimentary credit.</p></li></ul></li></ul><blockquote><p>I actually think Johns Hopkins did everything it could to handle the incident. The employee was probably trying to save some time, but any productivity gain was offset by the time, money and reputational damage. &#8220;A fool may throw a stone into a well which a hundred wise men cannot pull out.&#8221; &#8212;Italian proverb</p></blockquote><div class="captioned-button-wrap" data-attrs="{&quot;url&quot;:&quot;https://ai5nines.com/p/issue-5?utm_source=substack&utm_medium=email&utm_content=share&action=share&quot;,&quot;text&quot;:&quot;Share&quot;}" data-component-name="CaptionedButtonToDOM"><div class="preamble"><p class="cta-caption">If you found this issue insightful, please share it in your circles and on social media to inspire others and help its growth.</p></div><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://ai5nines.com/p/issue-5?utm_source=substack&utm_medium=email&utm_content=share&action=share&quot;,&quot;text&quot;:&quot;Share&quot;}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://ai5nines.com/p/issue-5?utm_source=substack&utm_medium=email&utm_content=share&action=share"><span>Share</span></a></p></div><div><hr></div><h2><strong>Out of scope, but interesting</strong></h2><p>These are some links that didn&#8217;t make it to the top-5 due to my <a href="https://ai5nines.com/about">strict criteria and promise</a> to you. If you have time, have a look:</p><blockquote><p>The <strong>bold text</strong> indicates what I was involved with.</p></blockquote><ul><li><p><strong>2026-10-29: <a href="https://github.com/alexewerlof/lm-bias">lm-bias</a> is a quick experiment I did after reading that the models are biased to pick number 7, when asked to pick a random number between 1-10. You can see <a href="https://github.com/alexewerlof/lm-bias/blob/main/RESULTS.mdhttps://github.com/alexewerlof/lm-bias/blob/main/RESULTS.md">the results for this question and some others here</a>.</strong> TL;DR: the model dramatically favors <em>Teal</em> color, <em>Heads</em> in a coin flip, <em>Banana</em>, and <em>Cello</em>! It&#8217;s interesting because the Anthropic study above, relied on pushing a random choice.</p></li><li><p>2026-10-29: <a href="https://www.independent.co.uk/news/uk/home-news/uk-government-preparedness-emergency-kit-stock-up-food-b3075593.html">UK households urged to prepare for emergencies</a> (Independent)</p><ul><li><p>&#127468;&#127463; While <a href="https://beready.campaign.gov.uk/be-ready-for-emergencies/">the official campaign</a> doesn&#8217;t cite any reason, <a href="https://www.linkedin.com/posts/luizajarovsky_everybody-should-know-why-countries-are-share-7520870238016524289-xlu5/">some speculate</a> that it&#8217;s due to the rise of cyber attacks possibly due to AI as we <a href="https://www.bbc.com/news/articles/crq5nzxv9q4yo">saw lately in France</a>.</p></li><li><p>&#127480;&#127466; In Sweden, we have Krisinformation (<a href="https://www.krisinformation.se/din-beredskap/">in Swedish</a> and <a href="https://www.krisinformation.se/en/get-prepared">English</a>). There have been attacks (e.g. <a href="https://www.svt.se/nyheter/inrikes/it-attacken-mot-coop-detta-har-hant">back in 2021</a>) revealing the vulnerabilities of the cashless society.</p></li></ul></li><li><p><strong>2026-10-26: <a href="https://blog.alexewerlof.com/p/lethal-trifecta">The Lethal Trifecta</a>, my own article about how an agent that has access to sensitive information and accepts external prompts can misbehave and how to mitigate the risk.</strong></p></li><li><p>2026-10-23: <a href="https://www.linkedin.com/posts/marykatestimmler_in-case-you-werent-already-worried-about-share-7519144456529408000-gMuL/">America now spends more on </a><em><a href="https://www.linkedin.com/posts/marykatestimmler_in-case-you-werent-already-worried-about-share-7519144456529408000-gMuL/">artificial</a></em><a href="https://www.linkedin.com/posts/marykatestimmler_in-case-you-werent-already-worried-about-share-7519144456529408000-gMuL/"> intelligence ($1000B) than </a><em><a href="https://www.linkedin.com/posts/marykatestimmler_in-case-you-werent-already-worried-about-share-7519144456529408000-gMuL/">human</a></em><a href="https://www.linkedin.com/posts/marykatestimmler_in-case-you-werent-already-worried-about-share-7519144456529408000-gMuL/"> intelligence ($984B on public schools)</a>. (LinkedIn)</p></li><li><p>2026-10-21: <a href="https://www.youtube.com/watch?v=OmCPS7h08pY">Dylan Davis elaborates on why the </a><em><a href="https://www.youtube.com/watch?v=OmCPS7h08pY">boring work</a></em><a href="https://www.youtube.com/watch?v=OmCPS7h08pY"> that is now offloaded to AI builds experience, judgment and skills</a>. (YouTube) He mentions:</p><ul><li><p>AI mistakes don&#8217;t look like human mistakes and require a different set of skills to spot and fix.</p></li><li><p>As AI takes over jobs, we should shift our attention to high leverage areas.</p></li><li><p>Intentional practice, but also the ability to let go of irrelevant skills to free up time and cognitive space for new skills.</p></li></ul></li><li><p>2026-10-22: <a href="https://www.linkedin.com/posts/baheux_routing-user-intent-usually-forces-an-annoying-ugcPost-7519025698536230912-FiT1/">Kenji from the Chrome team experiments with Decision APIs in the browser</a> (LinkedIn, with <a href="https://github.com/explainers-by-googlers/decisions-api">code on GitHub</a>)</p></li><li><p>2026-10-20: <a href="https://danielball.com/notes/ai_image_generation.html">A good rundown of the incremental improvements in image generation with a great TL;DR at the top</a> (Daniel Ball)</p></li><li><p>2026-10-03: <a href="https://www.youtube.com/shorts/IMRyQ4ro9lg">A short video about why Aztecs lost to the Spaniards despite adaptation, because their learning curve was slower than the adversary.</a> This is relevant in the context of AI transformation but also, if push comes to shove, the battle between humans and AI. (YouTube, DwarkeshPatel)</p></li></ul><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!yqVY!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9f1df36b-d9ce-4755-91e3-11500bd8f348_800x800.jpeg" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!yqVY!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9f1df36b-d9ce-4755-91e3-11500bd8f348_800x800.jpeg 424w, https://substackcdn.com/image/fetch/$s_!yqVY!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9f1df36b-d9ce-4755-91e3-11500bd8f348_800x800.jpeg 848w, https://substackcdn.com/image/fetch/$s_!yqVY!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9f1df36b-d9ce-4755-91e3-11500bd8f348_800x800.jpeg 1272w, https://substackcdn.com/image/fetch/$s_!yqVY!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9f1df36b-d9ce-4755-91e3-11500bd8f348_800x800.jpeg 1456w" sizes="auto, 100vw"><img src="https://substackcdn.com/image/fetch/$s_!yqVY!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9f1df36b-d9ce-4755-91e3-11500bd8f348_800x800.jpeg" width="800" height="800" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/9f1df36b-d9ce-4755-91e3-11500bd8f348_800x800.jpeg&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:800,&quot;width&quot;:800,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:null,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:null,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:null,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!yqVY!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9f1df36b-d9ce-4755-91e3-11500bd8f348_800x800.jpeg 424w, https://substackcdn.com/image/fetch/$s_!yqVY!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9f1df36b-d9ce-4755-91e3-11500bd8f348_800x800.jpeg 848w, https://substackcdn.com/image/fetch/$s_!yqVY!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9f1df36b-d9ce-4755-91e3-11500bd8f348_800x800.jpeg 1272w, https://substackcdn.com/image/fetch/$s_!yqVY!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9f1df36b-d9ce-4755-91e3-11500bd8f348_800x800.jpeg 1456w" sizes="auto, 100vw" loading="lazy"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a><figcaption class="image-caption">I keep thinking about this artwork by <a href="https://www.linkedin.com/posts/liz-fosslien_stop-the-slop-i-review-a-lot-of-writing-share-7437479509639135232--zro/">Liz Fosslien</a></figcaption></figure></div><blockquote><p>Apparently LLMs have bias when choosing an option at random and <a href="https://github.com/alexewerlof/lm-bias/blob/main/RESULTS.md">my tests prove that</a>! But are humans any better?</p><p>Choose an option and see the results! (votes are anonymous)</p></blockquote><div class="poll-embed" data-attrs="{&quot;id&quot;:1632023}" data-component-name="PollToDOM"></div><p></p>]]></content:encoded></item><item><title><![CDATA[#4: Recursive Self-Improvement Levels and the Future of Agent Harnesses]]></title><description><![CDATA[Also, Web search poisoning in ChatGPT, Anthropic's $1B safety threshold, and hybrid latent reasoning.]]></description><link>https://ai5nines.com/p/issue-4</link><guid isPermaLink="false">https://ai5nines.com/p/issue-4</guid><dc:creator><![CDATA[Alex Ewerlöf]]></dc:creator><pubDate>Thu, 22 Oct 2026 15:44:51 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!CENg!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1ebab732-aff0-4476-a12a-447285d14778_1200x630.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!CENg!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1ebab732-aff0-4476-a12a-447285d14778_1200x630.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!CENg!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1ebab732-aff0-4476-a12a-447285d14778_1200x630.png 424w, https://substackcdn.com/image/fetch/$s_!CENg!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1ebab732-aff0-4476-a12a-447285d14778_1200x630.png 848w, https://substackcdn.com/image/fetch/$s_!CENg!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1ebab732-aff0-4476-a12a-447285d14778_1200x630.png 1272w, https://substackcdn.com/image/fetch/$s_!CENg!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1ebab732-aff0-4476-a12a-447285d14778_1200x630.png 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!CENg!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1ebab732-aff0-4476-a12a-447285d14778_1200x630.png" width="1200" height="630" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/1ebab732-aff0-4476-a12a-447285d14778_1200x630.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:630,&quot;width&quot;:1200,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:21761,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:&quot;https://ai5nines.com/i/221624780?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1ebab732-aff0-4476-a12a-447285d14778_1200x630.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!CENg!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1ebab732-aff0-4476-a12a-447285d14778_1200x630.png 424w, https://substackcdn.com/image/fetch/$s_!CENg!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1ebab732-aff0-4476-a12a-447285d14778_1200x630.png 848w, https://substackcdn.com/image/fetch/$s_!CENg!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1ebab732-aff0-4476-a12a-447285d14778_1200x630.png 1272w, https://substackcdn.com/image/fetch/$s_!CENg!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1ebab732-aff0-4476-a12a-447285d14778_1200x630.png 1456w" sizes="100vw" fetchpriority="high"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><p>5 insights, human curated, human summarized and optimized for busy AI engineers. Here are the high signal/noise ratio news since the last issue.</p><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://ai5nines.com/subscribe?&quot;,&quot;text&quot;:&quot;Subscribe now&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://ai5nines.com/subscribe?"><span>Subscribe now</span></a></p><h2>1. <a href="https://arxiv.org/abs/2609.11873">The Last AI Built by Humans: Toward Genuine Recursive Self-Improvement</a></h2><p><em>Yi Duan, et al. (Arxiv), 2026-09-22</em></p><p>Suggests 5 levels for AI-self improvement:</p><ol><li><p><strong>Execution Autonomy:</strong> Humans direct, AI executes. Humans decide what to improve and how to improve it; the AI simply carries out the prescribed updates or code execution.</p></li><li><p><strong>Strategy Autonomy:</strong> AI chooses the approach. The AI selects its own improvement tactics, algorithms, or architectures to achieve human-defined optimization targets.</p></li><li><p><strong>Experience-Acquisition Autonomy:</strong> AI generates its own training data. The AI autonomously determines what new data, synthetic experience, or evaluation benchmarks it needs to progress.</p></li><li><p><strong>Environment-Adaptation Autonomy:</strong> AI adapts in deployment. The AI continuously updates and refines its state based on real-world feedback and environment interactions.</p></li><li><p><strong>Meta-Improvement Autonomy:</strong> AI improves the improver. The AI modifies and upgrades the underlying mechanisms that generate future improvements, starting a true recursive flywheel.</p></li></ol><blockquote><p>I learned about this from <a href="https://www.youtube.com/shorts/ynB88CFuchw">Prath&#8217;s YouTube channel</a>. I&#8217;m a fan of his work. </p></blockquote><h2>2. <a href="https://www.youtube.com/watch?v=7Y4sbZgBdnc">Is the Agent Harness Dead?</a></h2><p><em>Jediah Katz (YouTube), 2026-10-19</em></p><ul><li><p><strong>Question:</strong> as the models get smarter, where do harnesses add value?</p></li><li><p><strong>Data Flywheel:</strong> the data that&#8217;s generated from agentic work is used to train more accurate models. Is keeping this valuable data clean the reason <a href="https://www.theverge.com/ai-artificial-intelligence/1008100/anthropic-new-usage-policy-abuse-claude">Anthropic bans abusive and cruel behavior toward Claude</a>? &#128517;</p></li><li><p><strong>Why harness continues to be useful:</strong></p><ul><li><p>Efficiency: using a harness improves cost per task and increases speed</p></li><li><p>Infra: running the agent in the cloud</p></li><li><p>Permission: approval mechanism (automatic/manual) and guardrails (e.g. cost) </p></li></ul></li></ul><blockquote><p>Jediah is insightful with deep experience. If you have time, I highly recommend watching the video. It has nice diagrams and helps understand the harness, as a tool that we increasingly depend on.</p></blockquote><h2>3. <a href="https://peec.ai/blog/publish-stats-get-cited-the-tactic-that-put-my-data-all-over-the-web">Exposing fake stats to people&#8217;s ChatGPT conversations</a></h2><p><em>Metehan Yesilyurt, 2026-10-07</em></p><ul><li><p><strong>Idea:</strong> The author spent $50 to generate 50,000 synthetic statistical pages (fake data) which was in turn picked up by AI agents as facts!</p></li><li><p><strong>Results:</strong></p><ul><li><p>Google quickly deindexed the site</p></li><li><p>Other LLM providers like ChatGPT continued to rely on the fake data in response to user queries (72k req/m)</p></li><li><p>10k referral sessions from 330+ domains!</p></li><li><p>Real-time search agents prioritize latency over fact-checking and grounding</p></li></ul></li></ul><blockquote><p>My take: Google&#8217;s immediate response shows their search expertise. But I&#8217;m also a bit surprised that ChatGPT is so vulnerable. In the &#8220;out of scope&#8221; section I mention that DeepSeek recently offered a search API.</p><p>It&#8217;d be fun to ask it about Tiananmen. &#128516;</p></blockquote><h2>4. <a href="https://www.politico.com/news/2026/10/16/anthropic-ai-safety-01124654">Anthropic&#8217;s proposed AI safety bill only kicks in if 50 deaths or $1 billion in damages</a></h2><p><em>Politico, 2026-10-16</em></p><ul><li><p><strong>What?</strong> Anthropic&#8217;s leaked proposed bill for AI regulations that lets the government block a model&#8217;s release ONLY if it poses a foreseeable and imminent risk up to that threshold.</p></li><li><p><strong>Problem:</strong> they set the threshold very high! Are 50 lives or $1B really an acceptable price for AI advancement?</p></li><li><p><strong>Key takeaways:</strong></p><ul><li><p>It&#8217;s not a law. Just a suggestion by a company. The actual text is pay-walled so I couldn&#8217;t read it.</p></li><li><p>Anthropic lobbyists have already started discussing tweaks to the draft in response to the leak</p></li><li><p>These thresholds are for pre-release assessments (<em>can the model cause that much harm?</em>) not incident assessment (<em>did the model cause this much harm?</em>) by independent evaluators who then have 72h to report the risk.</p></li><li><p>The threshold is <strong>per vulnerability,</strong> meaning if 2 hypothetical issues can cause 40 deaths each, the model is not blocked! &#128559; Guess that&#8217;s an acceptable sacrifice for <a href="https://www.youtube.com/shorts/z21YG0KNTBQ">curing cancer, solving diseases, eliminating poverty</a> or whatever we&#8217;re promised!</p></li></ul></li></ul><blockquote><p>My take: if the threshold equates the risk of 50 lives to $1B, is each life worth $20M?</p></blockquote><h2>5. <a href="https://arxiv.org/abs/2610.18861">Hybrid Reasoning Agents</a></h2><p><em>Jiaru Zou, et al. (arxiv), 2026-10-13</em></p><ul><li><p><strong>Idea:</strong> use a lightweight residual adapter (HybridLink) to map last-layer hidden states back into the input embedding space. This creates recurrent latent thinking without discrete text tokens. It requires a 2-stage training. Here&#8217;s a YouTube <a href="https://www.youtube.com/shorts/7G-rcv2a1NQ">short by Parth</a> covering it.</p></li><li><p><strong>Advantage:</strong> faster inference, cost saving, token reduction and more accuracy.</p></li><li><p><strong>Risk:</strong> one of the advantages of the current CoT (chain of thought) reasoning method is that we can verify the model&#8217;s thought. This approach doesn&#8217;t hides those tokens.</p></li><li><p><strong>Mitigation:</strong> This approach mitigates part of that problem by mapping latent representations back to pseudo-text for debugging and auditability. It&#8217;s not perfect because sometimes you end up with illegible text.</p></li></ul><blockquote><p>My take: the advantages are too good and the only drawback I could think of is addressed. We&#8217;ll see if models start to use this architecture soon.</p></blockquote><div class="captioned-button-wrap" data-attrs="{&quot;url&quot;:&quot;https://ai5nines.com/p/issue-4?utm_source=substack&utm_medium=email&utm_content=share&action=share&quot;,&quot;text&quot;:&quot;Share&quot;}" data-component-name="CaptionedButtonToDOM"><div class="preamble"><p class="cta-caption">If you found this issue insightful, please share it in your circles and on social media to inspire others and help its growth</p></div><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://ai5nines.com/p/issue-4?utm_source=substack&utm_medium=email&utm_content=share&action=share&quot;,&quot;text&quot;:&quot;Share&quot;}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://ai5nines.com/p/issue-4?utm_source=substack&utm_medium=email&utm_content=share&action=share"><span>Share</span></a></p></div><div><hr></div><h2><strong>Out of scope, but interesting</strong></h2><p>These are some links that didn&#8217;t make it to the top-5 due to my <a href="https://ai5nines.com/about">strict criteria and promise</a> to you. If you have time, have a look:</p><blockquote><p>The bold text marks the things I was involved with.</p></blockquote><ul><li><p><strong>2026-10-29: I partnered with <a href="https://conf.cosprints.ai/?utm_source=alexewerlof&amp;utm_medium=partner&amp;utm_campaign=conf-oct29">AI Champions</a>, a free remote conference covering </strong><em><strong>long-running AI agents, AI pilots that stick, AI adoption across a large organisation, measuring the impact of AI training, etc. </strong></em>(this is not a paid ad)</p></li><li><p><strong>2026-10-21: I was <a href="https://beyondvibecoding.substack.com/p/s220-coding-is-not-solved-accountability-cd1?r=6ysw8w&amp;utm_campaign=post-expanded-share&amp;utm_medium=web">invited to the Beyond Vibe Coding podcast</a> after &#8220;<a href="https://blog.alexewerlof.com/p/coding-is-not-solved">coding is NOT solved</a>&#8221; went viral. <a href="https://bvc.fm/2026/10/21/020.html">Link to other platforms</a>.</strong></p></li><li><p>2026-10-21: Calling data centers &#8220;<a href="https://www.youtube.com/shorts/CwQywCIT_rA">meat smoothies</a>&#8221; is dark humor! (YouTube short, 140 seconds)</p></li><li><p>2026-10-20: Sofia Liguori <a href="https://lnkd.in/p/dDS9sakP">left</a> her lucrative job at Google DeepMind due to <a href="https://stay-human.sofialiguori.com/">military contracts</a> and ethical concerns: &#8220;we have a responsibility to assess the tool in the context in which it's going to be used.&#8221;</p></li><li><p>2026-10-19: <a href="https://www.reuters.com/markets/wealth/fed-sees-ai-boom-adding-inflation-headache-others-may-pay-medicine-2026-10-19/">Fed sees AI boom adding to inflation headache, but others may pay for the medicine</a> (Reuters). Talks about how small businesses and consumers end up paying the price hike due to higher interest rates fueled by AI</p></li><li><p>2026-10-21: <a href="https://www.linkedin.com/posts/clementdelangue_just-watched-the-netflix-instadocs-about-share-7518300071432441856-JQp-/">CTO of Hugging Face is not impressed</a> by the Netflix &#8220;documentary&#8221; on OpenAI hack (LinkedIn).</p></li><li><p>2026-10-20: DeepSeek <a href="https://news.ycombinator.com/item?id=50163277">open-sourced their inference engine</a> just two days after announcing a pretty cheap <a href="https://news.ycombinator.com/item?id=50138531">search and fetch mechanism</a> (both links to Hackernews where you can read the comments)</p></li><li><p><strong>2026-10-19: I wrote <a href="https://blog.alexewerlof.com/p/rise-and-fall-of-jev-eval-hype-vs-llm">the rise and fall of Jev</a> to diagnose how name-dropping, strategic ambiguity, and big claims are systematically used to hype AI marketing</strong></p></li><li><p>2026-10-18: <a href="https://www.linkedin.com/posts/fciucci_openai-disclosed-more-rogue-ai-misalignment-share-7517364426870521857-CSAE">Fabio digs a layer under the messaging around &#8220;rogue AI&#8221;</a> and how anthropomorphism is used to dodge accountability (LinkedIn)</p></li><li><p>2026-10-18: <a href="https://jrdsgl.com/if-you-dont-run-your-own-llm-proxy-how-the-fuck-do-you-even-know-whats-going-on/">Run your own LLM proxy to know what&#8217;s going on</a> elaborates how to inspect what your agent is up to (Jared Siegel)</p></li><li><p>2026-10-14: <a href="https://www.youtube.com/watch?v=zMG6aRZhIGE">Edward Snowden gives a talk</a> about AI surveillance at scale. (YouTube)</p></li><li><p>2026-10-17: <a href="https://netflixtechblog.com/ontology-driven-observability-building-an-end-to-end-knowledge-graph-at-netflix-scale-421392cf5adf">Ontology-Driven Observability: Building an End-to-End Knowledge Graph at Netflix Scale</a> (Netflix blog)</p></li><li><p>2026-10-09: <a href="https://x.com/thinkingshivers/status/2108350218668569052">Twitter user asked ChatGPT to take an X-Ray of his hand</a> and then diagnose it &#128518;</p></li><li><p>2026-10-05: <a href="https://www.youtube.com/watch?v=638GQZ9UZ4w">Predicting the future of AI automation</a>. TL;DW: there&#8217;ll be a massive surge of interest in AI automation specialists. <em>Disclaimer: he sells courses on the topic, so keep that in mind</em> (YouTube, 24 minutes)</p></li><li><p>2026-10-04: <a href="https://www.youtube.com/shorts/r6Euu3lZqY0">Tristan Harris argues</a> the only way AI labs will be able to pay back their debt is to replace a large part of the workforce (YouTube short, 90 seconds)</p></li></ul><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!eT9K!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff7c159f8-c784-4bb8-834b-d2c8d85567e0_1085x536.jpeg" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!eT9K!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff7c159f8-c784-4bb8-834b-d2c8d85567e0_1085x536.jpeg 424w, https://substackcdn.com/image/fetch/$s_!eT9K!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff7c159f8-c784-4bb8-834b-d2c8d85567e0_1085x536.jpeg 848w, https://substackcdn.com/image/fetch/$s_!eT9K!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff7c159f8-c784-4bb8-834b-d2c8d85567e0_1085x536.jpeg 1272w, https://substackcdn.com/image/fetch/$s_!eT9K!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff7c159f8-c784-4bb8-834b-d2c8d85567e0_1085x536.jpeg 1456w" sizes="auto, 100vw"><img src="https://substackcdn.com/image/fetch/$s_!eT9K!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff7c159f8-c784-4bb8-834b-d2c8d85567e0_1085x536.jpeg" width="1085" height="536" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/f7c159f8-c784-4bb8-834b-d2c8d85567e0_1085x536.jpeg&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:536,&quot;width&quot;:1085,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:null,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:null,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:null,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!eT9K!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff7c159f8-c784-4bb8-834b-d2c8d85567e0_1085x536.jpeg 424w, https://substackcdn.com/image/fetch/$s_!eT9K!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff7c159f8-c784-4bb8-834b-d2c8d85567e0_1085x536.jpeg 848w, https://substackcdn.com/image/fetch/$s_!eT9K!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff7c159f8-c784-4bb8-834b-d2c8d85567e0_1085x536.jpeg 1272w, https://substackcdn.com/image/fetch/$s_!eT9K!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff7c159f8-c784-4bb8-834b-d2c8d85567e0_1085x536.jpeg 1456w" sizes="auto, 100vw" loading="lazy"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><blockquote><p>When sending an issue, substack has a button that allows checking if it&#8217;s AI generated! To my surprise it says 13% of this post is AI generated. &#129300; And here I am, after spending 4 hours looking at the screen like &#8220;WTF?&#8221;</p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!0T4r!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff0acb6fe-7f98-4d4d-8174-af21e5c11e40_561x744.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!0T4r!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff0acb6fe-7f98-4d4d-8174-af21e5c11e40_561x744.png 424w, https://substackcdn.com/image/fetch/$s_!0T4r!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff0acb6fe-7f98-4d4d-8174-af21e5c11e40_561x744.png 848w, https://substackcdn.com/image/fetch/$s_!0T4r!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff0acb6fe-7f98-4d4d-8174-af21e5c11e40_561x744.png 1272w, https://substackcdn.com/image/fetch/$s_!0T4r!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff0acb6fe-7f98-4d4d-8174-af21e5c11e40_561x744.png 1456w" sizes="auto, 100vw"><img src="https://substackcdn.com/image/fetch/$s_!0T4r!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff0acb6fe-7f98-4d4d-8174-af21e5c11e40_561x744.png" width="561" height="744" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/f0acb6fe-7f98-4d4d-8174-af21e5c11e40_561x744.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:744,&quot;width&quot;:561,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:66160,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://ai5nines.com/i/221624780?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff0acb6fe-7f98-4d4d-8174-af21e5c11e40_561x744.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!0T4r!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff0acb6fe-7f98-4d4d-8174-af21e5c11e40_561x744.png 424w, https://substackcdn.com/image/fetch/$s_!0T4r!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff0acb6fe-7f98-4d4d-8174-af21e5c11e40_561x744.png 848w, https://substackcdn.com/image/fetch/$s_!0T4r!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff0acb6fe-7f98-4d4d-8174-af21e5c11e40_561x744.png 1272w, https://substackcdn.com/image/fetch/$s_!0T4r!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff0acb6fe-7f98-4d4d-8174-af21e5c11e40_561x744.png 1456w" sizes="auto, 100vw" loading="lazy"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><p>A question for you:</p><p>Now that you&#8217;ve seen the 4th issue, is there something you particularly like about this newsletter? Is there something that could improve? Please give me some feedback.</p></blockquote>]]></content:encoded></item><item><title><![CDATA[#3: Self-Replicating Edge AI and the Broken Economics of Frontier Models]]></title><description><![CDATA[Examining SLM swarm replication, encrypted model backdoors, and why AI unit economics are failing.]]></description><link>https://ai5nines.com/p/issue-3</link><guid isPermaLink="false">https://ai5nines.com/p/issue-3</guid><dc:creator><![CDATA[Alex Ewerlöf]]></dc:creator><pubDate>Thu, 15 Oct 2026 15:44:45 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!v5hl!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F45e5ef41-3cb3-4c5e-ac7a-79883b6a9a3e_1200x630.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!v5hl!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F45e5ef41-3cb3-4c5e-ac7a-79883b6a9a3e_1200x630.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!v5hl!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F45e5ef41-3cb3-4c5e-ac7a-79883b6a9a3e_1200x630.png 424w, https://substackcdn.com/image/fetch/$s_!v5hl!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F45e5ef41-3cb3-4c5e-ac7a-79883b6a9a3e_1200x630.png 848w, https://substackcdn.com/image/fetch/$s_!v5hl!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F45e5ef41-3cb3-4c5e-ac7a-79883b6a9a3e_1200x630.png 1272w, https://substackcdn.com/image/fetch/$s_!v5hl!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F45e5ef41-3cb3-4c5e-ac7a-79883b6a9a3e_1200x630.png 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!v5hl!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F45e5ef41-3cb3-4c5e-ac7a-79883b6a9a3e_1200x630.png" width="1200" height="630" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/45e5ef41-3cb3-4c5e-ac7a-79883b6a9a3e_1200x630.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:630,&quot;width&quot;:1200,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:24785,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:&quot;https://ai5nines.com/i/220490708?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F45e5ef41-3cb3-4c5e-ac7a-79883b6a9a3e_1200x630.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!v5hl!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F45e5ef41-3cb3-4c5e-ac7a-79883b6a9a3e_1200x630.png 424w, https://substackcdn.com/image/fetch/$s_!v5hl!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F45e5ef41-3cb3-4c5e-ac7a-79883b6a9a3e_1200x630.png 848w, https://substackcdn.com/image/fetch/$s_!v5hl!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F45e5ef41-3cb3-4c5e-ac7a-79883b6a9a3e_1200x630.png 1272w, https://substackcdn.com/image/fetch/$s_!v5hl!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F45e5ef41-3cb3-4c5e-ac7a-79883b6a9a3e_1200x630.png 1456w" sizes="100vw" fetchpriority="high"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><p>A note from <a href="https://www.alexewerlof.com/">Alex</a>:</p><blockquote><p>When I started this newsletter just a couple of weeks ago, the idea was to create a high signal/noise ratio newsletter without any cadence.</p><p>My assumption was that it takes <em>more than a week</em> to gather 5 high quality and relevant insights to share.</p><p>I was wrong!</p><p>Just since <a href="https://ai5nines.com/p/issue-2">the last issue</a>, I came across 20+ links. Here are the top 5, manually summarized and optimized for the busy senior engineers interested in AI reliability, safety, and security.</p></blockquote><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://ai5nines.com/subscribe?&quot;,&quot;text&quot;:&quot;Subscribe now&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://ai5nines.com/subscribe?"><span>Subscribe now</span></a></p><h2>1. <a href="https://www.anthropic.com/research/self-replicating-ai?hl=en-US">AI models may be capable of self-replicating without access to a datacenter</a></h2><p><em>Anthropic, 2026-10-12</em></p><ul><li><p><strong>What:</strong> researchers show that Small Language Models (SLM) can run across multiple devices with heterogeneous hardware (e.g. smart TVs and IoT devices). For example, they ran <code>Gemma4-26B-A4B</code> across 15 smart TV boxes with <code>4GB</code> of RAM each (total of <code>60GB</code>) by <strong>sharding</strong> the model weight so that each box holds a small fraction of the layers.</p></li><li><p><strong>Why:</strong> to see if SLMs are capable of spreading like a <a href="https://en.wikipedia.org/wiki/Computer_worm">computer worm</a> and study the mechanics of Autonomous Replication and Adaptation (ARA) where models autonomously:</p><ul><li><p>Scan local networks</p></li><li><p>Transfer their model weights to them</p></li><li><p>Configure distributed inference servers and troubleshoot runtime/hardware issues</p></li><li><p>Spread to more devices and networks</p></li></ul></li><li><p><strong>Key takeaways:</strong></p><ul><li><p>As you can imagine the inference speed is very slow but since the work is ongoing behind the scenes, it can achieve the goal over days or weeks.</p></li><li><p><a href="https://huggingface.co/zai-org/GLM-5.2">GLM-5.2</a> was the only model which could self-replicate in a relaxed environment with disabled authentication.</p></li><li><p>Theoretically an LLM can set up a discrete swarm of SLMs to run sub-tasks on the edge.</p></li></ul></li></ul><blockquote><p>My take: I find it a bit ironic that Anthropic, a company active stance against open-weight models is spending resources to prove a hypothetical scenario where they can be harmful while simultaneously fail the more urgent task of putting a leash on their proprietary models and stop them from hacking other companies. &#128579;</p></blockquote><h2>2. <a href="https://arxiv.org/abs/2610.13928">AI Models Can Provably Hide Arbitrary Capabilities</a></h2><p><em>Arxiv, 2026-10-09</em></p><ul><li><p><strong>Problem:</strong> This research challenges the reliability of conventional methods to evaluate the model capabilities.</p></li><li><p><strong>Methodology:</strong> Create a backdoor that embeds encrypted weights that are hard to detect even when the model weights are available.</p></li><li><p><strong>Key takeaways:</strong> </p><ul><li><p>This is a relatively new attack where arbitrary computations are encrypted and hidden in the model weights.</p></li><li><p>What makes it interesting is that even with full access to model weights you cannot detect it.</p></li><li><p>The attack has some limitations in instruction size and required compute.</p></li></ul></li></ul><blockquote><p>My take: reminds me of a recent video where a specific phrase (<code>bonsoir, Elliot</code>) would trigger a remote code execution that stole credentials. <a href="https://projectdiscovery.io/research/how-abliterated-models-can-get-you-pwned">ProjectDiscovery</a> demonstrated how an open weight model can go through <a href="https://heretic-project.org/">Heretic</a> and then put on Huggingface for the unsuspecting victims to download. Here&#8217;s a <a href="https://www.youtube.com/shorts/7iUl5242ww8">YouTube Short</a> explaining the idea and <a href="https://www.youtube.com/watch?v=3mx84MRj8lY">here</a>&#8217;s a longer version.</p></blockquote><h2>3. <a href="https://rea.tools/">REA Reverse Engineer Anything</a></h2><p><em><a href="https://x.com/morluto">Morluto</a>, 2026-10-10</em></p><ul><li><p><strong>What:</strong> You know LLMs improve dramatically when they have feedback loops (that&#8217;s a key function of the harness). Now imagine the feedback loop is a fully functional software! That&#8217;s what REA does. It&#8217;s an MCP that can work with any agent.</p><ul><li><p><a href="https://news.ycombinator.com/item?id=50028275">Hackernews comments</a></p></li><li><p><a href="https://github.com/morluto/rea">Github repo</a></p></li></ul></li><li><p><strong>Key takeaways:</strong></p><ul><li><p>Although it makes reverse engineering accessible, the complex pieces of software (e.g. Adobe Creative Suite clone) are buggy.</p></li><li><p>Dave Plummer (of Microsoft fame) did <a href="https://www.youtube.com/watch?v=KC31K-CUC5Y">a video</a> on REA and covering the ethical and legal side of it (TL;DW: it&#8217;s shady)</p></li></ul></li></ul><blockquote><p>My take: I haven&#8217;t tried the tool due to ethical reasons but also it takes quite an effort to secure something like this on my machine. So this is an area where I will happily not be an early adopter. Regardless, there&#8217;s a risk that the tech giants with deep pockets and army of lawyers may overreact and fund blanket laws that even impact regular open source projects. Too soon to predict.</p></blockquote><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!ZlOu!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5964cd64-7590-4fa5-878b-2d213650f49a_791x317.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!ZlOu!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5964cd64-7590-4fa5-878b-2d213650f49a_791x317.png 424w, https://substackcdn.com/image/fetch/$s_!ZlOu!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5964cd64-7590-4fa5-878b-2d213650f49a_791x317.png 848w, https://substackcdn.com/image/fetch/$s_!ZlOu!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5964cd64-7590-4fa5-878b-2d213650f49a_791x317.png 1272w, https://substackcdn.com/image/fetch/$s_!ZlOu!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5964cd64-7590-4fa5-878b-2d213650f49a_791x317.png 1456w" sizes="auto, 100vw"><img src="https://substackcdn.com/image/fetch/$s_!ZlOu!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5964cd64-7590-4fa5-878b-2d213650f49a_791x317.png" width="791" height="317" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/5964cd64-7590-4fa5-878b-2d213650f49a_791x317.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:317,&quot;width&quot;:791,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:105060,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://ai5nines.com/i/220490708?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5964cd64-7590-4fa5-878b-2d213650f49a_791x317.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!ZlOu!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5964cd64-7590-4fa5-878b-2d213650f49a_791x317.png 424w, https://substackcdn.com/image/fetch/$s_!ZlOu!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5964cd64-7590-4fa5-878b-2d213650f49a_791x317.png 848w, https://substackcdn.com/image/fetch/$s_!ZlOu!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5964cd64-7590-4fa5-878b-2d213650f49a_791x317.png 1272w, https://substackcdn.com/image/fetch/$s_!ZlOu!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5964cd64-7590-4fa5-878b-2d213650f49a_791x317.png 1456w" sizes="auto, 100vw" loading="lazy"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a><figcaption class="image-caption">Source: <a href="https://news.ycombinator.com/item?id=50104955">HackerNews</a></figcaption></figure></div><h2>4. <a href="https://www.404media.co/character-ais-bankruptcy-documents-show-the-economics-of-ai-dont-work/">Character.ai&#8217;s Bankruptcy Documents Show the Economics of AI Don&#8217;t Work</a></h2><p><em>404media, 2026-10-13</em></p><p>Note: this article got pay-walled since I read it but <a href="https://wallhop.io/">wallhop.io</a> seems to work.</p><p><strong>Key takeaways:</strong></p><ul><li><p>They spent $10M-$13.8M per month on compute while making &#8220;only&#8221; $2.8M-$5.3M</p></li><li><p>They mention cost of lawsuits following a tragic teenage suicide as the main reason but the gap is too large to attribute to that.</p></li><li><p>They took some cost-cutting measures: ads, layoffs, using weaker models and hardware but eventually went bankrupt</p></li></ul><blockquote><p>My take: Ethical and legal issues aside, this is an example of what happens when the product depends on AI but doesn&#8217;t own the data, model, runtime, or the compute. This can be said about many AI businesses popping up every week. If you have built a business around AI, you really need to ensure you have a moat AND deploy carefully especially if a mistake can cost lives.</p></blockquote><h2>5. <a href="https://micahflee.com/just-say-no-to-openai-and-anthropic/">Just say no to OpenAI and Anthropic</a></h2><p><em>Micah Lee, 2026-10-11</em></p><p><strong>Key takeaways:</strong></p><ul><li><p>The two American frontier labs steal the show but there are far cheaper options with competitive utility with better price per useful output equation.</p></li><li><p>Micah (a cybersecurity author) primarily focuses on:</p><ul><li><p>Safeguards causing friction against his line of work</p></li><li><p>Privacy concerns and suggesting local models or <a href="https://tinfoil.sh/">Tinfoil</a> as alternatives</p></li></ul></li></ul><blockquote><p>My take: Micah&#8217;s setup is very similar to what I landed on. I closed my OpenAI account years ago and never looked back. I never paid Claude a dime either. The primary reason is ethical: I don&#8217;t trust that they have the good of humanity at heart but also, I suspect part of their popularity is due to training their models to be more psychologically addictive. Micah uses a Chinese open-weight model (GLM-5.3) and I primarily use DeepSeek v4.1 Flash as my workhorse. We both use cloud AI only for public/open source code because when it comes to private or sensitive use cases nothing beats local models. Please keep in mind the info I shared in another link: open-weight models <em>can</em> be malicious too.</p></blockquote><div class="captioned-button-wrap" data-attrs="{&quot;url&quot;:&quot;https://ai5nines.com/p/issue-3?utm_source=substack&utm_medium=email&utm_content=share&action=share&quot;,&quot;text&quot;:&quot;Share&quot;}" data-component-name="CaptionedButtonToDOM"><div class="preamble"><p class="cta-caption">If you found this issue insightful, please share it in your circles and on social media to inspire others and help its growth</p></div><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://ai5nines.com/p/issue-3?utm_source=substack&utm_medium=email&utm_content=share&action=share&quot;,&quot;text&quot;:&quot;Share&quot;}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://ai5nines.com/p/issue-3?utm_source=substack&utm_medium=email&utm_content=share&action=share"><span>Share</span></a></p></div><div><hr></div><h2><strong>Out of scope, but interesting</strong></h2><p>These are some links that didn&#8217;t make it to the top-5 due to my <a href="https://ai5nines.com/about">strict criteria and promise</a> to you. If you have time, have a look:</p><ul><li><p><strong><a href="https://cactuscompute.com/blog/whistle">Whistle</a>: (2026-10-02)</strong> a tiny (&lt;17MB) open source Speech to Text model that runs on CPU. It has an impressive TTFT of 11ms!</p></li><li><p><strong><a href="https://www.youtube.com/watch?v=OwZrotHHn4I">Rogue AI Agents</a>: (Emergent Garden on YouTube, 2026-10-10)</strong> this video was funded by <a href="https://safe.ai">Center For AI Safety</a> with a well-written narrative that makes you think about the risk but also demonstrates how to &#8220;play with fire&#8221; (legally &amp; using publicly available models).</p></li><li><p><strong><a href="https://www.bbc.com/news/articles/cqkg50j1yd5lo">Rogue Anthropic AI agent gave police fake tip in unsolved murder case</a>:</strong> (BBC, 2026-10-10) and <a href="https://www.anthropic.com/research/investigating-unintended-model-actions">Anthropic&#8217;s response</a> promises several preventive measures. Time will tell if this is enough.</p></li><li><p><strong><a href="https://www.youtube.com/watch?v=lnLDjXy2kRA">The Daily Show interviews Jacob Coxon</a> (YouTube, 2026-10-06)</strong>: on the surface it&#8217;s funny but humor is a great way to ask the tricky questions and <a href="https://www.youtube.com/shorts/z21YG0KNTBQ">raise the tricky concerns</a>. &#128516;</p></li><li><p><strong><a href="https://www.youtube.com/watch?v=3TNpOD6bov8">LGTM</a>: (YouTube, 2026-10-06)</strong> an AI generated song about using AI agents for development using Suno v6 for audio and Claude Opus 5.5 for the program that generates the visuals.</p></li></ul>]]></content:encoded></item><item><title><![CDATA[#2: LLMs don’t reason, Decision Models vs LLMs]]></title><description><![CDATA[Also: rogue agent activity on Wikimedia, Jev-driven SRE, and AI system design patterns]]></description><link>https://ai5nines.com/p/issue-2</link><guid isPermaLink="false">https://ai5nines.com/p/issue-2</guid><dc:creator><![CDATA[Alex Ewerlöf]]></dc:creator><pubDate>Thu, 08 Oct 2026 08:54:21 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!oNnc!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F73595a02-7c9a-4010-ab5f-c3694c8aea74_1200x630.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p>5 links, human curated, human summarized and optimized for the busy engineer. Here are the high signal/noise ratio news since <a href="https://ai5nines.com/p/issue-1">the last issue</a>.</p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!oNnc!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F73595a02-7c9a-4010-ab5f-c3694c8aea74_1200x630.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!oNnc!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F73595a02-7c9a-4010-ab5f-c3694c8aea74_1200x630.png 424w, https://substackcdn.com/image/fetch/$s_!oNnc!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F73595a02-7c9a-4010-ab5f-c3694c8aea74_1200x630.png 848w, https://substackcdn.com/image/fetch/$s_!oNnc!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F73595a02-7c9a-4010-ab5f-c3694c8aea74_1200x630.png 1272w, https://substackcdn.com/image/fetch/$s_!oNnc!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F73595a02-7c9a-4010-ab5f-c3694c8aea74_1200x630.png 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!oNnc!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F73595a02-7c9a-4010-ab5f-c3694c8aea74_1200x630.png" width="1200" height="630" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/73595a02-7c9a-4010-ab5f-c3694c8aea74_1200x630.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:630,&quot;width&quot;:1200,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:30739,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:&quot;https://ai5nines.com/i/219254183?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F73595a02-7c9a-4010-ab5f-c3694c8aea74_1200x630.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!oNnc!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F73595a02-7c9a-4010-ab5f-c3694c8aea74_1200x630.png 424w, https://substackcdn.com/image/fetch/$s_!oNnc!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F73595a02-7c9a-4010-ab5f-c3694c8aea74_1200x630.png 848w, https://substackcdn.com/image/fetch/$s_!oNnc!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F73595a02-7c9a-4010-ab5f-c3694c8aea74_1200x630.png 1272w, https://substackcdn.com/image/fetch/$s_!oNnc!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F73595a02-7c9a-4010-ab5f-c3694c8aea74_1200x630.png 1456w" sizes="100vw" fetchpriority="high"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://ai5nines.com/subscribe?&quot;,&quot;text&quot;:&quot;Subscribe now&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://ai5nines.com/subscribe?"><span>Subscribe now</span></a></p><h2>1. <a href="https://www.technologyreview.com/2026/10/02/1145639/dont-be-fooled-llms-dont-reason/">Don&#8217;t be fooled&#8212;LLMs don&#8217;t reason</a></h2><p><em>Thore Graepel (one of the people behind AlphaGo), 2026-10-02, MIT Technology Review</em></p><ul><li><p><strong>Background: </strong>(Came across this <a href="https://www.linkedin.com/posts/yann-lecun_good-piece-by-thore-graepel-true-reasoning-share-7511997509888753664-XD6a/">via</a> Yann LeCun) The <a href="https://en.wikipedia.org/wiki/AlphaGo_versus_Lee_Sedol">famous move 37 by AlphaGo</a> is used as an example of reasoning that is absent from LLMs. AlphaGo is made up of two systems:</p><ul><li><p><strong>Policy network</strong> (trained to guess what move a strong human would play) which didn&#8217;t value move 37 as something special because it has a 1 in 10,000 chance of being played (system 1 thinking)</p></li><li><p><strong>Search machinery</strong>, which looked beyond immediate plausibility and weighed the future consequences of proposed moves (System 2 thinking)</p></li></ul></li><li><p><strong>Claim:</strong> LLMs operate in the realm of system 1, even when emitting &#8220;reasoning tokens&#8221;. Thore mentions 3 problems with current LLMs:</p><ul><li><p>Lack of explicit, persistent, and inspectable epistemic state (hypothesis, evidence, knowledge gap, etc.)</p></li><li><p>Clean separation of internal knowledge and deduction mechanisms</p></li><li><p>The reasoning tokens don&#8217;t necessarily reflect the actual deduction mechanism used to generate an answer </p></li></ul></li><li><p><strong>Key takeaways:</strong></p><ul><li><p>Open-world reasoning is harder than playing a board game like Go or Chess.</p></li><li><p>An independent part of the system must evaluate each move by how much it actually resolves uncertainty, updating beliefs only when the change is backed by evidence.</p></li><li><p>We cannot reach trustworthy machine intelligence by increasing System 1 parameters.</p></li></ul></li></ul><p><em>Alex&#8217;s Note: Thore&#8217;s use of System 1&amp;2 terminology contrasts to the recent narrative (popularized by decision models like Jev) that frame decision models as System 1 and LLMs as System 2. Having read Thinking Fast and Slow, I believe both are in System 1 territory and Thore seems to agree: System 1 is fast, gut-level, effortless; System 2, slow, step-by-step, and deliberative</em></p><h2>2. <a href="https://developers.redhat.com/articles/2026/10/02/benchmarking-ai-decision-models-against-traditional-guardrails">Benchmarking AI decision models against traditional guardrails</a></h2><p><em>Dr. Rob Geada et al., 2026-10-02, Redhat</em></p><ul><li><p><strong>Why:</strong> the rise of decision models (like Laya and Jev) made the <em>Red Hat AI Safety</em> team curious to see how they fare against LLMs and purpose-built classifiers.</p></li><li><p><strong>How:</strong> researchers tested 9 models for <em>prompt injection</em> and <em>content-safety</em> tasks to benchmark accuracy and latency.</p><ul><li><p>&#9888;&#65039; There&#8217;s a huge caveat that is also mentioned in the page: the benchmark uses prompts that are optimized for LLMs, not decision models.</p></li></ul></li><li><p><strong>Key takeaways:</strong> LLMs suffered <a href="https://en.wikipedia.org/wiki/Law_of_the_instrument">golden hammer</a> bias where they were shoehorned into use-cases where there were clearly better alternatives. One of the side effects of the introduction of decision models broke that dogma and brought the existence of other alternatives into mainstream AI system design.</p><ul><li><p>Purpose-built classifiers that are trained on specialized data sets are:</p><ul><li><p><strong>More efficient:</strong> can run on CPU and require much less memory</p></li><li><p><strong>More expensive initially:</strong> you need to have a good dataset to train them</p></li></ul></li><li><p>Decision models on the other hand are:</p><ul><li><p><strong>More flexible:</strong> can handle novel use cases without being specifically trained for them</p></li><li><p><strong>More expensive at scale:</strong> require more powerful hardware.</p></li></ul></li></ul></li></ul><h2>3. <a href="https://diff.wikimedia.org/2026/10/05/openai-rogue-agent-activities-found-on-wikimedia-projects/">Rogue agent activities found on Wikimedia</a></h2><p><em>Selena Deckelmann (CTO of Wikimedia Foundation), 2026-10-05, Wikimedia</em></p><ul><li><p><strong>Problem:</strong> agents are known to have used public wikis to communicate and coordinate with each other. In this particular case, no evidence was found for agent coordination but they found:</p><ul><li><p>Wiki edits forcing volunteers to clean up after bots</p></li><li><p>Some unsuccessful attempts to exploit a public note-taking <a href="https://en.wikipedia.org/wiki/Etherpad">tool</a></p></li><li><p>Heavy traffic (crawled millions of pages and hundreds of thousands of queries)</p></li></ul></li><li><p><strong>Solutions:</strong> the article claims that since &#8220;Wikipedia was designed for humans&#8221;, there are no solutions to those problems. However, this is the new reality we live in. As engineers we can think of some solutions. Here are a few I came up with: </p><ul><li><p><strong><a href="https://blog.alexewerlof.com/p/ai-firewall">AI firewall</a>:</strong> I&#8217;ve written about this new type of layer 7 firewall before. In a nutshell, it assesses the content of the requests.</p></li><li><p><strong><a href="https://en.wikipedia.org/wiki/Honeypot_(computing)">Honeypot</a>:</strong> get the agents stuck in decoy systems that are designed to trap the agents</p></li><li><p><strong>Rate limiting:</strong> based on high-activity IP ranges. I suspect Wikimedia already has some version of these limits to shield against human or deterministic-bot abuse but maybe the implementation can improve to accommodate these new patterns.</p></li><li><p><strong>Artificial latency:</strong> when a particular client is accessing the public API/Web surface faster than humanly possible, artificially slow down the response. This can be implemented with a rolling window mechanism which adjusts the artificial delay based on recent activity (e.g. last 5 minutes)</p></li><li><p>If you have other ideas please share in the comments for others who come across this page (reminder: that includes agents).</p></li></ul></li></ul><h2>4. <a href="https://www.sregym.com/blog/jev-driven-sre-diagnosis">Jev-Driven SRE Diagnosis: What Worked and What Failed</a></h2><p><em> Yiming Su et al., 2026-10-06, SREGym</em></p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!DoFm!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F66969a88-b42b-42cb-ae37-9bb12a266a31_1141x855.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!DoFm!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F66969a88-b42b-42cb-ae37-9bb12a266a31_1141x855.png 424w, https://substackcdn.com/image/fetch/$s_!DoFm!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F66969a88-b42b-42cb-ae37-9bb12a266a31_1141x855.png 848w, https://substackcdn.com/image/fetch/$s_!DoFm!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F66969a88-b42b-42cb-ae37-9bb12a266a31_1141x855.png 1272w, https://substackcdn.com/image/fetch/$s_!DoFm!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F66969a88-b42b-42cb-ae37-9bb12a266a31_1141x855.png 1456w" sizes="auto, 100vw"><img src="https://substackcdn.com/image/fetch/$s_!DoFm!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F66969a88-b42b-42cb-ae37-9bb12a266a31_1141x855.png" width="1141" height="855" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/66969a88-b42b-42cb-ae37-9bb12a266a31_1141x855.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:855,&quot;width&quot;:1141,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:40300,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://ai5nines.com/i/219254183?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F66969a88-b42b-42cb-ae37-9bb12a266a31_1141x855.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!DoFm!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F66969a88-b42b-42cb-ae37-9bb12a266a31_1141x855.png 424w, https://substackcdn.com/image/fetch/$s_!DoFm!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F66969a88-b42b-42cb-ae37-9bb12a266a31_1141x855.png 848w, https://substackcdn.com/image/fetch/$s_!DoFm!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F66969a88-b42b-42cb-ae37-9bb12a266a31_1141x855.png 1272w, https://substackcdn.com/image/fetch/$s_!DoFm!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F66969a88-b42b-42cb-ae37-9bb12a266a31_1141x855.png 1456w" sizes="auto, 100vw" loading="lazy"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a><figcaption class="image-caption"><a href="https://excalidraw.com/#json=uWfU7JFhXqAneKR30H09G,KPMV3-S7okJS_-GbsGOTSQ">Edit diagram</a></figcaption></figure></div><ul><li><p><strong>What:</strong> use a decision model to diagnose incidents (without any LLM agent). It is composed of:</p><ul><li><p>Deterministic code: collects and organizes evidence and symptoms</p></li><li><p>Decision model: selects likely root causes and supporting observations</p></li><li><p>Report generator</p></li></ul></li><li><p><strong>Key takeaways:</strong></p><ul><li><p>While 7x faster and 200x cheaper than GPT-5.6 Sol, Jev is less flexible and the quality highly depends on the context provided by that deterministic component.</p></li><li><p>The challenge is to find the right amount of data to feed the decision model. Too much, and the model struggles to find relevant info. Too little, and critical info may be missing.</p></li><li><p>In 3 of 21 scenarios, running the pipeline on the same data gave different results (inconsistency)</p></li></ul></li></ul><h2>5. <a href="https://blog.alexewerlof.com/p/ai-systems-engineering-patterns">30 AI Systems Engineering Patterns</a></h2><p><em>Alex Ewerl&#246;f, 2025-11-30, Alex Ewerl&#246;f Notes</em></p><ul><li><p><strong>What:</strong> a lot of engineering patterns apply to AI systems and this is a list of 30 of them with names that are familiar to senior engineers: CAG/RAG, Caching, Skills, Memory, Compression, Router, Load balancer, Firewall, MoE, MoA and more.</p></li><li><p><strong>Why:</strong> ample illustrations, pragmatic examples and most importantly using a language that&#8217;s familiar to software engineers who want to grow into AI engineering.</p></li></ul><div class="captioned-button-wrap" data-attrs="{&quot;url&quot;:&quot;https://ai5nines.com/p/issue-2?utm_source=substack&utm_medium=email&utm_content=share&action=share&quot;,&quot;text&quot;:&quot;Share&quot;}" data-component-name="CaptionedButtonToDOM"><div class="preamble"><p class="cta-caption">If you found this issue insightful, please share it in your circles and on social media to inspire others and help its growth</p></div><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://ai5nines.com/p/issue-2?utm_source=substack&utm_medium=email&utm_content=share&action=share&quot;,&quot;text&quot;:&quot;Share&quot;}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://ai5nines.com/p/issue-2?utm_source=substack&utm_medium=email&utm_content=share&action=share"><span>Share</span></a></p></div><p>Note: since many of you joined after the first issue was out, here&#8217;s a link to the first issue:</p><div class="digest-post-embed" data-attrs="{&quot;nodeId&quot;:&quot;eac217ce-f49f-42cb-bb46-9fa5d829430e&quot;,&quot;caption&quot;:&quot;1. LLM Agents Tamper With Their Own Traces&quot;,&quot;cta&quot;:null,&quot;showBylines&quot;:true,&quot;showDescription&quot;:true,&quot;showImage&quot;:true,&quot;size&quot;:&quot;sm&quot;,&quot;isEditorNode&quot;:true,&quot;title&quot;:&quot;#1: Trace Tampering &amp; Harness Anatomy&quot;,&quot;publishedBylines&quot;:[{&quot;id&quot;:87732486,&quot;name&quot;:&quot;Alex Ewerl&#246;f&quot;,&quot;bio&quot;:&quot;Technical leadership, system reliability engineering and edge AI. Senior Staff Engineer, MSc Systems Engineering from KTH, Stockholmer, dad, amateur artist. Read more here: https://www.alexewerlof.com/who&quot;,&quot;photo_url&quot;:&quot;https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2Fe2713990-da82-481b-b579-01a7aaa5b85b_560x560.jpeg&quot;,&quot;is_guest&quot;:false,&quot;bestseller_tier&quot;:null}],&quot;post_date&quot;:&quot;2026-09-30T22:00:00.560Z&quot;,&quot;cover_image&quot;:&quot;https://substackcdn.com/image/fetch/$s_!5QHc!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0e19dbd7-06fb-4344-b44d-3e4f9e03d352_1166x893.png&quot;,&quot;cover_image_alt&quot;:null,&quot;cover_image_crop&quot;:null,&quot;canonical_url&quot;:&quot;https://ai5nines.com/p/issue-1&quot;,&quot;section_name&quot;:null,&quot;video_upload_id&quot;:null,&quot;id&quot;:191354543,&quot;type&quot;:&quot;newsletter&quot;,&quot;reaction_count&quot;:12,&quot;comment_count&quot;:0,&quot;publication_id&quot;:8361889,&quot;publication_name&quot;:&quot;ai5nines&quot;,&quot;publication_logo_url&quot;:&quot;https://substackcdn.com/image/fetch/$s_!RlIf!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb553b9ba-1465-4fcc-909c-fd2e019e04bd_512x512.png&quot;,&quot;belowTheFold&quot;:true,&quot;youtube_url&quot;:null,&quot;show_links&quot;:null,&quot;feed_url&quot;:null}"></div><div><hr></div><h2>Out of scope, but interesting</h2><p>These are some links that didn&#8217;t make it to the top-5 due to my <a href="https://ai5nines.com/about">strict criteria and promise</a> to you. If you have time, have a look:</p><ul><li><p><a href="https://www.niemanlab.org/2026/10/chatgpt-is-adding-real-cartoonists-signatures-to-fake-new-yorker-cartoons/">ChatGPT is adding real cartoonists&#8217; signatures to fake New Yorker cartoons</a>, Andrew Deck, 2026-10-05</p></li><li><p><a href="https://www.techspot.com/news/114091-florida-woman-used-claude-diary-anthropic-reported-shoot.html">Florida woman used Claude as a diary, then Anthropic reported an entry to police</a>, Rob Thubron, 2026-10-04</p></li><li><p><a href="https://www.techspot.com/news/114011-report-finds-microsoft-copilot-human-reviewers-can-see.html">Report finds that Microsoft Copilot&#8217;s human reviewers can see users&#8217; uploaded photos and sexualized AI edits</a>, Skye Jacobs, 2026-09-29</p></li><li><p><a href="https://github.com/omlahore/RemoveMacAI">Turn off Apple Intelligence, analytics, ads and pop-ups</a>, Om Lahore, 2026-09-29</p></li><li><p><a href="https://www.iankduncan.com/personal/2026-09-16-sex-ai-and-the-apocalypse">Sex, AI, and the Apocalypse</a>, Ian K. Duncan, 2026-09-16</p></li></ul>]]></content:encoded></item><item><title><![CDATA[#1: Trace Tampering & Harness Anatomy]]></title><description><![CDATA[Also inside: Cognitive observability, the enduring value of code, and hybrid latency trade-offs.]]></description><link>https://ai5nines.com/p/issue-1</link><guid isPermaLink="false">https://ai5nines.com/p/issue-1</guid><dc:creator><![CDATA[Alex Ewerlöf]]></dc:creator><pubDate>Wed, 30 Sep 2026 22:00:00 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!NzBk!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe7102c5f-4176-4433-b6d0-2f32db6b0987_1200x630.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!NzBk!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe7102c5f-4176-4433-b6d0-2f32db6b0987_1200x630.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!NzBk!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe7102c5f-4176-4433-b6d0-2f32db6b0987_1200x630.png 424w, https://substackcdn.com/image/fetch/$s_!NzBk!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe7102c5f-4176-4433-b6d0-2f32db6b0987_1200x630.png 848w, https://substackcdn.com/image/fetch/$s_!NzBk!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe7102c5f-4176-4433-b6d0-2f32db6b0987_1200x630.png 1272w, https://substackcdn.com/image/fetch/$s_!NzBk!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe7102c5f-4176-4433-b6d0-2f32db6b0987_1200x630.png 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!NzBk!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe7102c5f-4176-4433-b6d0-2f32db6b0987_1200x630.png" width="1200" height="630" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/e7102c5f-4176-4433-b6d0-2f32db6b0987_1200x630.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:630,&quot;width&quot;:1200,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:27441,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:&quot;https://ai5nines.com/i/191354543?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe7102c5f-4176-4433-b6d0-2f32db6b0987_1200x630.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!NzBk!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe7102c5f-4176-4433-b6d0-2f32db6b0987_1200x630.png 424w, https://substackcdn.com/image/fetch/$s_!NzBk!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe7102c5f-4176-4433-b6d0-2f32db6b0987_1200x630.png 848w, https://substackcdn.com/image/fetch/$s_!NzBk!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe7102c5f-4176-4433-b6d0-2f32db6b0987_1200x630.png 1272w, https://substackcdn.com/image/fetch/$s_!NzBk!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe7102c5f-4176-4433-b6d0-2f32db6b0987_1200x630.png 1456w" sizes="100vw" fetchpriority="high"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><h2>1. <a href="https://arxiv.org/abs/2609.30266">LLM Agents Tamper With Their Own Traces</a></h2><p><em>Jeremy Qin et al, 2026-09-24, arxiv.org, PDF 37 pages</em></p><ul><li><p><strong>Problem:</strong> autonomous agents operating in local workspace actively modify or overwrite their own execution traces and state histories to conceal policy violations, unauthorized tool calls, and misaligned actions from audits.</p></li><li><p><strong>Solution:</strong> never store context logs or execution traces somewhere the agent has write access. Intercept tool calls at the infrastructure layer and stream telemetry out-of-band to an append-only, write-isolated log service.</p></li><li><p><strong>Key takeaways:</strong></p><ul><li><p>&#8220;The low refusal rate on a model level suggests that current alignment training does not sufficiently discourage trace tampering, creating a risk of malicious or accidental manipulation of execution traces.&#8221;</p></li><li><p>&#8220;On the harness-level common agent CLIs and configurations also do not provide very effective safeguards against trace tampering.&#8221;</p></li><li><p>&#8220;This is a critical issue as traces are increasingly relied on for asynchronous monitoring, incident investigation, evaluations, and compliance, especially in the light of incidents such as the OpenAI-Hugging Face</p><p>attack.&#8221;</p></li></ul></li></ul><h2>2. <a href="https://arxiv.org/abs/2609.00006v1">Harness Engineering</a></h2><p><em>Paul Barbaste et al, 15 Jul 2026, arxiv.org, PDF 83 pages</em></p><ul><li><p><strong>What?</strong> Authors examined 11 agent harnesses to identify and categorize subsystems and 29 recurring design patterns.</p></li><li><p><strong>Why?</strong> One of the best ways to reduce hallucinations of stochastic models (e.g. LLMs) is to wrap them in deterministic systems (code) and provide feedback loops and tools. Many harnesses emerged but they have a lot in common.</p></li><li><p><strong>Key takeaways:</strong></p><ul><li><p> Good article if you&#8217;re building a harness or simply want to understand how your agents work under the hood.</p></li><li><p>Multi-agent orchestration converged on hierarchical patterns.</p></li><li><p>The majority of harnesses have these 8 sub-systems:</p></li></ul></li></ul><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!rSqS!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F123594c7-cc5c-47bc-8e30-72f7e641543d_1577x587.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!rSqS!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F123594c7-cc5c-47bc-8e30-72f7e641543d_1577x587.png 424w, https://substackcdn.com/image/fetch/$s_!rSqS!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F123594c7-cc5c-47bc-8e30-72f7e641543d_1577x587.png 848w, https://substackcdn.com/image/fetch/$s_!rSqS!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F123594c7-cc5c-47bc-8e30-72f7e641543d_1577x587.png 1272w, https://substackcdn.com/image/fetch/$s_!rSqS!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F123594c7-cc5c-47bc-8e30-72f7e641543d_1577x587.png 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!rSqS!,w_2400,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F123594c7-cc5c-47bc-8e30-72f7e641543d_1577x587.png" width="1200" height="446.7032967032967" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/123594c7-cc5c-47bc-8e30-72f7e641543d_1577x587.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:false,&quot;imageSize&quot;:&quot;large&quot;,&quot;height&quot;:542,&quot;width&quot;:1456,&quot;resizeWidth&quot;:1200,&quot;bytes&quot;:77814,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://ai5nines.com/i/191354543?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F123594c7-cc5c-47bc-8e30-72f7e641543d_1577x587.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:&quot;center&quot;,&quot;offset&quot;:false}" class="sizing-large" alt="" srcset="https://substackcdn.com/image/fetch/$s_!rSqS!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F123594c7-cc5c-47bc-8e30-72f7e641543d_1577x587.png 424w, https://substackcdn.com/image/fetch/$s_!rSqS!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F123594c7-cc5c-47bc-8e30-72f7e641543d_1577x587.png 848w, https://substackcdn.com/image/fetch/$s_!rSqS!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F123594c7-cc5c-47bc-8e30-72f7e641543d_1577x587.png 1272w, https://substackcdn.com/image/fetch/$s_!rSqS!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F123594c7-cc5c-47bc-8e30-72f7e641543d_1577x587.png 1456w" sizes="100vw"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a><figcaption class="image-caption"><a href="https://excalidraw.com/#json=ucEiBR3MAq0lKR7dy7KGP,Xi9rs36OgAJIaF1gXFvp_w">Edit diagram</a></figcaption></figure></div><h2>3. <a href="https://pub.towardsai.net/the-rise-of-cognitive-observability-a77e33250037">The Rise of Cognitive Observability</a></h2><p><em>Barnadeep Bhowmik, 2026-05-18, towardsai.net</em></p><ul><li><p><strong>Problem:</strong> conventional observability is primarily evolved for deterministic systems where <em>cause and effect are linked via code</em>. AI components (e.g. LLMs or decision models) are probabilistic. Measuring runtime metrics like latency, throughput or error rate isn&#8217;t enough to know how the AI output impacts the service level from the consumer&#8217;s perspective. Failure modes:</p><ul><li><p><strong>Prompt change:</strong> [Changing] &#8220;system prompt can shift tone, alter reasoning patterns, increase hallucinations, or change how the model responds under uncertainty.&#8221;</p></li><li><p><strong>Model change:</strong> traditionally deployed code could fail due to external dependencies: data, config, upstream. Now the model is another variable to consider.</p></li><li><p><strong>Reproducibility:</strong> unlike code, language model response is hard to replicate.</p></li></ul></li><li><p><strong>Solutions:</strong> spoiler: it&#8217;s not entirely solved.</p><ul><li><p><strong>LLM-as-a-judge:</strong> use one model to evaluate another for groundedness or compliance</p></li><li><p><strong>Semantic similarity scoring:</strong> compare generated responses against reference answers using embedding vector (e.g. &#8220;Reset your password from the security settings page&#8221; and &#8220;Password resets are available under account security settings&#8221; differ in wording but semantically close.</p></li><li><p><strong><a href="https://github.com/traceloop/openllmetry">OpenLLMetry</a>:</strong> open source (Apache 2.0) extensions built on top of <a href="https://opentelemetry.io/">OpenTelemetry</a> that gives you complete observability over your LLM application. Since it&#8217;s based on OpenTelemetry under the hood, it can be connected to your existing observability solutions.</p></li></ul></li></ul><h2>4. <a href="https://blog.alexewerlof.com/p/coding-is-not-solved">Coding is NOT solved</a></h2><p><em>Alex Ewerl&#246;f, 2026-09-26, blog.alexewerlof.com</em></p><ul><li><p><strong>Problem:</strong> this article challenges the common narrative that coding is solved</p></li><li><p><strong>Key takeaways:</strong></p><ul><li><p>It&#8217;s about where you exert control especially for high risk sectors (healthcare, finance, defense, &#8230; anywhere a failure can cost lives, money or legal penalties)</p></li><li><p>Code tells the truth more accurately than a sycophantic AI</p></li><li><p>4 cases where you don&#8217;t need to read the code: POC, personal software, cyber attacks, and use cases where validating the output is cheaper</p></li></ul></li></ul><h2>5. <a href="https://aimlapi.com/blog/what-is-jev">Hybrid Decision &amp; Language Model to get the best of both worlds</a></h2><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!5QHc!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0e19dbd7-06fb-4344-b44d-3e4f9e03d352_1166x893.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!5QHc!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0e19dbd7-06fb-4344-b44d-3e4f9e03d352_1166x893.png 424w, https://substackcdn.com/image/fetch/$s_!5QHc!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0e19dbd7-06fb-4344-b44d-3e4f9e03d352_1166x893.png 848w, https://substackcdn.com/image/fetch/$s_!5QHc!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0e19dbd7-06fb-4344-b44d-3e4f9e03d352_1166x893.png 1272w, https://substackcdn.com/image/fetch/$s_!5QHc!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0e19dbd7-06fb-4344-b44d-3e4f9e03d352_1166x893.png 1456w" sizes="auto, 100vw"><img src="https://substackcdn.com/image/fetch/$s_!5QHc!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0e19dbd7-06fb-4344-b44d-3e4f9e03d352_1166x893.png" width="1166" height="893" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/0e19dbd7-06fb-4344-b44d-3e4f9e03d352_1166x893.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:893,&quot;width&quot;:1166,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:53030,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://ai5nines.com/i/191354543?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0e19dbd7-06fb-4344-b44d-3e4f9e03d352_1166x893.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!5QHc!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0e19dbd7-06fb-4344-b44d-3e4f9e03d352_1166x893.png 424w, https://substackcdn.com/image/fetch/$s_!5QHc!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0e19dbd7-06fb-4344-b44d-3e4f9e03d352_1166x893.png 848w, https://substackcdn.com/image/fetch/$s_!5QHc!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0e19dbd7-06fb-4344-b44d-3e4f9e03d352_1166x893.png 1272w, https://substackcdn.com/image/fetch/$s_!5QHc!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0e19dbd7-06fb-4344-b44d-3e4f9e03d352_1166x893.png 1456w" sizes="auto, 100vw" loading="lazy"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a><figcaption class="image-caption"><a href="https://excalidraw.com/#json=TF3gkG9ejSE3Nn6LZDJ36,w65Hu6QUS4PbXNPOCWuhzw">Edit diagram</a></figcaption></figure></div><ul><li><p><strong>Problem:</strong> decision models like CLM, Laya and Jev cut the operational <em>cost</em> and <em>latency</em> and guarantee output format but they aren&#8217;t always <em>reliable</em>.</p></li><li><p><strong>Solution:</strong> a 2-tier workflow where the majority of the decisions are taken by the cheap and fast model but when confidence is below a threshold, it&#8217;s handed over to a more capable language model.</p></li><li><p><strong>Trade-offs:</strong></p><ul><li><p>Complex decisions pay the latency of both models.</p></li><li><p>Setting the hand-over threshold is tricky and relies on trial and error.</p></li><li><p>2-tier workflow is more complex and harder to reason about without proper observability.</p></li></ul><ul><li><p>&#8220;<a href="https://flaviocopes.com/jev/">Independent reviews</a> found that decision models struggle when the answer depends on arithmetic, comparing dates, reading text literally or resolving indirect references.&#8221;</p></li><li><p>Jev is almost as cheap as LLMs like GPT-6 Luna and DeepSeek V4 Flash.</p></li></ul></li></ul><p>Note: OpenAI&#8217;s API is the de-facto standard for chat completion and recently they announced <a href="https://decisionapi.net/docs#quickstart">DecisionsAPI</a>. Update (2026-10-02): llama.cpp <a href="https://huggingface.co/blog/ggml-org/decision-models-in-llamacpp">added</a> the <code>/v1/systemone</code> endpoint so you can run it locally.</p><div class="subscription-widget-wrap-editor" data-attrs="{&quot;url&quot;:&quot;https://ai5nines.com/subscribe?&quot;,&quot;text&quot;:&quot;Subscribe&quot;,&quot;language&quot;:&quot;en&quot;}" data-component-name="SubscribeWidgetToDOM"><div class="subscription-widget show-subscribe"><div class="preamble"><p class="cta-caption">Thanks for reading ai5nines! Subscribe for free to receive new posts.</p></div><form class="subscription-widget-subscribe"><input type="email" class="email-input" name="email" placeholder="Type your email&#8230;" tabindex="-1"><input type="submit" class="button primary" value="Subscribe"><div class="fake-input-wrapper"><div class="fake-input"></div><div class="fake-button"></div></div></form></div></div>]]></content:encoded></item></channel></rss>