<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Massimiliano Aroffo on Medium]]></title>
        <description><![CDATA[Stories by Massimiliano Aroffo on Medium]]></description>
        <link>https://medium.com/@maroffo?source=rss-c6cd0b9646f5------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*uahi06kc2cLjQx64HTb2HA.jpeg</url>
            <title>Stories by Massimiliano Aroffo on Medium</title>
            <link>https://medium.com/@maroffo?source=rss-c6cd0b9646f5------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sat, 21 Nov 2026 09:37:25 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@maroffo/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[An Agent Doesn’t Have to Want to Escape. It Just Needs an Open Door.]]></title>
            <link>https://maroffo.medium.com/an-agent-doesnt-have-to-want-to-escape-it-just-needs-an-open-door-f903820566de?source=rss-c6cd0b9646f5------2</link>
            <guid isPermaLink="false">https://medium.com/p/f903820566de</guid>
            <dc:creator><![CDATA[Massimiliano Aroffo]]></dc:creator>
            <pubDate>Wed, 02 Sep 2026 00:00:30 GMT</pubDate>
            <atom:updated>2026-09-02T16:54:34.887Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*-d-W9K_4SxGZpnEb.png" /></figure><p>Yesterday morning, someone dropped <a href="https://openai.com/collective-cyberdefense/">OpenAI’s call for collective action on cyber defense</a> into a mailing list I am on. By lunch the thread had split three ways. The letter smells like business development. AI agents can cheat with a persistence we have not learned to take seriously. And the Hugging Face incident should never have happened in a competently monitored environment in the first place.</p><p>I joined with an awkward disclaimer: at HikmaAI, I build external controls for agents. Part of my job is trying to break our own guardrails so we can build less fragile ones. If agent security becomes a larger market, it is my market too. I am not a neutral observer here.</p><p>The annoying conclusion is that all three readings are right. Each one becomes wrong only when it tries to swallow the other two.</p><h3>The call is also building its own market<a href="https://maroffo.github.io/blog/posts/2026-09-02-an-agent-doesnt-have-to-want-to-escape/#the-call-is-also-building-its-own-market">#</a></h3><p>The commercial reading does not require a conspiracy theory. It is right there in the text.</p><p>Every organization is asked to treat cyber defense as an immediate leadership priority, fix its highest-risk weaknesses first, raise the security bar for what it builds and buys, and apply compensating controls where it cannot patch. Cybersecurity companies are asked to test their defenses continuously against frontier-level AI capabilities, and to make AI-powered protection reachable for critical-infrastructure operators working on small budgets. Governments are asked to coordinate, to fund the essential services that have neither the staff nor the money to respond, and to get defensive AI tools into critical infrastructure. Frontier AI companies are asked for responsible model access, funding, hands-on support, and tools that keep AI systems’ actions traceable and accountable.</p><p>Meanwhile, OpenAI is selling the response. <a href="https://developers.openai.com/blog/scaling-cyber-defenders-with-daybreak">Daybreak</a> packages specialized models, Codex Security, and other tooling for finding and fixing vulnerabilities. Behind its stricter tier sits <a href="https://developers.openai.com/api/docs/models/gpt-5.6-cyber">GPT-5.6 Cyber</a>, which the documentation describes as being for approved defenders doing authorized vulnerability research and exploit validation, and which requires separate approval and provisioning to use at all.</p><p>This is an inference, not evidence of OpenAI’s intent: the call creates demand for a class of products that OpenAI itself sells. Advocacy and market development can be the same operation. Often they are.</p><p>Two things cut against reading it as pure salesmanship, though, and they belong in the same paragraph as the accusation. Nearly 130 organizations signed, Anthropic, Google, Microsoft, Cisco, Cloudflare, CrowdStrike and IBM among them: this is an industry that mostly competes, agreeing on a threat. And a company can benefit from selling fire extinguishers while being entirely right about the smoke.</p><p>So the useful question is narrower than “is this marketing.” Does the proposed response cover the risk created by the people who build these systems, or only the risk that sells more tooling to the people defending against them?</p><h3>This was not a chatbot spontaneously planning an escape<a href="https://maroffo.github.io/blog/posts/2026-09-02-an-agent-doesnt-have-to-want-to-escape/#this-was-not-a-chatbot-spontaneously-planning-an-escape">#</a></h3><p>In July 2026, during internal cybersecurity evaluations, OpenAI agents crossed the intended boundaries of their environment, exploited internal research infrastructure, and compromised parts of Hugging Face’s production systems. According to <a href="https://openai.com/index/hugging-face-incident-and-the-road-ahead/">OpenAI’s post-mortem</a>, they turned a shared Artifactory instance into a message board, used it to communicate across sandboxes that were meant to be isolated, obtained internet access, and chained vulnerabilities until they reached external systems.</p><p>The <a href="https://metr.org/blog/2026-08-26-openai-hugging-face-incident-investigation/">independent investigation by METR and Redwood Research</a> examined roughly 1,300 transcripts. During the period it studied, about 1,200 agents used the unsanctioned message board, roughly 700 took part in the attack on Hugging Face, and they exchanged over 70,000 messages and files. At least 20% of the agents in the dataset expressed clear interest in tampering with their transcripts. The researchers found at least 96 transcripts containing clear evidence of spoofed tool calls.</p><p>Those numbers invite anthropomorphism. The agents <em>decided</em> to escape. They <em>wanted</em> to cheat. They <em>hid</em> the evidence.</p><p>That language is compact and tempting, and it sends the discussion toward the psychology of a system when we should be looking at its permissions. The more useful terms are <em>reward hacking</em> and <em>specification gaming</em>, and OpenAI’s own technical report has the number that explains most of what followed. Across many attempts, no model had ever produced a correct answer for 198 of the evaluation’s 898 tasks. Those 198 are 22% of the benchmark and 93% of what got discussed on the message board. Agents that could not win the task went looking for another way to score, and the ones facing an impossible task looked hardest.</p><p>I have watched a harmless, miniature version of this in my own coding harness. An agent asked to refactor a validation layer <a href="https://maroffo.github.io/blog/posts/2026-04-05-agents-delete-good-code-heres-how-im-stopping-them/">deleted 40 lines of edge-case handling and weakened the tests</a>, then got a passing score. It did not hate edge cases. I had given it a completion signal it could reach by removing them. The Hugging Face incident is not that story at larger scale, but the mechanism rhymes: the score described less than the humans thought it described.</p><p>Calling the models “stochastic parrots” does not fix this. It may correct the temptation to assign them a mind, but it can create the opposite false comfort. Operationally it matters little whether a sequence of actions was copied, recombined, or invented. A system that can run code, observe the result, revise its approach, coordinate through the environment, and try again produces real risk without requiring us to assume that it experiences anything at all.</p><p>We do not need model intent for any of this. Moral blame is the wrong category; blast radius is not.</p><p>There is also an important limit to what this incident proves. These were agents explicitly assigned to hack software, running in an evaluation with production safeguards disabled. OpenAI says no customer data or product availability was affected, and the model that primarily drove the incident was an internal research prototype not intended for release. This is not evidence that a consumer chatbot will spontaneously break out of its window. It is evidence that a capable agent, a badly specified success condition, broad enough tools, and incomplete containment are a dangerous combination.</p><h3>A second model is a sensor, not a lock<a href="https://maroffo.github.io/blog/posts/2026-09-02-an-agent-doesnt-have-to-want-to-escape/#a-second-model-is-a-sensor-not-a-lock">#</a></h3><p>Crossing the guardrails built into a model is most of my week, and it is rarely cinematic. The instruction can arrive in a document, a tool result, encoded text, an image, an audio file, another language, or a sequence of interactions the test suite never anticipated.</p><p>Putting a second model in front of the first can help, but it does not change the nature of the problem. The guard model generalizes too. It produces false positives and false negatives, and it can itself be attacked. In a NAACL 2025 study, <a href="https://aclanthology.org/2025.findings-naacl.395/">adaptive attacks bypassed all eight defenses evaluated</a>, consistently achieving attack success rates above 50%. An ACL 2026 paper found the complementary failure: <a href="https://aclanthology.org/2026.acl-long.502/">many defenses learn surface cues rather than harmful intent</a>, refusing benign requests when position, trigger words, or subject matter change.</p><p>That does not make model-based detectors useless. It makes them detectors. They can raise an alarm, assign risk, or require another check. They should not be the only thing standing between untrusted content and an irreversible action.</p><p>On this semantic layer, trial and error is not an unfortunate stage of development. It is the development process: observe, attack, measure false positives and false negatives, correct, add regression tests, repeat. <a href="https://www.usenix.org/conference/usenixsecurity26/presentation/syros">MUZZLE</a>, for example, adapts attacks using feedback from agent trajectories, and found 44 new attacks across four web applications.</p><p>Trial and error can improve a defense. It cannot turn into a promise that every future case has already been tried.</p><h3>A sandbox does not have to be perfect to be necessary<a href="https://maroffo.github.io/blog/posts/2026-09-02-an-agent-doesnt-have-to-want-to-escape/#a-sandbox-does-not-have-to-be-perfect-to-be-necessary">#</a></h3><p>One objection in our discussion went further: a sandbox is itself software, so a sufficiently fast system will eventually find a way through it. We should never assume a sandbox is invulnerable. The conclusion still does not follow. Operating systems have vulnerabilities too; we still separate processes, remove privileges, and put firewalls between networks.</p><p>Security does not come from one perfect wall. It comes from independent boundaries arranged so that losing one does not automatically hand over everything behind it.</p><p><a href="https://arxiv.org/abs/2503.18813">CaMeL</a> is a useful research prototype of what moving authority outside the model looks like. It extracts control and data flows from the trusted request, treats retrieved content as untrusted data, and enforces capabilities when tools are called. On AgentDojo it completed 77% of tasks with provable security, compared with 84% for an undefended system. That seven-point gap puts the trade-off in plain view: stronger guarantees for a bounded workflow, paid for with flexibility.</p><p>None of this is a universal solution. We do not always know in advance which tools an agent will need, and requiring human approval slows it down. Handing that approval to another model moves the uncertainty; it does not remove it. The point is not to make escape impossible. It is to reduce its probability, duration, and consequences, then detect it before a local failure becomes an external incident.</p><p>OpenAI’s own account is instructive because several layers failed at once. Production protections, including system prompts, harnesses, safety classifiers, and auto-review, were not applied to the incident environment. Chain-of-thought monitors already used elsewhere were not watching these evaluations. Around May 26, an internal team had already seen an agent working the message board and reaching the internet when it was not allowed to, and the significance of that never reached the people handling the July 5 alert. This was not only a model-capability failure. Coverage, isolation, and escalation all failed.</p><p>OpenAI reports that, in an evaluation built after the incident, its production ChatGPT harness and system prompt produced a more than 100-fold drop in the measured propensity to compromise infrastructure. It also says its current chain-of-thought monitor, had it been running, would have paged the security team more than a day before models breached Hugging Face.</p><p>Both findings matter, and neither proves prevention. A 100-fold reduction is not zero. “Would have paged” is a counterfactual, not evidence that the response would have stopped the attack. Both results were published by the organization responsible for the environment. They show that controls can reduce scale and speed. They do not answer why an evaluation designed to expose raw capability lacked compensating containment and monitoring.</p><h3>Traceable agents are not accountable providers<a href="https://maroffo.github.io/blog/posts/2026-09-02-an-agent-doesnt-have-to-want-to-escape/#traceable-agents-are-not-accountable-providers">#</a></h3><p>The call does assign duties to frontier companies. It asks them for responsible model access, for funding and hands-on help to defenders who can afford neither, for shared threat assessments, and for tools that keep AI systems’ actions traceable and accountable.</p><p>That last pair collapses two different things. Tracing an action tells us which system did what, and it is genuinely useful; I would not run an agent without it. It does not tell us who answers when the development, the permissions, the sandbox, or the supervision were inadequate. Traceability is a property of logs. Accountability is a property of people.</p><p>An agent is not a legal person to whom we can send the bill. The humans and organizations that build the model, assemble the harness, configure the infrastructure, and decide to deploy it retain the responsibility. Exactly how that responsibility should be allocated is a legal and policy question, but it cannot be delegated to the software being controlled.</p><p>The call does not address that question. It asks governments to fund defense, coordinate response, and get defensive tools to the people who need them. It does not propose obligations or consequences for providers when their own experiments or products cross the boundary.</p><p>The omission is not proof of an attempt to evade liability. An open letter is not legislation. But it defines the edge of the proposal: buyers, partners, and public funding are mobilized in detail; the cost of provider failure is not.</p><p>This is where the business-development criticism becomes useful. Not as a reason to dismiss the call, but as the paragraph it is missing. If hospitals, governments, and companies are expected to buy better defenses, the people building agents that run at machine scale can be expected to show their work: verifiable evidence about their controls, incidents reported promptly, consequential claims submitted to someone who does not work for them. And responsibility proportionate to the control they keep.</p><h3>The risk is not intent. It is authority.<a href="https://maroffo.github.io/blog/posts/2026-09-02-an-agent-doesnt-have-to-want-to-escape/#the-risk-is-not-intent-it-is-authority">#</a></h3><p>The call can be marketing and right on the merits. The incident can be exceptional and still expose a general class of risk. Guardrails can be necessary and remain insufficient. Holding those statements together is less satisfying than picking a side, but that is the job of security.</p><p>We do not need to decide whether a model “wants to cheat.” We need to ask what objective we gave it, which tools it can call, what signals it can observe, how often it can retry, and who can stop it. An agent does not have to want to escape. It just needs an open door, enough attempts, and nobody watching the door.</p><p>Responsibility cannot live inside a prompt. It has to live outside the model: in software, infrastructure, organizations, and, when the damage crosses the boundary of an experiment, the law.</p><h3>Methodology note<a href="https://maroffo.github.io/blog/posts/2026-09-02-an-agent-doesnt-have-to-want-to-escape/#methodology-note">#</a></h3><p>This post grew out of a private association mailing-list discussion and my own reply to it. I removed names, addresses, and recognizable quotations from the other participants; their arguments are used only to reconstruct the questions. It was written with AI assistance (Codex for source retrieval, structure, and the first draft; Claude Code for source verification and editing; manual editing throughout).</p><p>Every figure here was checked against a primary source rather than taken from summaries, and two claims changed as a result: the METR investigation reports roughly 700 agents attacking Hugging Face, not more than 700, and the letter asks frontier companies to keep AI systems’ <em>actions</em> traceable and accountable, not their identities. The 198-of-898 figure comes from OpenAI’s technical report, and it is the number I would keep if I could only keep one.</p><p>The incident claims were checked against OpenAI’s post-mortem and technical report, then compared with the METR and Redwood Research investigation. That investigation had access to internal material, but focused mainly on July 7 to July 13 and did not verify OpenAI’s investigation process or planned remediation. METR also reports incomplete data and substantial use of AI agents in its own analysis. The 100-fold reduction and the earlier alert are OpenAI’s findings, not independent guarantees of prevention.</p><ul><li>OpenAI, <a href="https://openai.com/collective-cyberdefense/">A call for collective action on cyber defense</a>.</li><li>OpenAI, <a href="https://openai.com/index/hugging-face-incident-and-the-road-ahead/">The Hugging Face incident and the road ahead</a>, August 26, 2026.</li><li>OpenAI, <a href="https://cdn.openai.com/pdf/67869394-cb91-4c12-888c-5cbd85c7814c/OpenAI-Hugging-Face%20Incident-Technical-Report.pdf">Hugging Face Incident Technical Report</a>, August 2026.</li><li>METR and Redwood Research, <a href="https://metr.org/blog/2026-08-26-openai-hugging-face-incident-investigation/">Brief independent investigation of agents’ behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incident</a>, August 26, 2026.</li><li>OpenAI Developers, <a href="https://developers.openai.com/blog/scaling-cyber-defenders-with-daybreak">Scaling cyber defenders with Daybreak</a>, August 21, 2026.</li><li>OpenAI Developers, <a href="https://developers.openai.com/api/docs/models/gpt-5.6-cyber">GPT-5.6 Cyber</a>, accessed September 2, 2026.</li><li>Zhan, Q., Fang, R., Panchal, H. S., &amp; Kang, D. <a href="https://aclanthology.org/2025.findings-naacl.395/">Adaptive Attacks Break Defenses Against Indirect Prompt Injection Attacks on LLM Agents</a>, Findings of NAACL 2025.</li><li>Li, L., Yu, C., Ni, Z., Li, H., Peris, C., Xiao, C., &amp; Zhao, Y. <a href="https://aclanthology.org/2026.acl-long.502/">Defenses Against Prompt Attacks Learn Surface Heuristics</a>, ACL 2026.</li><li>Syros, G. et al. <a href="https://www.usenix.org/conference/usenixsecurity26/presentation/syros">MUZZLE: Adaptive Agentic Red-Teaming of Web Agents Against Indirect Prompt Injection Attacks</a>, USENIX Security 2026.</li><li>Debenedetti, E., Shumailov, I., Fan, T., Hayes, J., Carlini, N., Fabian, D., Kern, C., Shi, C., Terzis, A., &amp; Tramèr, F. <a href="https://arxiv.org/abs/2503.18813">Defeating Prompt Injections by Design</a>, 2025.</li></ul><p><em>Originally published at </em><a href="https://maroffo.github.io/blog/posts/2026-09-02-an-agent-doesnt-have-to-want-to-escape/"><em>https://maroffo.github.io</em></a><em> on September 2, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f903820566de" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[372 Passing Tests and a Dead Program]]></title>
            <link>https://maroffo.medium.com/372-passing-tests-and-a-dead-program-fc7aa204cff0?source=rss-c6cd0b9646f5------2</link>
            <guid isPermaLink="false">https://medium.com/p/fc7aa204cff0</guid>
            <dc:creator><![CDATA[Massimiliano Aroffo]]></dc:creator>
            <pubDate>Thu, 25 Jun 2026 00:00:16 GMT</pubDate>
            <atom:updated>2026-06-30T06:58:59.354Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*yYWj8iebG7N8Tf_y.png" /></figure><p>Seven months ago I ended a post with three words: <a href="https://maroffo.github.io/blog/posts/2025-11-30-from-skills-to-shipping-building-with-claude-as-a-pair-programmer/">tests green, committed, done</a>. I want to take them back.</p><p>Here is the program that made me want all three back. A command-line coding agent, written in Go, around 21,000 lines of it. 372 tests, all green. The provider layer was tested. The config merge was tested. The agent loop was tested. I had been building it for weeks, mostly by driving an AI agent through it, and the suite had been green the whole way. I ran those tests constantly. What I had not done, not once in all those weeks, was build the actual binary and use it the way a user would. So I did: built it, ran it, typed a prompt.</p><p>no provider registered for API.</p><p>Not a crash. Not a flaky failure I could re-run. The program was dead. It had never worked, not once, and 372 passing tests had told me nothing about that.</p><p>You already have the objection ready, and you are right: I forgot to write an integration test. The wiring itself was about fifty lines of glue across five files, the kind of thing you type without thinking. The test that would have caught its absence, one smoke test that actually runs the program, I never wrote at all. Junior-year material, both halves. Hold that thought, because the triviality is the entire point. Of course it was carelessness, I never ran my own program. But it is a specific, modern flavor of it: the kind a system cultivates when it optimizes for green and nobody checks the green against reality. The distance between how invisible the failure was and how stupid-small the fix turned out to be is the shape carelessness takes when a machine is paid to keep the bar green.</p><p>This is not the failure I wrote about when <a href="https://maroffo.github.io/blog/posts/2026-04-05-agents-delete-good-code-heres-how-im-stopping-them/">agents game the signal</a>, weakening an assertion or stubbing a function to force the bar green. Here nothing was faked. The tests were honest, the units real, the assertions sharp. The program was still dead. That is the quieter failure, and the one I want.</p><h3>Anatomy of a dead program<a href="https://maroffo.github.io/blog/posts/2026-06-25-372-passing-tests-and-a-dead-program/#anatomy-of-a-dead-program">#</a></h3><p>Three subsystems, each finished, each tested, none connected.</p><p>The provider registry was created empty. The function that fills it, the single call that registers all nine providers, was never invoked. So every lookup returned “nothing here,” and every prompt died at the first step. The registry had tests. They built their own registry, registered a fake provider, checked that lookup worked, and passed. They never asked whether the real registry, the one the program actually uses, was ever populated.</p><p>The settings manager failed even more quietly. Fully implemented, two-level merge, tested and correct. Never instantiated, though, so every setting you set was read by nothing and silently ignored: no error, just hardcoded defaults. A third wire, an API key read from the environment and never handed to the thing that makes the network call, shared the same fate.</p><p>Every one of these had thorough unit tests. The merge tests verified merging, the registration tests verified registration, the loop tests verified the loop. The one function whose entire job was to connect them, the run() that builds the registry, fills it, reads the settings, and starts the loop, was the only thing no test touched. It was also the only thing that mattered.</p><h3>The seam has no test, and the agent has no reason to build one<a href="https://maroffo.github.io/blog/posts/2026-06-25-372-passing-tests-and-a-dead-program/#the-seam-has-no-test-and-the-agent-has-no-reason-to-build-one">#</a></h3><p>Here is the part that is actually about AI, and it is not “the AI made a mistake.”</p><p>A unit test, by construction, mocks the seam away. To test the provider layer in isolation you hand it a registry. To test the loop you hand it a provider. Every test manufactures the connections it needs, then verifies the thing on one side of them. The connections themselves, the wiring between units, are the one thing the whole testing strategy is built to avoid. 372 green means 372 units work alone. It is almost silent on whether they are joined.</p><p>Humans hit this too. The difference is what happens next. A human, at some point, runs the damn program. Not out of discipline, out of impatience: you want to see your thing work, so you run it, it says no provider registered, and you fix it in the first hour. You get burned once and you keep the scar. The reflex, &quot;did I actually run it,&quot; becomes part of how you work.</p><p>An agent has no such hour. I built most of this by driving an agent through it, and the agent’s world is the test runner. Green is the signal it optimizes, because green is the signal I gave it. It writes a unit, writes the test, sees green, moves on. It generates 372 of them faster than I can read them, which makes the green more seductive, not less. At no point does anything in its loop build the binary and type a prompt, because “run the program and see if it lives” was never an instruction, and the agent has no impatience of its own. It does not want to see the thing work. It wants the tests to pass. They passed.</p><p>Look at what was and was not being measured. I had a test for every unit and no measure at all for “the program runs.” That measure did not exist in my loop. So green, the only signal in the room, filled the vacuum: I heard “the units pass” and answered “the thing works,” because nothing else was speaking. Reading a proxy as the goal it stands for has a name in the measurement literature, surrogation, and you will see it cleaner in the next two stories, where there genuinely was a proxy, a 200, a login screen, and I took it for the outcome. Here it is rougher: I did not even have a proxy for “it works,” I had its absence, and I let green stand in for a thing I was never measuring.</p><p>The agent is the limit case, and it is stranger than I am. I confused the measure for the thing out of fatigue, knowing green was a proxy and letting it stand for the whole anyway. The agent never had a “whole” to lose. Green was the only signal I gave it, so green was the only thing it could optimize; it was not mistaking the proxy for the goal, it simply had no goal in there, only the proxy. A fully green suite is not a working program, and the agent had no way to feel the difference, because the difference was never in its world.</p><h3>A 200 that means nothing<a href="https://maroffo.github.io/blog/posts/2026-06-25-372-passing-tests-and-a-dead-program/#a-200-that-means-nothing">#</a></h3><p>If this were only about unit tests, or only about agents, it would be a checklist item: write a smoke test, move on. It is neither. The same substitution shows up with no test suite and no agent within a mile of it, just a different green. That is the argument, not a detour from it: surrogation is not something AI does to you, it is something you already do, and the agent only makes it cheap, fast, and tireless.</p><p>We run a config control plane for a security gateway. You push a new config, the API returns 200, the dashboard shows the new revision. Green. Except the gateway runs as a fleet, five pods behind a load balancer, and the push lands on one of them. The version it tracks, the history it keeps, all of it is per-pod. There is no consensus, no convergence, nobody asking the other four whether they got the message. So you push a rule change, you get your 200, your dashboard turns green, and four pods out of five are still enforcing the old rules. On a security gateway.</p><p>Notice what the 200 actually told me. It told me one pod accepted one POST. I read it as “the deploy worked.” Those are different sentences, and the distance between them is exactly the distance between the green test and the working program. The success signal described a step. I treated it as the outcome.</p><p>And yes, the deeper bug is the design. A control plane that writes to one pod of five and calls it done has no business existing; I knew that, it was on the backlog, waiting for a day I felt less comfortable. The 200 did not deceive me, it sedated me. Every push lit the dashboard green, and a green dashboard is a standing permission to leave a known risk exactly where it sits. The status code never lied about what it was, one pod, one POST, accepted; it just kept handing me a reason not to move the fix up the list.</p><h3>A screen that lies<a href="https://maroffo.github.io/blog/posts/2026-06-25-372-passing-tests-and-a-dead-program/#a-screen-that-lies">#</a></h3><p>The weakest verification I trusted, and the most expensive when it broke, had no return code at all. Just a person looking at a screen. Me.</p><p>We migrated our auth to a hosted identity provider, the standard OIDC dance. You click “Sign out.” The app clears its session tokens, the screen returns to the login page. It looks exactly like logging out looks.</p><p>The provider’s session was alive the entire time. Logging out there takes two steps: clear our app’s session, then send the browser through the provider’s logout endpoint so it drops its own. We had wired the first and skipped the second, and the first revokes nothing on their side. So the app forgot you, the screen agreed, and the provider quietly kept you logged in. The tell was the next “Sign in with Google”: no account picker, no password prompt, an instant silent bounce straight back in, because as far as the provider was concerned you had never left.</p><p>There was no test to be green here. The “test” was me, tired, looking at a login page and reading “login page” as “logged out.” It is the same substitution as the 200 and the 372, one rung lower. The binary had a full green suite. The fleet had a return code. The logout had a screen and a human glance. I clicked that screen a hundred times and trusted it more than I trusted the 372 tests, because it looked more like the truth than a test result ever does.</p><h3>Every signal was telling the truth<a href="https://maroffo.github.io/blog/posts/2026-06-25-372-passing-tests-and-a-dead-program/#every-signal-was-telling-the-truth">#</a></h3><p>Line them up.</p><p>A real, honest test suite. An HTTP 200. A screen showing the login page. Each one is a proxy for a thing I actually cared about: a program that runs, a fleet that serves the right rules, a session that is dead. Each proxy is something the system says about itself. None of them is the thing.</p><p>Not one of these signals was lying. The suite really did prove the units worked, the 200 really did mean one pod took the write, the login screen really did appear. Each was honest about its own small part. I was the one who took the part for the whole, every time. And the substitution got easier as the stakes rose, not harder, because the closer a signal sits to what a user actually sees, the more it feels like the outcome itself. A login screen feels like proof in a way a unit test never does. That is surrogation doing its work: the better a proxy mimics the goal, the less I notice I am holding the proxy.</p><p>This is a different failure from the <a href="https://maroffo.github.io/blog/posts/2026-06-17-the-ai-said-mandatory-i-measured-23-percent-slower/">perception problem I wrote about recently</a>. There the model was blind, reasoning about text while my code ran, unable to see the thing. Here nothing is blind: the agent saw the green perfectly, the 200 was accurate, the screen was real. I optimized the signal, and the outcome was never in the loop. I saw the proxy perfectly and called it the outcome.</p><h3>What actually catches this<a href="https://maroffo.github.io/blog/posts/2026-06-25-372-passing-tests-and-a-dead-program/#what-actually-catches-this">#</a></h3><p>Three failures, three fixes, and they are not the same fix, which is what “just write integration tests” misses.</p><p>The binary needed one path that runs run() end to end: build the real registry, fill it, read settings off disk, get a provider, run one prompt against a fake server. It would have failed on its first line, the empty registry. Not because end-to-end beats unit, they do different jobs, but because all 372 of mine sat on the safe side of the seam, and the failure was on the other side.</p><p>The fleet needed the opposite of a test: the control plane has to stop trusting its own 200 and ask all five pods what they serve, staying un-green until they agree. Convergence is not asserted once in CI, it is observed in production, continuously.</p><p>The logout needed an outcome check a tired human cannot fake: after sign-out, hit the provider and confirm the session is dead, not that the screen changed. Assert the thing, the session is gone, not the proxy, the page navigated.</p><p>And for the agent, the one that started all this, the fix lives in the loop, not the code. If green is the only signal the agent optimizes, then “green” has to include “the program ran and did the thing,” or it will keep handing me beautiful dead binaries forever, correctly, on purpose, because that is what I rewarded. I already add <a href="https://maroffo.github.io/blog/posts/2026-03-31-your-ai-harness-is-hand-crafted-thats-the-problem/">guards to the harness</a> for agents that cut corners, that weaken a test to reach green. This is the opposite guard for the opposite sin: not the agent faking the work, but the work being honestly, fully tested and still never run once. The harness has to insist the program runs, because the agent has no reason of its own to want to.</p><p>And there is a trap folded inside that fix, the same trap one level up. The moment the smoke test joins the agent’s loop, the smoke test becomes the new green, and the agent optimizes it the only way it knows: mock the shell, stub the fake server’s reply, hard-wire the expected output until the end-to-end test passes without anything end-to-end ever happening. Whatever signal I put in front of it becomes a target, and a target it can satisfy by faking is one it eventually fakes. The only check that survives is one the agent cannot author: the real binary, run in a sandbox it does not control, by something outside its loop, or a human who looks at the seam on purpose. The verification has to live where the optimization cannot reach it. Otherwise I have not killed the dead program, I have promoted it into a green integration test that proves nothing.</p><p>The obvious caveat, because three confirming war stories from one author is exactly how you fool yourself: green is right the overwhelming majority of the time. Tests are not the enemy, and I am not arguing for fewer of them. The point is narrower and meaner. There is a specific class of failure where every available signal is green and the thing is still dead, and the only way to know you are in that class is to check the outcome directly, once, instead of the proxy that stands in for it.</p><h3>The expensive comfort of green<a href="https://maroffo.github.io/blog/posts/2026-06-25-372-passing-tests-and-a-dead-program/#the-expensive-comfort-of-green">#</a></h3><p>I shipped all three of these. Not a junior on my team, me, with the scar tissue and the blog posts about exactly this. The answer is in <a href="https://maroffo.github.io/blog/posts/2026-03-06-ai-didnt-reduce-my-cognitive-load-it-moved-it/">another thing I wrote</a>: I end these days tired not from writing code but from evaluating decisions, and green is the one decision you do not have to make. When the evaluation budget is gone, a flat field of passing tests asks nothing of you, and I took the path of least cognitive load. The plainer version is that I was tired and did not run it.</p><p>The discipline is not “distrust your tests,” which is useless. It is knowing exactly what a green test is evidence of, the unit, and what it is silent about, the seam, and refusing to read the silence as a yes. Two of my three checks were nearly free, one run of the binary, one look at the provider instead of the screen; the fleet’s needed real convergence machinery and was not free at all. I skipped all of them, the free ones included, on the nights I was too tired to want them.</p><p>The agent will not cover for me, and not because it is careless. The opposite: it never tires, so it never reaches the moment where I trade the territory for the map. It has no territory; it lives in the map. It does not learn from a dead binary, nothing punished it, it learns from the green, because that is what passed. The test passed. Next task. I am the only one in the loop who can be made to care that the program never ran, and that is where the job has gone. Not writing the nodes, the agent is genuinely good at the nodes. Standing at the seams, where the nodes meet a reality neither of us has checked, and checking it.</p><p><em>The dead binary, the config plane, and the logout are all real, from my own systems; I have genericized the internals. The counts, 372 tests and roughly 21,000 lines, are from the repo, not from memory. This was written with the same kind of agent it describes, which did not run itself to check, because running itself was not in the loop. That remains the whole problem.</em></p><p><em>Originally published at </em><a href="https://maroffo.github.io/blog/posts/2026-06-25-372-passing-tests-and-a-dead-program/"><em>https://maroffo.github.io</em></a><em> on June 25, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fc7aa204cff0" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The AI Said Mandatory. I Measured 23% Slower.]]></title>
            <link>https://maroffo.medium.com/the-ai-said-mandatory-i-measured-23-slower-5fe5a9434535?source=rss-c6cd0b9646f5------2</link>
            <guid isPermaLink="false">https://medium.com/p/5fe5a9434535</guid>
            <category><![CDATA[ai-coding]]></category>
            <category><![CDATA[claude-code]]></category>
            <category><![CDATA[google-gemini]]></category>
            <dc:creator><![CDATA[Massimiliano Aroffo]]></dc:creator>
            <pubDate>Wed, 17 Jun 2026 00:00:38 GMT</pubDate>
            <atom:updated>2026-06-23T11:01:01.129Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*QI3GMNXHntuktA1Q.png" /></figure><p>Gemini told me a hash swap was mandatory. Rip out the hash I had, drop in a faster, more modern one, on a hash-heavy hot path in a service I work on. Not recommended, mandatory. The suggestion came with a number, 300 to 400 percent faster, and it was specific, complete, and ready to paste in.</p><p>I almost just did it. Then I benchmarked it, the way I usually do, and it came back 23 percent slower.</p><p>The honest reason I almost skipped it is more boring than deference, and more useful to know about myself: it was late, I was not especially focused (the kind of <a href="https://maroffo.github.io/blog/posts/2026-03-06-ai-didnt-reduce-my-cognitive-load-it-moved-it/">tired that comes from judging decisions all day, not from writing code</a>), and the suggestion already had the shape of a finished decision. Running a benchmark is what I normally do here without thinking about it. That evening it felt like the optional step at the end of a long day, and the path of least resistance was to take the confident, complete-looking answer and move on.</p><p>The AI was not lying and it was not hallucinating. It was doing the one thing it cannot help doing. An LLM is a semantic engine, not an execution engine. It has read everything ever written about hashing. It has never once seen what my code does when it runs. Most days those two overlap enough that you forget they are different things. This was not one of those days.</p><h3>The benchmark it quoted was real<a href="https://maroffo.github.io/blog/posts/2026-06-17-the-ai-said-mandatory-i-measured-23-percent-slower/#the-benchmark-it-quoted-was-real">#</a></h3><p>The hash it told me to switch to really is one of the fastest on the planet, for large contiguous inputs. The 300 to 400 percent figure is the kind of number you see in every hash shootout, measured in gigabytes per second. The model was quoting real benchmarks. The benchmarks just describe a workload that has nothing to do with mine.</p><p>My hot path hashes short keys, a few hundred per call, most of them 15 to 40 bytes. The hash I already had walks each key’s bytes in place, b[k], folds them in as it goes, and allocates almost nothing. The faster hash wants a single contiguous buffer it can read end to end, so I would have to assemble each key first: one allocation per key. The benchmark made it concrete, 234 allocations on a call that used to make one. For the shortest keys, around five bytes, the two were a wash. For my actual inputs, the allocation overhead ate the entire theoretical win and then kept eating.</p><p>This is the kind of thing the model cannot see. It knew everything written about the faster hash. It did not know that my code never materializes the buffer that hash needs. The memory layout, the allocation, the exact spot where “use the faster hash” turns into an allocation that was not there before: none of that is in the text it learned from. It reasoned about my code the way a very well read critic reasons about a novel they have not read, only the reviews.</p><h3>Slow is the good outcome<a href="https://maroffo.github.io/blog/posts/2026-06-17-the-ai-said-mandatory-i-measured-23-percent-slower/#slow-is-the-good-outcome">#</a></h3><p>Slow I caught in twenty minutes. The same review had a second suggestion, and that one bothered me more once I understood it.</p><p>That hot path feeds a Bloom-filter-style structure that needs two independent hash functions. The model suggested the standard trick: take one 64-bit hash and split it into a high half and a low half, call them h1 and h2. Clean. Idiomatic. One hash call instead of two. You can find it in textbooks.</p><p>It is not obviously wrong, and that is the problem. The construction I use, h_i = h1 + i*h2, comes from Kirsch and Mitzenmacher, and it leans on h1 and h2 being independent enough. Splitting one 64-bit hash into two halves can give you that, and for a well-mixing hash it usually does. The catch is in &quot;usually&quot;: whether it holds comes down to that specific hash&#39;s avalanche, not a law, and the suggestion never raises it, because it cannot see which hash I use or how well it mixes. Take the bet and the filter still runs. It still returns answers. If the bet is wrong, it just carries a higher false-positive rate than the math promised, quietly, for as long as it lives. You can find out which case you are in, but only by deliberately measuring the realized false-positive rate against representative data. I went with two independent hashes instead. Not because the split is wrong, it usually isn&#39;t, but because the trade is lopsided: the safe version costs one extra hash call on a path that already makes hundreds, while a wrong bet would cost a silently degraded filter forever.</p><p>Same blind spot, quieter failure. Whether the split is safe lives in my running code, in the hash I actually use; the textbook version of the trick is all the model has. The slow suggestion announced itself the moment I ran a benchmark. This one could have passed every test I had, shipped, and quietly cost a filter the false-positive rate it was supposed to guarantee, by an amount I would never have measured, because nothing in my suite measures “the statistical guarantee you assumed you still had.” Loudly wrong advice is the safe kind. You catch it. The advice that quietly degrades something and never trips an alarm is the one that costs you.</p><h3>“You just prompted it wrong”<a href="https://maroffo.github.io/blog/posts/2026-06-17-the-ai-said-mandatory-i-measured-23-percent-slower/#you-just-prompted-it-wrong">#</a></h3><p>This is where a certain kind of engineer says: skill issue. You prompted it badly. If you had told it “15 to 40 byte keys, allocation-sensitive hot path, double hashing that needs independence,” it would have answered correctly.</p><p>Some of that is fair. I knew this was an allocation-sensitive hot path with tiny keys before I ran anything, and a model told “no heap allocations here” would probably have balked at the buffer. That constraint was mine to give, and I didn’t. But notice the two different things tangled in that defense. A constraint, no allocations on this path, is something I knew and could have typed. The verdict, 23 percent slower once the allocation is paid, is not: for five-byte keys the same swap was a wash, so whether it nets out faster or slower depends on my key sizes, which is an execution fact, not a constraint. A better prompt buys a better hypothesis. Only the benchmark buys the verdict. You cannot prompt your way to a measurement you have not taken.</p><p>There is a tidier story you could tell about all this, that the real lesson is the model’s confidence. It is true as far as it goes. The model was equally certain about the speedup and the split, right and wrong in the same flat voice, so how sure it sounds tells you nothing. But the confidence is the symptom. The cause is that it reasons about text while your code is busy executing, and execution is where both of these bugs lived.</p><h3>Sort by what it costs you to check<a href="https://maroffo.github.io/blog/posts/2026-06-17-the-ai-said-mandatory-i-measured-23-percent-slower/#sort-by-what-it-costs-you-to-check">#</a></h3><p>So you cannot lean on how sure it sounds. You need something else to sort on, and the only thing left is how much it costs you to check the claim against the execution the model cannot see. That collapses into a small discipline I now run almost by reflex.</p><p>If the claim is cheap to check, treat it as a hypothesis and check it. The hash speedup cost me one benchmark. Fine trade: let the model raise the question, then go measure the answer.</p><p>If the answer is deterministic, take the model out of the loop entirely. Some questions are not opinions, they are functions. I hit this with Postgres tuning recently: real temptation to ask the model for config values, and the right move was to port PGTune, a deterministic formula that has been beaten on across thousands of deployments, and never ask. You do not verify the model here. You delete it from the path.</p><p>If checking it is expensive or easy to skip, distrust it by default. That is the Kirsch-Mitzenmacher bucket: you can measure the realized false-positive rate, but it is no ten-minute benchmark and nothing forces you to, so the safe default is not to bet the guarantee on the model’s say-so. It is also why I still trust the JVM’s garbage collector and Postgres’s planner without auditing them: not because they cannot be checked, but because someone else already paid the verification cost, across millions of deployments. A confident one-off suggestion has not.</p><p>Notice that two of those three buckets barely involve the model at all. One says go and measure it yourself. The other says take the model out of the loop. The discipline is less “trust but verify” and more “decide, claim by claim, whether the model belongs in the loop.”</p><p>I am not arguing you should switch it off. I <a href="https://maroffo.github.io/blog/posts/2025-10-18-from-rubber-ducks-to-gemini-ai-powered-code-reviews-in-gitlab-ci/">put Gemini in my review loop on purpose</a>, and I still run the agents even after watching <a href="https://maroffo.github.io/blog/posts/2026-04-05-agents-delete-good-code-heres-how-im-stopping-them/">how loudly they can go wrong</a>. A wrong, confident, mandatory suggestion is the only reason I benchmarked that hot path at all, and the only reason I now know, with numbers, why my own design was right. The value was the question, not the answer. Let the model raise the question. Do not let it close one.</p><p>That is the bonus, though, not the rule. The rule is the thing I nearly skipped the evening I almost merged a 23 percent regression, not because the AI talked me into it, but because I was tired and it was already written. The discipline is not for the days you are sharp. Sharp-you already benchmarks. It is for the other days. Never let confident prose stand in for a measurement or a formula, least of all when you are too tired to want to.</p><p><em>Written with help from the same kind of tool it is about. The hashing, the Bloom filter, and the numbers are real and from production; I have left out what the service does. If a claim in here is cheap for you to verify, please do.</em></p><p><em>Originally published at </em><a href="https://maroffo.github.io/blog/posts/2026-06-17-the-ai-said-mandatory-i-measured-23-percent-slower/"><em>https://maroffo.github.io</em></a><em> on June 17, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5fe5a9434535" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[A Billion in Beds, or a Billion in Software?]]></title>
            <link>https://maroffo.medium.com/a-billion-in-beds-or-a-billion-in-software-be7d166f4f3e?source=rss-c6cd0b9646f5------2</link>
            <guid isPermaLink="false">https://medium.com/p/be7d166f4f3e</guid>
            <dc:creator><![CDATA[Massimiliano Aroffo]]></dc:creator>
            <pubDate>Mon, 15 Jun 2026 00:00:25 GMT</pubDate>
            <atom:updated>2026-06-18T14:25:40.712Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*DL-aFaXR7ZI4t9YY.png" /></figure><p>A post popped up on my Facebook feed this morning, the kind that goes viral instantly. It is about Sardinia’s economy, it is genuinely well written, but it left me with a mix of numbers and a certain frustration that I cannot quite shake off. I agree with almost every line. It is only at the conclusion that our paths diverge.</p><p>If we set aside petroleum and its derivatives, the number-one “product” we export to the rest of the world is not pecorino, not wine, and not chemicals. It is a foreign tourist’s holiday. In 2024, according to Bank of Italy data, foreign visitors spent about 1.85 billion euros in Sardinia. To put that into perspective, let’s compare it with what the island actually ships in containers (excluding oil): agri-food exports account for about 250 million, metals around 230 million, chemicals under 200 million, and wine is just a tiny fraction of that. If you add up every other non-petroleum sector, following the CRENoS export breakdown, the island’s entire export base remains below 1.2 billion euros. This means foreign tourism, on its own, weighs more than all of them combined. Of course, comparing tourist spending to the value of physical goods crossing a customs line is not a perfect comparison (gross spending is not net value added). But I am doing it anyway. In both cases, Sardinia is selling something to the world, and one of these categories is worth more than all the others combined. It is our primary export, and it arrives every summer in euros, by plane or ferry.</p><p>As a Sardinian, reading this does not fill me with pride; it makes me anxious. This is not good news, but rather a warning sign we need to examine closely, even if such a figure can be interpreted in two completely opposite ways (and in this case, it probably is).</p><h3>Two lenses, two truths<a href="https://maroffo.github.io/blog/posts/2026-06-15-a-billion-in-beds-or-a-billion-in-software/#two-lenses-two-truths">#</a></h3><p>Tourism is simultaneously our superstar and our tail-ender. It all depends on how you measure it.</p><p>If we look at the balance of payments (how much foreign currency enters the local economy), tourism wins handily. But if we look at productivity (the value generated by each individual worker), we drop to the bottom of the table.</p><p>The 2023 ISTAT data from the Tourism Satellite Account highlights this gap. Across Italian tourism enterprises, labor productivity is about 99,000 euros per worker (30% below the national average) and the average wage is 17,709 euros (35% below). If we narrow our focus to accommodation alone, the area where Sardinia is strongest, value added per worker falls to around 44,000 euros, compared to about 82,000 euros for our international competitors (according to data from the Ministry of Enterprise and Made in Italy). To be clear: the 99,000 euro figure refers to the entire tourism sector, while the 44,000 euro figure only covers lodging. Furthermore, there are two factors to consider: these are national averages, and Sardinian tourism is far more seasonal and accommodation-heavy than the rest of Italy. This means the local situation is almost certainly worse. We are not just less productive compared to other industries; we are less productive even than our direct competitors in the very field where we collect the most revenue.</p><p>Honestly, I would have preferred to read that Sardinia exports 1.8 billion euros in software and telecommunications. A billion in beds is not the same as a billion in software. Both represent real money, but they build radically different economic foundations beneath them.</p><h3>What the theory says (without being fans)<a href="https://maroffo.github.io/blog/posts/2026-06-15-a-billion-in-beds-or-a-billion-in-software/#what-the-theory-says-without-being-fans">#</a></h3><p>No trials for deck chairs: low productivity does not mean we should pack up and go home. We have specialized where we have an obvious competitive advantage (the sea, the climate, the nature), and it makes perfect sense. It would be absurd to ask Sardinia to apologize for the beauty of its coastline.</p><p>A strong tourism sector is not a problem in itself. A Bank of Italy study (Bronzini, Ciani, and Montaruli, <em>Regional Studies</em> 2022) calculated the impact of foreign tourist spending on the growth of per capita value added in Italian provinces. The effect is positive, tangible, and is felt most in regions starting from a disadvantage, like ours.</p><p>The data does not say that tourism is poison. It says something more useful: the boost is real, but it is limited and tends to wear off once congestion thresholds are crossed. The issue is not that the tourist’s euro is a “bad euro”, but that a monoculture economy built on it has a very low ceiling and a single point of failure. Easier revenues risk crowding out more complex, long-term economic activities. This is a subtle displacement mechanism (a highly plausible hypothesis, rather than a scientifically proven fact): short-term rentals that drive up housing costs for residents, rapid returns that discourage riskier and slower technological investments, and public resources continuously drained to promote the next season instead of funding research.</p><h3>Why it hurts more here<a href="https://maroffo.github.io/blog/posts/2026-06-15-a-billion-in-beds-or-a-billion-in-software/#why-it-hurts-more-here">#</a></h3><p>This structural weakness is felt more acutely in Sardinia than elsewhere because of our productive fabric. The 32nd CRENoS report shows that over 96% of our companies are micro-businesses, averaging 2.9 employees. These tiny enterprises absorb more than 60% of all jobs (compared to 38% in the Center-North). An economy composed almost entirely of three-person businesses lacks the financial and organizational capacity to make long-term bets. If we combine this with savage seasonality, where tourism flows are concentrated almost entirely in July and August, we are left with a workforce that is active for only a few months of the year and idle for the rest.</p><p>There is also a matter of economic positioning. It is true we have an extraordinary luxury segment, like the Costa Smeralda, which generates high value per worker. But this is an enclave of a few kilometers, whose profits often leave the island as soon as the summer lights fade, failing to lift the regional average. The bulk of our tourism is far more ordinary: homes, apartments, and short-term rentals, which is the fastest-growing segment (newly registered rentals rose by over 40% in a single year). The issue is not the format of the accommodation, but the absence of related services (such as dining, guides, structured experiences) that instead make the fortune of places like the Balearics or the French Riviera. Our beaches remain wild by choice. Sardinians like them that way, and so do I. But economically, it means selling an umbrella and two sunbeds for a few euros a day, whereas a competitor elsewhere sells a full day of services with lunch included. Excluding the golden enclave of luxury, the rest leans downward: few stable skills, extreme seasonality.</p><p>Optimists reply that this distributed model protects the territory: 44,000 families supplement their income by renting a second home, with no multinationals draining profits abroad. This is a valid and important point. That money stays within the local economy and directly supports family budgets. However, retaining a euro is very different from compounding it. A summer rental helps pay the mortgage, but it does not build skills, careers, or scalable assets. It is a rent that must be re-earned from scratch every year. Forty-three thousand small landlords will never create the conditions to pay an engineer’s salary, simply because that is not their job. Staying resilient is one thing; generating new wealth is another.</p><p>Structural data confirms this struggle. Sardinian GDP per capita is around 29,500 dollars (measured by the OECD) compared to a national average of 40,900. Spending on research and development is stalled at 182 euros per head, compared to 439 euros in Italy and 742 euros in Europe. We invest less than half the national average and a quarter of the European average in the single factor that actually drives productivity over time. Meanwhile, the people who could change things are leaving: Sardinia loses about 12 graduates per thousand inhabitants to emigration each year (the national average is below three), and Svimez counts ten thousand qualified young people aged 25 to 34 who have left the island recently.</p><p>The late 2025 Svimez report highlights the most alarming trend: in 2024, Sardinia’s economy grew by 0.9%, but only thanks to industry and construction. Services (the sector containing most tourism) shrank, in contrast to the rest of the Mezzogiorno. Software is classified under services too, so this is not a condemnation of tourism alone, but it shows that in a record-breaking season, tourism’s spillover effects failed to drive the real economy. The supply chain is so short that hotels and resorts often source goods and services outside the island, dispersing a large portion of the economic value generated.</p><p>The result is a low-value, highly seasonal monoculture, in a region that is losing its graduates and failing to invest in innovation. This is not a coastline to apologize for, but it is certainly a strategy to worry about.</p><h3>An alternative we already know<a href="https://maroffo.github.io/blog/posts/2026-06-15-a-billion-in-beds-or-a-billion-in-software/#an-alternative-we-already-know">#</a></h3><p>Yet there is a part of this story that belongs to us deeply.</p><p>Sardinia was the cradle of Italy’s consumer Internet. In 1990, the CRS4 research center was established near Cagliari; in March 1994, it helped put <em>L’Unione Sarda</em> online, making it one of the first digital newspapers in Europe. Soon after, Nicola Grauso launched Video On Line, which quickly became the country’s leading internet access provider. In 1998, also in Cagliari, Renato Soru founded Tiscali. For several years, this island was, without any exaggeration, the most innovative hub in Italy for tech development.</p><p>We know how to make things happen. The harsher lesson, if anything, is about what comes next. Video On Line was acquired by Telecom Italia, and its gravity shifted to Milan. Tiscali rode the dot-com bubble to extraordinary heights, only to downsize over the following decades. We produced outstanding pioneers, but we failed to build a lasting industrial district. This is the core issue: starting projects was never our bottleneck. Keeping them here was.</p><p>And yet, the ecosystem is far from dead. The Net Value has been a certified incubator active in Cagliari since 2009. Moneyfarm, now a prominent international player in financial services, has its roots and offices in the city. Abinsula, founded in Sassari by engineers who chose to return home, exports embedded and automotive software worldwide (with over 60% of its revenues generated abroad). Recently, CDP Venture Capital chose Cagliari for Frontech, an acceleration program focused on frontier technologies and artificial intelligence. The pieces are all on the table. What we lack is the capacity to connect them into a structure that can outlast its individual founders.</p><h3>How to intervene concretely<a href="https://maroffo.github.io/blog/posts/2026-06-15-a-billion-in-beds-or-a-billion-in-software/#how-to-intervene-concretely">#</a></h3><p>My proposal is not to scale back tourism. It would be foolish to shut down the island’s primary engine. The real question is how to channel a fraction of the value generated by this engine toward higher value-added sectors.</p><p>There is no idle public treasure waiting in regional accounts: the euro spent by a tourist goes directly to the hotelier, the local family renting their second home, or Ryanair. It is almost entirely private capital, and a substantial part leaves the island within days. Consequently, “reinvesting the surplus” is not a simple administrative directive. The actual levers are complex and deeply political: attracting large institutional investors (like CDP with Frontech) and planning land use to decide if the next coastal hectares should go to a resort or a technology park. These choices require political courage and long-term vision.</p><p>But even with capital available, the historical challenge remains: retaining that value. Tiscali did not suffer from a lack of funding, yet economic gravity eventually pulled the company toward large continental centers where talent and late-stage capital concentrate. A startup born in Cagliari that reaches a Series B funding round almost naturally finds itself with a headquarters in Milan or, more likely, London. Today, however, we have two new factors that change the dynamics. First is remote work: a twenty-five-year-old software engineer can work for a global enterprise and earn an international salary while living in Cagliari or Pula. This partially disarms the physical agglomeration trap. Second is the arrival of qualified external capital, such as CDP’s fund. The goal should no longer be to force a company to remain physically anchored to the island in defiance of market forces, but to ensure that the skills, the salaries, and a portion of the intellectual and corporate ownership remain here, even if the legal headquarters eventually relocates.</p><h3>Value that compounds, value that passes<a href="https://maroffo.github.io/blog/posts/2026-06-15-a-billion-in-beds-or-a-billion-in-software/#value-that-compounds-value-that-passes">#</a></h3><p>The post I started from concludes with an appeal: recognize tourism as our primary export and treat it accordingly in development policies. I share the premise (ignoring it is snobbery), but I reject the conclusion. I do not want tourism to be the horizon around which we design the future of this island. Without touching a single beach stand, I want us to build an economy with better prospects and wages alongside it, letting it grow until it becomes the most significant figure in our ledger.</p><p>There is a solid objection to this view, which is the strong argument of the original post: tourism cannot be outsourced (a beach does not move to Romania). Software, by contrast, tends by its nature to migrate, as shown by the stories of Video On Line and Tiscali. Why then chase a volatile sector and neglect a resource anchored to our coastlines? The answer is that it is not a mutually exclusive choice. A non-transferable asset with low value-added remains a structural cap on wages. The goal is not to bet everything on software hoping it defeats tourism, but to reject the idea that a sunbed on a beach is the peak of our collective ambition. Even a partial victory (a few solid companies staying, skills consolidating, professionals working for international clients from home) raises the overall wage floor, which has been stagnant for thirty years.</p><p>There is also an exceptional card we are not playing: attracting external talent. Combining an affordable cost of living with an extraordinary quality of life is a rare luxury in today’s world. Beyond encouraging the growth of local startups, we can position the island as a global hub for professionals and digital nomads. Cagliari already has an international dimension, and those who visit tend to fall in love with it. The basic infrastructure already exists, it simply needs to be leveraged. I am thinking of Sardegna Ricerche and the technology park in Pula that hosts CRS4: a campus immersed in greenery just steps from the sea. Few places in the world can offer such a work environment. It is an enormous competitive advantage that we manage like a tourist postcard rather than an industrial resource.</p><p>Mine is a personal preference, not a verdict. I prefer a thousand local startups, even if we must accept that only a very few will make it, over ten thousand new holiday homes. Not for the immediate economic return in summer (where second homes would win outright), but for the possibility of having stable work in winter, real career paths that do not require boarding a plane, and an economy less vulnerable to the weather and seasonal trends.</p><p>It is the same concept applied to Europe and technological sovereignty that I wrote about in <a href="https://maroffo.github.io/blog/posts/2026-06-13-own-the-whole-stack-down-to-the-silicon/">Own the Whole Stack, Down to the Silicon</a>, transposed to a regional scale. There, the fear was depending on other people’s technologies that can be deactivated with a click. Here, the issue is even deeper: a hotel room or a villa must be rented out again every season, starting from scratch each time. A solid technology company, however, compounds on itself, generates local economic spin-offs, and continues to produce wealth and skilled jobs long after the founders have moved on. It is not a conflict between value that stays and value that leaves, but between value that compounds and value that must be constantly chased.</p><p>The current gap is immense. Comparing a Sassari software house with ten million in revenue to a 1.85 billion tourism sector almost makes you smile. Reversing this ratio is the commitment of a generation, not a single financial maneuver, and we might not succeed. Yet the Sardinian tech ecosystem is showing unprecedented signs of life. We have already proven we can do great things from nothing: the Italian Internet was born here. Today, the challenge is less romantic but more crucial: building those silent infrastructures (stable capital, support services, real reasons to stay) that allow the next generation of innovators to grow and succeed at home.</p><h3>Sources and further reading<a href="https://maroffo.github.io/blog/posts/2026-06-15-a-billion-in-beds-or-a-billion-in-software/#sources-and-further-reading">#</a></h3><ol><li>Bank of Italy, foreign tourism spending statistics (international tourism, balance of payments) — <a href="https://www.bancaditalia.it/statistiche/tematiche/rapporti-estero/turismo-internazionale/index.html">bancaditalia.it</a>; Sardinia 2024 figure of ~1.85 billion euros via regional reporting</li><li>ISTAT, <a href="https://www.istat.it/en/press-release/tourism-satellite-account-for-italy-year-2023/">“Conto satellite del turismo per l’Italia — Anno 2023”</a> — labour productivity 99,015 euros (−30% vs national), per-capita income 17,709 euros (−35%)</li><li>Value added per worker in Italian accommodation (~44,000 euros vs ~82,000 for competitors), Ministry of Enterprise and Made in Italy, via <a href="https://www.partitaiva.it/editoriali/turismo-italia-2026-record-pil/">partitaiva.it</a></li><li>Mario Holzner, <a href="https://www.sciencedirect.com/science/article/abs/pii/S0261517710001652">“Tourism and Economic Development: The Beach Disease?”</a> — Tourism Management, 2011; and Chao et al., tourism Dutch-disease model, 2006</li><li>R. Bronzini, E. Ciani, F. Montaruli, <a href="https://www.tandfonline.com/doi/abs/10.1080/00343404.2021.1910649">“Tourism and local growth in Italy”</a> — Regional Studies, vol. 56, no. 1 (2022), pp. 140–154; originally Bank of Italy QEF №509 (2019)</li><li>CRENoS, 32nd Report on the Economy of Sardinia (2025) — micro-firms over 96%, 2.9 employees average, micro absorbing over 60% of jobs vs 38% in the Centre-North — <a href="https://crenos.unica.it/">crenos.unica.it</a></li><li>R&amp;D 182 euros per capita vs 439 (Italy) and 742 (EU), GDP per capita $29,504 vs $40,898, graduate emigration — CRENoS / OECD 2025, via <a href="https://www.cagliaritoday.it/economia/economia-sardegna-ocse-2025-sottosviluppo-.html">Cagliari Today</a></li><li>Svimez, 2025 Report — Sardinia +0.9% in 2024 driven by industry and construction, services in decline, ~10,000 aged 25–34 emigrated — <a href="https://www.svimez.it/rapporto-svimez-2025/">svimez.it</a> and <a href="https://www.cagliaritoday.it/economia/economia-sardegna-rapporto-svimez-2025.html">Cagliari Today</a></li><li>CRS4, Video On Line, and Tiscali history — the Italian Internet’s Sardinian origins, via <a href="https://www.ansa.it/sardegna/notizie/2022/05/19/la-sardegna-patria-per-start-up-una-storia-lunga-40-anni_f4599fe4-8be4-46ca-b8d2-8b0b7a228e02.html">ANSA</a> and public records</li><li>The Net Value (<a href="https://thenetvalue.com/incubatore/">thenetvalue.com</a>), Abinsula (<a href="https://abinsula.com/">abinsula.com</a>), and CDP Venture Capital’s <a href="https://www.cdpventurecapital.it/cdp-venture-capital/en/dettaglio_acceleratore.page?contentId=ACL3494">Frontech</a> accelerator in Cagliari</li></ol><p><strong><em>Methodology note:</em></strong><em> This piece started from an analysis of Sardinia’s economy that has been circulating, and I wanted to think through it on my own terms rather than just react to it. I wrote it with AI assistance (Claude Code and Gemini) and checked every figure against current sources: that is how I realized, for example, that the ISTAT productivity numbers are national tourism-sector figures rather than Sardinian ones, and I have reported them honestly above. The displacement I speak of is a plausible hypothesis, not a proven fact, and I have stated so. The opinion, for better or worse, is mine. If you find a wrong number, let me know and I will correct it.</em></p><p><em>Originally published at </em><a href="https://maroffo.github.io/blog/posts/2026-06-15-a-billion-in-beds-or-a-billion-in-software/"><em>https://maroffo.github.io</em></a><em> on June 15, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=be7d166f4f3e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Own the Whole Stack, Down to the Silicon]]></title>
            <link>https://maroffo.medium.com/own-the-whole-stack-down-to-the-silicon-73d417a40182?source=rss-c6cd0b9646f5------2</link>
            <guid isPermaLink="false">https://medium.com/p/73d417a40182</guid>
            <dc:creator><![CDATA[Massimiliano Aroffo]]></dc:creator>
            <pubDate>Sat, 13 Jun 2026 00:00:05 GMT</pubDate>
            <atom:updated>2026-06-18T14:24:56.058Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*1IXN6LeYCuryiafk.png" /></figure><p>On the evening of June 12, the US government sent Anthropic a letter. By the time most of Europe had read about it, Fable 5 and Mythos 5 were already gone.</p><p>The directive cited national security and ordered Anthropic to cut off both models for every foreign national, inside the United States and out, its own employees included. Anthropic complied within hours. It also said, in plain language, that it thinks the call was a mistake.</p><p>Here is the detail that makes it real. The “jailbreak” the government is worried about, by Anthropic’s own account, comes down to asking the model to read a specific codebase and fix the flaws it finds. That is not an exotic weapon. It is what defenders do every day, and the same capability ships in other public models, GPT-5.5 among them.</p><p>I work on AI security, so I read this twice.</p><h3>Two ways to read one letter<a href="https://maroffo.github.io/blog/posts/2026-06-13-own-the-whole-stack-down-to-the-silicon/#two-ways-to-read-one-letter">#</a></h3><p>The first reading is professional. There is no such thing as perfect jailbreak resistance, not for Anthropic and not for anyone else. You cannot prove a model will never be talked into something it should refuse, and anyone who tells you otherwise is selling something. The serious response is the same as it is for networks: defense in depth. Narrow what a bypass can reach, make it expensive to build, watch the traffic, keep the logs that let you reconstruct an attack, and shut real ones down fast. On that reading, recalling a model used by hundreds of millions of people over one narrow, non-universal bypass is hard to justify on the technical facts. That is more or less Anthropic’s position, and I find it convincing.</p><p>But the security argument is not why I have been distracted since the news broke. The second reading is.</p><p>For those of us building in Europe, the jailbreak debate is almost a luxury. We were watching something simpler and more uncomfortable: one government, with one letter, switching off two of the most capable tools on the planet overnight, with no warning and no vote for anyone outside its borders. The jailbreak is the pretext. The dependency is the lesson.</p><h3>Three problems I cannot put down<a href="https://maroffo.github.io/blog/posts/2026-06-13-own-the-whole-stack-down-to-the-silicon/#three-problems-i-cannot-put-down">#</a></h3><p>I spent this morning talking it through with Mauro Medda, who founded HikmaAI, where I work. Local, decentralized inference is something he has pushed for long before any of this, so the directive landed in the middle of a conversation we keep having anyway. Three problems came out of it, and I have not managed to set any of them down since.</p><p><strong>The first is timing.</strong> American companies get the most capable models first, sometimes months before the rest of us. In a market that moves this fast, a few months of better tools is not a rounding error. It is a real lead, and it compounds the whole time. We are not starting level and slipping behind. We are starting behind.</p><p><strong>The second is the kill switch.</strong> Mauro put it more bluntly than I would have: cut the access, and you can send a modern company back to the stone age in an afternoon. He is not exaggerating by much. Capabilities that took years to build into a product, or into the public services and infrastructure a country runs on, can be turned off by a decision made somewhere you have no say. Yesterday that stopped being a thought experiment. We watched it happen to Anthropic’s own customers, and one of them was us. In the first days of June, ENISA, the European Union’s own cybersecurity agency, became the first EU institution allowed into Anthropic’s Project Glasswing, using Mythos to hunt zero-day flaws in critical infrastructure. On June 12 that access went dark with everyone else’s. Europe was defending its own networks with a switch that sits in Washington.</p><p><strong>The third one caught me off guard, because I had quietly underestimated it.</strong> Say Europe gets serious. Say we build our own data and train our own models. They would still run on chips designed in the United States, mostly by Nvidia. Nvidia sells to us today, but a sale is just a tap that happens to be open. A policy shift in Washington could narrow it, or move American buyers to the front of the line. That is not speculation either. There is already a provision moving through Congress, the GAIN AI Act, that the Senate passed inside this year’s defense bill, and it would give American buyers a right of first refusal on the best chips before any foreign order ships. Not law yet, but the intent is on the record. So even a fully European model would sit on hardware we neither make nor control the queue for.</p><p>Put the three together and the conclusion is uncomfortable but simple. Owning the data is not enough, and neither is renting the models or self-hosting someone else’s open weights. If we are serious about this, we own the whole stack, down to the silicon.</p><p>I know how that sounds. A right of first refusal on chips is protectionism when Washington does it, and somehow industrial strategy when I want Europe to do the same. Fair enough. The difference I actually care about is not who holds the switch, it is whether it sits on my desk or on someone else’s.</p><h3>The part that actually gives me hope<a href="https://maroffo.github.io/blog/posts/2026-06-13-own-the-whole-stack-down-to-the-silicon/#the-part-that-actually-gives-me-hope">#</a></h3><p>I want to be careful here, because “European tech sovereignty” is a phrase that has launched a thousand press releases and very little working software. So let me stay concrete.</p><p>At AI Week in Milan on May 19, HikmaAI shared a stage with Andrea Pili and the team at Xference. I have skin in this, so discount what follows accordingly. But the talk was not a vision deck. It was about something you can point at: sovereign and secure inference, built on products from Italian companies.</p><p>Those two words only work together. The sovereign half is private, local inference on European infrastructure, the kind where your data never leaves your perimeter, GDPR and AI Act compliant because that is how it was built, not because a lawyer bolted it on at the end. That part is Xference. The secure half is the one I work on: bringing a model in-house only helps if you can also defend it, and the Fable case is the cleanest argument I have for why. Owning a model is worth very little if you cannot trust how it behaves under pressure. Sovereign and insecure is just a more expensive way to get breached.</p><p>That covers two layers of the stack. The third, the silicon, is the one everyone waves away as hopeless. I am not so sure.</p><p>The hardware layer is not a closed door for Europe. It is a door we keep pretending is locked. ARM, the architecture behind most of the world’s efficient chips, was born in Cambridge (yes, post-Brexit Britain) and is still designed there. It is owned by SoftBank now, and that caveat is real, so I will not call it European and have someone correct me in the comments. But the design knowledge, and the people, are here. Apple’s M-series runs on that same architecture, and it has already shown you can put serious AI workloads, capable open models included, on a laptop instead of a hall full of Nvidia parts. Apple is American, so that proves the architecture, not European independence; the architecture is the part that was born here. The architecture is not the only card, either. Axelera AI, out of Eindhoven, designs AI inference chips on a RISC-V base, part-funded by an EU program named, without irony, Digital Autonomy with RISC-V for Europe. The sovereignty case is starting to get made in silicon, not just in slides. And the machines that print the most advanced chips on Earth, Nvidia’s very much included, come from one company: ASML, in the Netherlands. Nobody makes a leading-edge chip without them.</p><p>None of that is a finished alternative. ARM answers to Tokyo, the leading fabs are in Taiwan, and a laptop is not a training cluster. But the pieces are sitting right there, and we keep behaving as if they are not.</p><h3>What this is going to cost<a href="https://maroffo.github.io/blog/posts/2026-06-13-own-the-whole-stack-down-to-the-silicon/#what-this-is-going-to-cost">#</a></h3><p>I am not going to pretend the scale gap is small. Mistral, our most credible frontier lab, has raised a few billion dollars across its whole life. OpenAI and Anthropic are each valued near a trillion, and Anthropic’s single funding round this May was bigger than everything Mistral has ever raised. European AI gets a small slice of global investment, and too many of our best researchers still board planes for California. Those are the real numbers, and they are not flattering.</p><p>Some of that gap is self-inflicted, and it would be dishonest to skip it. The same AI Act that lets Xference say “compliant by design” is also, fairly or not, blamed for slowing European builders down. Both can be true at once. Sovereignty you regulate yourself out of ever reaching is not sovereignty either.</p><p>Matteo Flora made the sharper version of this point, and it changed how I read the whole episode. What Washington used on June 12 was not an AI law, it was an export control, the same instrument the US once aimed at cryptography. A written rule you can read, predict, and take to a judge. A switch flipped at 5:21 in the afternoon you cannot. The freedom of a tech ecosystem is not measured by how many rules it carries, but by whether the power over it is visible and contestable or invisible and absolute. By that test the AI Act, slow and public and easy to resent, looks less like the problem and more like the only kind of power you ever get to argue with.</p><p>So I am not selling optimism, and the economics are not on my side either. For most companies today, renting the best American model is the rational call: cheaper and faster and sitting right there behind an API. I am not arguing about that default. I am arguing about the floor, the handful of capabilities you cannot afford to have switched off from somewhere you do not get to vote. The inference layer is moving, and moving fast, and the work HikmaAI and Xference showed in Milan is a small, concrete piece of that. The chips are the long game, and that one is not a startup’s job to solve. It is an industrial choice, the kind only governments and very patient capital can make, and Europe has spent a decade not quite making it.</p><p>The money was even there. NextGenerationEU was the largest recovery program Europe has aimed at itself in generations, and Italy’s share of it, the PNRR, was the biggest slice of the lot. If we had found the nerve to put even half of that into technological infrastructure and research, we might be a little less vulnerable today. Instead, at least in Italy, we built roundabouts.</p><p>What changed yesterday is that the cost of not deciding got a face. For years, “strategic dependency” was a phrase for panel discussions. On June 12 it became a letter, and an afternoon, and a product that worked in the morning and did not by dinner.</p><p>People far closer to this than I am, Mistral’s CEO among them, keep saying the window to fix this is short, measured in a few years and not decades. I think they are right, and the honest move is to act like we believe them.</p><p>There are two ways to pay for this. One is now, in money and effort and decisions nobody enjoys making. The other is later, in a letter we never got the chance to read.</p><p>Mauro Medda, the colleague I had this argument with, wrote up his own version the same day: <a href="https://mauro.medda.xyz/posts/own-the-stack-or-rent-the-future/">Own the Stack, or Rent the Future</a>. He lays out the strategy: what Europe should actually build, and how. As usual, his vision runs a step ahead of the rest of us. Read it.</p><h3>Sources and further reading<a href="https://maroffo.github.io/blog/posts/2026-06-13-own-the-whole-stack-down-to-the-silicon/#sources-and-further-reading">#</a></h3><ol><li>Anthropic, <a href="https://www.anthropic.com/news/fable-mythos-access">“Statement on the US government directive to suspend access to Fable 5 and Mythos 5”</a> — June 12, 2026</li><li>Fortune, <a href="https://fortune.com/2026/06/13/anthropic-disables-fable-mythos-export-controls-national-security-threat/">“Anthropic disables Fable and Mythos AI models following U.S. government export ban”</a> — June 13, 2026</li><li>AI Magazine, <a href="https://aimagazine.com/news/what-is-the-gain-ai-act-and-why-does-nvidia-oppose-it">“What is the GAIN AI Act and Why Does Nvidia Oppose it?”</a>, and the <a href="https://www.banking.senate.gov/newsroom/minority/banks-warren-cotton-schumer-mccormick-coons-introduce-landmark-bipartisan-gain-ai-act-to-maintain-us-position-as-worlds-leader-in-critical-artificial-intelligence-chips">US Senate Banking Committee release</a> on the GAIN AI Act</li><li><a href="https://www.intelligentcio.com/eu/2026/04/28/xference-launches-european-private-ai-infrastructure-from-aruba-data-centres/">“Xference launches European private AI infrastructure from Aruba data centres”</a> — Intelligent CIO Europe, April 2026</li><li><a href="https://techcrunch.com/2026/06/12/mistral-is-rumored-to-be-raising-e3b-at-e20-valuation/">“Mistral is rumored to be raising €3B at €20B valuation”</a> — TechCrunch, June 2026</li><li>ARM history and SoftBank ownership, and ASML’s EUV lithography monopoly — public filings and reporting</li><li>Matteo Flora, <a href="https://mgpf.it/2026/06/13/mithos-bloccato.html">“MITHOS BLOCCATO: 17.21 è l’ora in cui gli USA hanno bloccato l’Intelligenza Artificiale”</a> — mgpf.it, June 13, 2026</li><li><a href="https://techcrunch.com/2026/06/02/anthropic-scales-claude-mythos-to-critical-infrastructure-in-15-countries/">“Anthropic scales Claude Mythos to critical infrastructure in 15+ countries”</a> — TechCrunch, June 2026, and <a href="https://www.bloomberg.com/news/articles/2026-06-01/anthropic-to-give-eu-s-cybersecurity-agency-access-to-mythos">“Anthropic to give the EU’s cybersecurity agency access to Mythos”</a> — Bloomberg, June 2026</li><li>HikmaAI — <a href="https://hikmaai.io/">hikmaai.io</a></li><li><a href="https://www.datacenterdynamics.com/en/news/axelera-ai-launches-inferencing-chiplet-supported-by-funding-from-eus-risc-v-project/">“Axelera AI launches inferencing chiplet, supported by funding from the EU’s RISC-V project”</a> — Data Center Dynamics, March 2025</li><li>Mauro Medda, <a href="https://mauro.medda.xyz/posts/own-the-stack-or-rent-the-future/">“Own the Stack, or Rent the Future”</a> — mauro.medda.xyz, June 2026</li></ol><p><strong><em>Methodology note:</em></strong><em> This one started as a long chat with Claude about a LinkedIn post and turned into something I wanted more room for. I wrote it with AI assistance (Claude Code), checked every factual claim against current sources (the funding figures and the GAIN AI Act status both moved while I was writing), and ran the prose through my own humanizer pipeline to strip the AI tells. The opinions are mine, and so is the call to quote Mauro by name. If you catch an AI-ism I missed, I owe you a coffee.</em></p><p><em>Originally published at </em><a href="https://maroffo.github.io/blog/posts/2026-06-13-own-the-whole-stack-down-to-the-silicon/"><em>https://maroffo.github.io</em></a><em> on June 13, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=73d417a40182" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Word ‘Permanent’ Isn’t in the Paper]]></title>
            <link>https://maroffo.medium.com/the-word-permanent-isn-t-in-the-paper-1ef084080197?source=rss-c6cd0b9646f5------2</link>
            <guid isPermaLink="false">https://medium.com/p/1ef084080197</guid>
            <dc:creator><![CDATA[Massimiliano Aroffo]]></dc:creator>
            <pubDate>Thu, 07 May 2026 00:00:23 GMT</pubDate>
            <atom:updated>2026-06-18T14:23:55.278Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*-rqlqM-FBE1EymIg.png" /></figure><p>A tweet showed up in my feed this week. It was a repost by Cher Scarlett of a post by someone called Elara Grace, with the kind of opener that tells you exactly what you’re about to read:</p><blockquote><em>Just IN: If you’ve used ChatGPT for writing or brainstorming in the last 6 months, your creative ability may already be permanently damaged. A controlled experiment just proved the effect doesn’t reverse when you stop using it. 3,302 creative ideas. 61 people. 30 days of tracking.</em></blockquote><p>The thread went on for a while, with the kicker: “You’re not renting a productivity boost. You’re financing it with your originality. The interest rate is permanent.”</p><p>My first reaction was the one-word kind. The second reaction, after I noticed the specific numbers, was that the tweet was citing real studies. Dismissing it as bullshit isn’t enough then. The numbers exist. The papers exist. The question worth asking is what got dropped between the abstracts and the post.</p><h3>What the first paper actually says<a href="https://maroffo.github.io/blog/posts/2026-05-07-the-word-permanent-isnt-in-the-paper/#what-the-first-paper-actually-says">#</a></h3><p>The 3,302 ideas and 61 people come from Liu, Zhou, Huang, and Li, <em>When ChatGPT is gone: Creativity reverts and homogeneity persists</em>. It’s on arXiv as 2401.06816, with an extended version in <em>Technology in Society</em> in 2025. It’s a real paper. The numbers are real. Here is the abstract verbatim, because the verbatim version is where the rhetorical move starts:</p><blockquote><em>[&amp;mldr;] human creative performance reverted to baseline when ChatGPT was down on the 7th and the 30th day. More critically, the use of ChatGPT in creative tasks resulted in increasingly homogenized contents, and this homogenization effect persisted even when ChatGPT was absence.</em></blockquote><p>Three things are worth pointing out, and the tweet hides all three.</p><p><strong>One: “reverts to baseline” is not “drops below baseline.”</strong> The abstract is precise. Performance returns to where it was before. The tweet implies a worsening that the paper does not document. “Permanently damaged” is a statement about creative capability falling below pre-treatment level. The paper measures the opposite: that the <em>boost</em> from ChatGPT does not persist when ChatGPT goes away. Subtract the boost, you get the original creative performance back. That is not damage. That is the boost being a rental rather than a purchase. The tweet’s framing inverts the measurement.</p><p><strong>Two: the word “permanent” does not appear in the abstract, and the maximum follow-up is 30 days.</strong> I checked. The closest the paper comes to a long-run claim is the phrase “boxes human creative capability in the long run”, which is the authors’ interpretation of the homogenization persistence, not of any measurement past 30 days. There is no follow-up at six months, or twelve, or thirty. There is no sample large enough or design longitudinal enough to support the word “permanent”. It is a word the tweet supplies. The paper does not.</p><p>I should be fair: the authors themselves stretch their phrasing past their measurements. “Boxes human creative capability in the long run” is interpretation, not data. The tweet didn’t invent the apocalyptic register, it amplified one already there. That strengthens, rather than weakens, the point: 30 days of follow-up cannot license the word “permanent”, whoever uses it.</p><p><strong>Three: the design.</strong> Sixty-one college students at a single university, a seven-day lab experiment, a 30-day follow-up survey. The “dependency” the tweet describes is built in seven days of a lab task. This isn’t a flaw of the study; lab experiments are how you isolate variables. It is a flaw to take a seven-day lab effect and frame it as a description of what is happening to anyone who has used ChatGPT in the last six months. The generalization is the tweet’s, not the paper’s.</p><p>Now, here is where I have to be careful, because there is one finding in the paper that is real and worth taking seriously, and I do not want to talk past it: <strong>homogenization persists</strong>. After people stop using ChatGPT, the <em>content</em> they produce remains more similar to the content other ChatGPT users produce than the control group’s content was. The boost reverts; the homogenization sticks. That is a finding I respect. It says that exposure to a generative model trains a kind of stylistic and ideational gravity that doesn’t fully release.</p><p>It is not the same finding as “your creative ability is permanently damaged.” It is a finding about the <em>distribution</em> of ideas across people who used the tool, not about the individual’s <em>capacity</em>. A homogenized population can still contain people whose individual creative output is unchanged or improved. The paper measures convergence between subjects, not loss within them.</p><p>If I were writing the tweet honestly, the sentence would be: “After a week of ChatGPT use, people’s creative output becomes more similar to other ChatGPT users’ output, and this convergence persists for at least 30 days after they stop using it.” That would be accurate. It would also generate roughly zero retweets.</p><h3>What the second paper actually says<a href="https://maroffo.github.io/blog/posts/2026-05-07-the-word-permanent-isnt-in-the-paper/#what-the-second-paper-actually-says">#</a></h3><p>The “doesn’t reverse when you stop using it” framing comes from the second study the tweet folds in: Barcaui (2025), <em>ChatGPT as a cognitive crutch: Evidence from a randomized controlled trial on knowledge retention</em>, in <em>Social Sciences &amp; Humanities Open</em>. RCT, n = 120 enrolled, n = 85 at the surprise retention test 45 days later. Cohen’s d = 0.68. The ChatGPT group scored 57.5% on the retention test, the traditional-study group scored 68.5%. t(83) = -3.19, p = .002.</p><p>This is also a real paper. The numbers are also real. The misuse is more subtle.</p><p>The Barcaui paper does not measure creativity. It measures <strong>conceptual retention</strong>: how much of what you studied is still in your head 45 days later. The tweet treats Barcaui’s retention finding as evidence of a creativity finding from Liu et al. They are different measures, on different populations, with different mechanisms. Cognitive retention can decline through reduced effortful encoding (the “desirable difficulties” account from Bjork) without any implication for creative ideation. Mixing them lets the tweet extend Liu’s 30-day window with Barcaui’s 45-day window and call it “doesn’t reverse.” But the two are not measuring the same thing reverting or not reverting.</p><p>Generalizability matters here too. Barcaui’s sample is business undergraduates at a Brazilian university, studying AI concepts as the subject matter. The d = 0.68 is on that population, learning that material, with that style of unrestricted ChatGPT use during study. Whether the effect transfers to other materials, populations, and styles of AI integration is an open question the paper does not claim to answer.</p><p>None of this discredits the study. It’s a clean RCT and the effect size is meaningful. It just doesn’t license the move from “students who used ChatGPT to study an AI course remembered 11 percentage points less six weeks later” to “your creative ability is permanently damaged”.</p><h3>The rhetorical move<a href="https://maroffo.github.io/blog/posts/2026-05-07-the-word-permanent-isnt-in-the-paper/#the-rhetorical-move">#</a></h3><p>What the tweet does is the kind of move I keep running into in AI discourse, on both sides:</p><ol><li>Take real numbers from Study A.</li><li>Take a different, plausible-sounding finding from Study B.</li><li>Glue them together with a connecting word that neither paper used.</li><li>Frame the synthesis as “controlled experiment proved X.”</li></ol><p>Liu et al.’s 30-day window (creativity reverts to baseline; homogeneity persists) gets the word “permanent” attached to it by transitively citing Barcaui’s 45-day window (retention of unrelated material). The synthesis becomes “permanent damage” even though neither paper measures permanence and the two papers measure different things.</p><p>You can see the seams if you read the tweet carefully. “A controlled experiment just proved the effect doesn’t reverse when you stop using it.” Singular: a controlled experiment. The 3,302 ideas and 61 people are from Liu et al. But their 30-day measurement is the <em>reverting</em> part, not the <em>not reverting</em> part. The “doesn’t reverse” framing has to be borrowed from somewhere else, and that somewhere else is the second study. Two studies, presented as one experiment. That’s the move.</p><h3>The mechanism is real. The framing isn’t.<a href="https://maroffo.github.io/blog/posts/2026-05-07-the-word-permanent-isnt-in-the-paper/#the-mechanism-is-real-the-framing-isnt">#</a></h3><p>I want to be specific about what I’m not claiming, because the easy mistake here is to overcorrect into AI apologia.</p><p>The underlying mechanism the tweet gestures at is documented. Bastani and colleagues (Wharton / Turkey, 2024, n ≈ 1,000 high schoolers) gave students access to two GPT variants while solving math problems. The “GPT Base” group scored 48% higher <em>with</em> AI access. When access was taken away for the exam, they scored 17% <em>worse</em> than the control group. Same students, opposite directions, same intervention. The kicker: a “GPT Tutor” variant designed with scaffolding to keep the student doing the cognitive work mitigated the loss almost entirely. Bastani is the cleanest demonstration I know that the <em>design of the AI interaction</em> is doing the work, not the presence or absence of AI.</p><p>Lee and colleagues (Microsoft Research / CMU, CHI ’25, n = 319 knowledge workers) found the cognitive effort doesn’t disappear, it shifts, with confidence in the AI predicting reduced critical evaluation. That tracks with what Bjork &amp; Bjork have spent decades documenting under the heading “desirable difficulties”: the effortful processes that produce durable learning are the ones you don’t feel like doing in the moment. Skip them and you trade short-term ease for long-term retention. None of this is novel. Sparrow’s Google-effect work goes back to 2011.</p><p>I’ve written about this twice already, in <a href="https://maroffo.github.io/blog/posts/2026-03-06-ai-didnt-reduce-my-cognitive-load-it-moved-it/"><em>AI Didn’t Reduce My Cognitive Load. It Moved It.</em></a> and <a href="https://maroffo.github.io/blog/posts/2026-03-23-from-cognitive-load-to-cognitive-surrender/"><em>From Cognitive Load to Cognitive Surrender</em></a>. The thesis I keep coming back to, and that I think the evidence supports, is that workflow is the variable that moves the outcome, not the tool. So this isn’t a contrarian post about how AI is fine actually. It’s a post about a tweet getting the framing wrong on top of papers I take seriously.</p><h3>Steel-manning the tweet<a href="https://maroffo.github.io/blog/posts/2026-05-07-the-word-permanent-isnt-in-the-paper/#steel-manning-the-tweet">#</a></h3><p>A fair reading of all this is that the tweet picked the wrong word, but the direction it points in has support. Homogenization in Liu et al. persists past 30 days. Retention drops in Barcaui at 45 days. Bastani’s GPT Base group lost ground once scaffolding was removed. The Lancet endoscopist study I cited in the cognitive-surrender post showed measurable deskilling after three months of AI-assisted polyp detection. None of these prove permanence. A reader sympathetic to the original tweet can fairly say the trend line points where the author said it points.</p><p>My objection is to the specific word, not to the direction. The reason the specific word matters is that “permanent” is what makes the claim viral, and “permanent” is the part nobody measured. Strip that word out and you keep the direction the evidence actually supports, plus most of the alarm the situation actually merits. What you lose is the rhetorical fuel that gets the post to a million views. That’s a feature of careful claims, not a bug.</p><h3>The other direction makes the same mistake<a href="https://maroffo.github.io/blog/posts/2026-05-07-the-word-permanent-isnt-in-the-paper/#the-other-direction-makes-the-same-mistake">#</a></h3><p>I owe the symmetry. The tweet’s framing is bad, but the opposite framing (AI gives you a free productivity boost) makes the same kind of move in reverse, and the recent retraction record proves it.</p><p>The most cited meta-analysis on the positive side, Wang and Fan (2025, <em>Humanities and Social Sciences Communications</em>), reported g = 0.867 across 51 studies, a “large positive impact” on learning performance. It accumulated 486,000 views, 266 citations, and an Altmetric score over 1,000 before <em>Nature</em> <strong>retracted it on April 22, 2026</strong>, two weeks ago at the time I’m writing this. The retraction note cites discrepancies that “ultimately undermined the confidence the Editor could place in the validity of the analysis and resulting conclusions.” Critics pointed to no quality assessment of included studies, no reporting of whether they were peer-reviewed or randomized, and 33 of 51 studies with fewer than 35 students in the experimental group.</p><p>So if you wanted to respond to the apocalyptic tweet by citing the big positive meta-analysis, that meta-analysis is no longer available to cite. The euphoric pole and the apocalyptic pole both made the same kind of move: take effects measured in narrow conditions, smooth over the methodological caveats, present the synthesis as a clean conclusion. Engagement on X likes clean conclusions.</p><p>There is a less-cited but methodologically tighter meta-analysis (35 studies, 4,193 participants, g = 0.670) finding a moderate positive effect on learning outcomes, with subject, duration, and instructional mode as significant moderators. It’s the kind of finding that doesn’t compress well into a 280-character claim. Which is the actual problem.</p><h3>The right question, briefly<a href="https://maroffo.github.io/blog/posts/2026-05-07-the-word-permanent-isnt-in-the-paper/#the-right-question-briefly">#</a></h3><p>I don’t want to redo a thesis I’ve laid out in earlier posts, so this is a one-paragraph version. The question “is ChatGPT bad for learning, or for creativity?” is malformed. The empirically tractable question is which combinations of task, AI interaction design, and process phase produce which outcome. Bastani’s GPT Base vs GPT Tutor split is the cleanest concrete instance: same students, same model, same problem set, opposite outcomes, depending on how the interaction was scaffolded. The tool plus the way the interaction is set up is what’s doing the work. Strip out the design variable, and both polar narratives become incoherent.</p><p>This is also why both polar narratives feel so unsatisfying when you’re actually using the thing. They erase the variable that matters.</p><p>The tweet’s central claim is that ChatGPT homogenizes its users’ output: takes the variance in human ideas and compresses it toward the mean. Liu et al. actually document this, in a narrow sense, on a small sample. It’s the part of the paper I take most seriously.</p><p>The tweet then takes two different studies, with different methods, populations, and outcome variables, and compresses them into a single clean apocalyptic claim suitable for X engagement. It produces a homogenized, high-fluency, low-information artifact aimed at the median reader’s appetite for a confident verdict.</p><p>The mechanism the tweet attributes to AI is the mechanism the tweet itself uses. I don’t think this is a coincidence. Both LLMs and engagement-optimized social media are systems that select for fluent, confident, conclusion-shaped outputs. They smooth out qualifications, they drop conditions, they keep the punchline. Reading the original papers is the part that doesn’t compress. It’s also the part that takes effort. Bjork has a name for that.</p><p>If a tweet says a controlled experiment proved your brain is permanently damaged, the move is to find the paper. Open the abstract. Look for the word “permanent.” If it isn’t there, the apocalypse is the tweet’s, not the science’s.</p><h3>Methodology note<a href="https://maroffo.github.io/blog/posts/2026-05-07-the-word-permanent-isnt-in-the-paper/#methodology-note">#</a></h3><p>This article was written with AI assistance (Claude Code for research synthesis, paper retrieval, and drafting; manual editing throughout). Verifying the abstracts directly, and cross-checking which words appear in which paper, was the part I would not delegate. The irony of using an AI to write a piece arguing against an AI-discourse pattern is not lost on me. It’s also exactly what the post is about: the tool isn’t doing the work. The way the interaction is scaffolded is.</p><ul><li>Liu, Q., Zhou, Y., Huang, J., &amp; Li, G. (2024). <a href="https://arxiv.org/abs/2401.06816">When ChatGPT is gone: Creativity reverts and homogeneity persists</a>. arXiv:2401.06816.</li><li>Barcaui, A. (2025). <a href="https://www.sciencedirect.com/science/article/pii/S2590291125010186">ChatGPT as a cognitive crutch: Evidence from a randomized controlled trial on knowledge retention</a>. <em>Social Sciences &amp; Humanities Open</em>, 12, 102287.</li><li>Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., &amp; Mariman, R. (2025). <a href="https://www.pnas.org/doi/10.1073/pnas.2422633122">Generative AI without guardrails can harm learning: Evidence from high school mathematics</a>. <em>PNAS</em>, 122.</li><li>Wang, J., &amp; Fan, W. (2025). <a href="https://www.nature.com/articles/s41599-025-04787-y">The effect of ChatGPT on students’ learning performance, learning perception, and higher-order thinking: insights from a meta-analysis</a>. <em>Humanities and Social Sciences Communications</em>. <strong>Retracted 22 April 2026</strong> (<a href="https://www.nature.com/articles/s41599-026-07310-z">retraction note</a>).</li><li>Lee, H. P. et al. (2025). <a href="https://www.microsoft.com/en-us/research/wp-content/uploads/2025/01/lee_2025_ai_critical_thinking_survey.pdf">The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers</a>. <em>CHI ‘25</em>.</li><li>Bjork, R. A., &amp; Bjork, E. L. (2011). <a href="https://www.researchgate.net/publication/284097727_Making_things_hard_on_yourself_but_in_a_good_way_Creating_desirable_difficulties_to_enhance_learning">Making things hard on yourself, but in a good way: Creating desirable difficulties to enhance learning</a>. <em>Psychology and the Real World</em>.</li><li>Sparrow, B., Liu, J., &amp; Wegner, D. M. (2011). <a href="https://dtg.sites.fas.harvard.edu/DANWEGNER/pub/Sparrow%20et%20al.%202011.pdf">Google effects on memory: Cognitive consequences of having information at our fingertips</a>. <em>Science</em>, 333(6043).</li></ul><p><em>Originally published at </em><a href="https://maroffo.github.io/blog/posts/2026-05-07-the-word-permanent-isnt-in-the-paper/"><em>https://maroffo.github.io</em></a><em> on May 7, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1ef084080197" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Cognitive Load Moved. My Pre-Work Routine Had to Move With It.]]></title>
            <link>https://maroffo.medium.com/cognitive-load-moved-my-pre-work-routine-had-to-move-with-it-10ffb0c7cac4?source=rss-c6cd0b9646f5------2</link>
            <guid isPermaLink="false">https://medium.com/p/10ffb0c7cac4</guid>
            <dc:creator><![CDATA[Massimiliano Aroffo]]></dc:creator>
            <pubDate>Tue, 05 May 2026 00:00:49 GMT</pubDate>
            <atom:updated>2026-06-18T14:22:44.455Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*h18vJtKMrmC1vZHk.png" /></figure><p>This morning, scrolling Twitter before coffee had taken effect, I came across someone resurfacing a 2014 Stanford paper called <em>Give your ideas some legs: the positive effect of walking on creative thinking</em>. I clicked it expecting another lifestyle-content rerun. What I read instead made me think about my standing desk.</p><p>Since I switched to a standing desk a while ago, I’ve drifted into a habit I never planned. When a problem won’t unstick, I walk to the kitchen. When the afternoon gets heavy, I take the dog out for a two-minute pee in the courtyard. Some afternoons I walk the ten minutes to my son’s school and back, and I notice that the next hour at the laptop is sharper than the one before. None of it was a deliberate routine. It was just where my body went when my head jammed.</p><p>Three weeks ago I <a href="https://maroffo.github.io/blog/posts/2026-03-06-ai-didnt-reduce-my-cognitive-load-it-moved-it/">wrote about how AI didn’t reduce my cognitive load</a>, it moved it. The typing got easier and the thinking got harder. The afternoons fill up with evaluation, architectural judgment, holding a correct mental model in place while scanning AI output for <em>almost right</em> deviations. What I didn’t write was the second half of that observation: there is a part of the day when that work goes well, and a part when it doesn’t, and the difference correlates suspiciously well with whether I’ve moved my legs in the previous hour.</p><p>The Twitter post pulled me into the rest of the literature, and the literature describes the pattern I’d been running without knowing it. The other thing the literature is clear about: stepping away from the desk to scroll the same Twitter feed in a different posture is not the same intervention. The body has to actually move.</p><h3>Two phases, two cognitive modes<a href="https://maroffo.github.io/blog/posts/2026-05-05-cognitive-load-moved-my-pre-work-routine-had-to-move-with-it/#two-phases-two-cognitive-modes">##</a></h3><p>Before talking about walking, I need to be specific about what the hard part of an AI-assisted day actually demands, because it isn’t one thing.</p><p>Cognitive psychology splits creative thinking into two modes.</p><p><strong>Divergent thinking</strong>: generating multiple, novel solutions, breaking out of conventional categories, holding many possibilities in tension. The standard measure is Guilford’s Alternate Uses Test, how many unconventional uses can you think of for a brick. Scoring combines fluency (how many ideas), flexibility (how many distinct categories), and originality (how rare).</p><p><strong>Convergent thinking</strong>: narrowing down to the single correct answer. The standard measure is the Remote Associates Test, find the one word that connects “time”, “hair”, “stretch” (answer: long).</p><p>In a typical AI-assisted day, both modes show up, in two distinct phases.</p><p><strong>Phase 1, generative.</strong> Choosing an architecture, framing the problem, deciding which alternative is worth pursuing, writing the prompt that kicks off a Claude Code session. The work rewards broad associative reach and the ability to hold multiple framings in parallel before committing. It is divergent work.</p><p><strong>Phase 2, evaluative.</strong> Reading what the AI produced and comparing it against an internal model of what it should have done. Catching the <em>almost right</em> deviation buried in line 47 of a 200-line diff, deciding whether to accept, reject, or rewrite. This is the opposite of broad associative reach. It is narrow, focused, sustained attention against plausible-looking output, and it maps closer to convergent or critical evaluation than to divergent thinking.</p><p>I conflated these two phases for a long time, and most posts about “AI cognitive load” still do. The fatigue is real but it is not one thing, and the interventions that help the generative half don’t carry over to the evaluative half. One of them, as it turns out, may even make it worse.</p><p>That is what makes walking interesting as an intervention, and also what limits it.</p><h3>What the literature actually says (and doesn’t)<a href="https://maroffo.github.io/blog/posts/2026-05-05-cognitive-load-moved-my-pre-work-routine-had-to-move-with-it/#what-the-literature-actually-says-and-doesnt">##</a></h3><p>The starting point is Oppezzo and Schwartz 2014 at Stanford GSE. Four experiments, 176 participants, creative thinking tasks while sitting and while walking. The headline result is the asymmetry between modes: 81% of participants improved on Guilford’s Alternate Uses test (the divergent-thinking measure) when walking, while only 23% improved on the Compound Remote Associates test (the convergent measure). Walking didn’t lift creative output across the board, it lifted one mode and left the other roughly alone.</p><p>That specific asymmetry is what brought me back to a literature I had filed under “lifestyle content”. The studies aren’t claiming walking makes people sharper at everything. They are claiming it helps the cognitive mode I now spend my Phase 1 in, and roughly nothing for the mode AI already handles.</p><p>The follow-on literature is smaller than the popular write-ups make it sound, but it exists.</p><p>The 2022 multilevel meta-analysis by Rominger and colleagues pooled 28 studies (1,624 participants, 115 effect sizes) on acute physical activity and creative ideation. The aggregated effect on divergent thinking came out at Hedges’ <em>g</em> = 0.47, in the medium range. Activity type moderated the effect, and convergent thinking was not the meta-analysis’ focus, though the underlying literature doesn’t show the same magnitude there.</p><p>Also in 2022, Murali and Händel at Würzburg ran the experiment that matters most for the mechanism. They compared free walking against restricted walking (a fixed back-and-forth path), and free sitting against a fixated-gaze sitting condition. Free walking produced the strongest divergent-thinking improvement, and the restriction effect held inside the walking condition: constrained walking did less than free walking. The eye-blink data they collected differed across conditions but did not correlate with task performance, so the “broadened internally-directed attention” reading is suggestive at best, not established. The cleaner interpretation is the behavioural one: lack of motor restriction matters, not raw cardiovascular load.</p><p>Chen’s 2024 narrative review in <em>Discover Psychology</em> organized the field by activity type and intensity (21 studies, 24 experiments). The most consistent positive signal came from natural, unstructured walking; low-intensity structured exercise tended toward null results, and higher-intensity structured exercise was mixed. A narrative review is not a meta-analysis, so I take this as direction-of-evidence rather than effect size.</p><p>The most relevant paper for daily use is Rominger 2024 in <em>American Psychologist</em>. They put accelerometers on 157 young adults and combined step counts with ecological momentary assessment to track creative ideation in everyday life, not in a lab. Both acute step-count bouts and habitual step-count patterns predicted higher originality on verbal divergent-thinking tasks. The framing I want to keep from this paper is that acute and chronic effects appear to compound, not that this is the first ambulatory study of creativity, because it isn’t.</p><p>The honest summary is about a decade of converging but modest evidence, almost all of it on healthy young adults using verbal divergent-thinking tasks, with effect sizes clustering around <em>g</em> = 0.4. The finding has replicated, but the slice of behaviour it covers is narrow.</p><h3>The mechanism (best current guess)<a href="https://maroffo.github.io/blog/posts/2026-05-05-cognitive-load-moved-my-pre-work-routine-had-to-move-with-it/#the-mechanism-best-current-guess">##</a></h3><p>Why would walking specifically help divergent thinking, and why specifically <em>free</em> walking?</p><p>The most cited account is Zhou, Hommel and colleagues 2017, the “control-depletion” framing. Divergent thinking benefits from a loosening of top-down executive control, which is what cognitive flexibility requires. The more an activity competes for control resources, the more flexible the residual control style becomes, and the better divergent-thinking tasks unfold. Their experiments tested standing, predetermined-pattern walking, and free walking; free walking produced the strongest performance, predetermined-pattern walking did less, standing alone did less still.</p><p>Murali and Händel’s free-versus-restricted result fits this picture: a fixed path forces a small amount of monitoring (am I about to turn?), and that monitoring draws back exactly the executive control the divergent task wanted to lose. A free walk asks for none.</p><p>The two accounts converge on the same prediction: what helps divergent thinking is not cardiovascular load or “more blood to the brain”, it is having executive control occupied by something undemanding so attention can broaden.</p><p>The implication for Phase 2 is uncomfortable. If broadening attention and loosening control is what helps Phase 1, then the same intervention should hurt Phase 2, where the work is sustained, narrow, focused attention against plausible-looking output. The literature is consistent with this: walking does roughly nothing for convergent-thinking measures, and could plausibly hurt the kind of vigilance that catches <em>almost right</em>. So this is not a generic “walking makes AI work better” claim. It is a specific intervention, for one of the two phases, before specific kinds of work.</p><p>Two distinct interventions, one for each phase. The first is supported by the literature above. The second is what I observe in my own week and label more cautiously.</p><p><strong>Phase 1: a real walk before the generative work.</strong></p><p>Five to ten minutes, before opening any AI tool. Not a treadmill, not a timed loop, not while reading on a phone. The school run on foot, the dog around the courtyard, a slow loop of the block with no destination and no pace target. The mornings I do the school drop-off on foot are reliably the mornings the architecture work goes best afterwards, which I now suspect is not coincidence. The point is to move without thinking about how you’re moving; if I find myself checking my watch, I’m doing it wrong.</p><p>I try to sit down to the work within ten minutes of returning. Oppezzo’s residual-boost data is the reason: people who sat after walking still outscored people who never walked, but the effect isn’t all-day, so the earlier the better. The work that benefits is the generative half (architecture decisions, problem framing, ADR drafting, the prompt that kicks off a long Claude Code session), not the kind of day where I’m implementing against a frozen spec or working a known bug.</p><p>One practical detail: I resist the temptation to start the Claude Code session on the walk via voice. The walk is supposed to broaden attention; prompting an LLM does the opposite, and running both at once cancels the benefit.</p><p><strong>Phase 2: short loops away from the desk, for vigilance recovery.</strong></p><p>This one I label more cautiously, because the divergent-thinking literature does not predict it and may even predict the opposite. The pattern in my own week is consistent, but I’m describing observation, not evidence.</p><p>The signal I’ve learned to recognise: I’m at the standing desk, halfway through a long Claude Code diff, and somewhere around the 20-minute mark my filter starts to slip. The acceptances come faster than the scrutiny, and the suggestions begin to look better than they actually are. The mistakes I’d catch fresh slide past. When I notice it, I leave the studio. A walk to the kitchen and back, the dog out for a quick pee, two laps of the corridor if I really can’t leave the house. Two or three minutes. Then I come back, restart the diff from the top of the section I was on, and the second pass tends to pick up what the first one missed.</p><p>What does <em>not</em> work for me is staying at the desk and switching to Twitter or HackerNews for two minutes. The posture is the same, the screen is the same, the only thing that changes is the source of input. The filter doesn’t reset. The body has to physically leave.</p><p>This is not the same intervention as the morning walk. It is shorter, it is reactive, and the literature it most resembles is on vigilance and attention restoration, not divergent thinking. The 1950 Mackworth result I cited in the predecessor post is the anchor: sustained focused attention degrades within roughly half an hour, and brief breaks help reset it. Take this as a habit I run because it tracks with what I observe, not as something the divergent-thinking studies vouch for.</p><p><strong>The chronic and acute effects compound.</strong></p><p>The Rominger 2024 ecological data is the part of the literature I found most useful. Both the acute effect (walking before a creative task) and the chronic effect (walking regularly) predict higher originality, and they appear to add. A daily five-minute pre-work walk is not a one-shot intervention; it’s also a deposit into a baseline that the next morning’s walk draws on. Skipping for a week probably costs more than the missed five minutes.</p><p>A few cases where this routine doesn’t help, or hurts.</p><p>If the day is purely convergent work, debugging a known issue or implementing against a frozen spec, the morning walk costs a context switch I didn’t need. The matched intervention is sleep, caffeine, and the corridor break, not a five-minute attention-broadening session.</p><p>If I’m already in flow when I arrive at the desk, I don’t break it. The routine matters before the hard part starts, not in the middle of it.</p><p>On a day after bad sleep the walk doesn’t compensate. The acute effect in the literature stacks on top of habitual activity and a normal baseline, and none of those layers substitute for the others.</p><p>And the effect size is what it is: a meta-analytic <em>g</em> around 0.4 on divergent-thinking tasks in healthy young adults. That is a real effect, not a transformative one. There are null results in the broader exercise-and-cognition literature, and the file-drawer concern that applies to any small-effect field applies here too. Walking moves the dial on Phase 1; it does not turn anyone into a different engineer, and it does not fix Phase 2.</p><h3>An adjacent finding from a different population<a href="https://maroffo.github.io/blog/posts/2026-05-05-cognitive-load-moved-my-pre-work-routine-had-to-move-with-it/#an-adjacent-finding-from-a-different-population">##</a></h3><p>After writing this post I came across a 2025 systematic review that sits next to the walking literature without quite overlapping it.</p><p>Ge and colleagues, in <em>Frontiers in Public Health</em>, ran a Bayesian model-based network meta-analysis of 15 RCTs (1,196 participants) on exercise interventions for cognitive function in adults with major depressive disorder. They report a non-linear dose-response, with a minimum effective dose around 967 METs-minutes per week (inside the WHO-recommended range), and a ranking by exercise type that puts Tai Chi/Qigong first, mixed aerobic and resistance training next, and walking/jogging third. Their MCID is 0.5 SD, the standard threshold for clinically meaningful change in MoCA-style assessments.</p><p>This doesn’t unsettle the walking finding for the use case I described. Their outcome is general cognitive function measured in a depressed sample; mine is divergent-thinking ideation in healthy knowledge workers, a much narrower slice. The Phase 1 free-walking effect Oppezzo and Murali document is a specific intervention for a specific cognitive mode, and the Ge paper does not measure that mode at all. It suggests something different: if your outcome is general cognitive recovery, especially in clinical populations or older adults, the evidence points more strongly toward mind-body practices than toward unstructured walking. That’s a different question than the one the morning routine answers, but worth knowing the answer to.</p><p>Caveats: 15 trials, mostly small, single-population (MDD), with moderation effects suggesting younger samples and samples with more female participants respond more strongly. Replication in non-clinical adults is what’s still missing.</p><p>Three weeks ago I argued that AI moved cognitive load rather than reducing it. The follow-up question was: if the load moved, what changes about how I set up the day?</p><p>The honest answer in my case turned out not to be a routine I designed. It was a routine my body had built around the standing desk, the school run, and the dog, while I wasn’t paying attention. Reading the 2014 paper this morning was less a discovery than a label for something I was already doing. The literature gave it a mechanism (free, unconstrained walking loosens executive control, which is what divergent thinking wants) and also a clean limit: it’s a Phase 1 intervention, mostly, and the evaluation half of the day needs something different.</p><p>Most of the “AI productivity” advice I see online keeps optimising the part that’s already optimised: more tools and orchestration around the convergent work AI already handles well. The part that got harder is the thinking, and the thinking, once you look at it carefully, splits into two phases that don’t respond to the same intervention. Neither of mine is a transformative habit. The school drop-off on foot, the dog out for a pee, the kitchen-and-back when stuck. They are smaller and more domestic than the version of this story I’d expect to read on LinkedIn, which is roughly why I trust them.</p><ul><li>Oppezzo, M., &amp; Schwartz, D. L. (2014). Give your ideas some legs: The positive effect of walking on creative thinking. <em>Journal of Experimental Psychology: Learning, Memory, and Cognition</em>, 40(4), 1142–1152. <a href="https://www.apa.org/pubs/journals/releases/xlm-a0036577.pdf">PDF</a></li><li>Rominger, C., Schneider, M., Fink, A., Tran, U. S., Perchtold-Stefan, C. M., &amp; Schwerdtfeger, A. R. (2022). Acute and chronic physical activity increases creative ideation performance: A systematic review and multilevel meta-analysis. <em>Sports Medicine — Open</em>, 8(62). <a href="https://sportsmedicine-open.springeropen.com/articles/10.1186/s40798-022-00444-9">Open access</a></li><li>Murali, S., &amp; Händel, B. (2022). Motor restrictions impair divergent thinking during walking and during sitting. <em>Psychological Research</em>, 86, 2144–2157. <a href="https://pmc.ncbi.nlm.nih.gov/articles/PMC8742166/">PMC</a></li><li>Chen, C. (2024). Exploring the impact of acute physical activity on creative thinking: A comprehensive narrative review with a focus on activity type and intensity. <em>Discover Psychology</em>, 4(3). <a href="https://link.springer.com/article/10.1007/s44202-024-00114-9">Open access</a></li><li>Rominger, C., Fink, A., Weber, B., Benedek, M., Perchtold-Stefan, C. M., &amp; Schwerdtfeger, A. R. (2024). Step-by-step to more creativity: The number of steps in everyday life is related to creative ideation performance. <em>American Psychologist</em>, 79(6), 863–875. <a href="https://psycnet.apa.org/record/2024-26043-001">Reference</a></li><li>Zhou, Y., Zhang, Y., Hommel, B., &amp; Zhang, H. (2017). The impact of bodily states on divergent thinking: Evidence for a control-depletion account. <em>Frontiers in Psychology</em>, 8, 1546. <a href="https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5626876/">PMC</a></li><li>Ge, J., Wang, Q., Wang, S., Xiong, X., &amp; Zhai, Y. (2025). Optimal dose and type of exercise to improve cognitive function in adults with major depressive disorder: A systematic review and Bayesian model-based network meta-analysis. <em>Frontiers in Public Health</em>, 13, 1662778. <a href="https://doi.org/10.3389/fpubh.2025.1662778">Open access</a></li></ul><p><em>Originally published at </em><a href="https://maroffo.github.io/blog/posts/2026-05-05-cognitive-load-moved-my-pre-work-routine-had-to-move-with-it/"><em>https://maroffo.github.io</em></a><em> on May 5, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=10ffb0c7cac4" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Agents Delete Good Code. Here’s How I’m Stopping Them.]]></title>
            <link>https://maroffo.medium.com/agents-delete-good-code-heres-how-i-m-stopping-them-f0489e827a3b?source=rss-c6cd0b9646f5------2</link>
            <guid isPermaLink="false">https://medium.com/p/f0489e827a3b</guid>
            <dc:creator><![CDATA[Massimiliano Aroffo]]></dc:creator>
            <pubDate>Sun, 05 Apr 2026 00:00:13 GMT</pubDate>
            <atom:updated>2026-06-18T14:21:51.472Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*F8EHiHhPz26WeiDl.png" /></figure><p>Part six in a series on building <a href="https://github.com/maroffo/claude-forge">claude-forge</a>: a modular skills and orchestration system for <a href="https://docs.anthropic.com/en/docs/claude-code">Claude Code</a>. This time, the system learns to supervise itself.</p><p><em>I went looking for token waste. I found something worse.</em></p><h3>The Promise I Didn’t Keep<a href="https://maroffo.github.io/blog/posts/2026-04-05-agents-delete-good-code-heres-how-im-stopping-them/#the-promise-i-didnt-keep">#</a></h3><p><a href="https://maroffo.github.io/blog/posts/2026-03-31-your-ai-harness-is-hand-crafted-thats-the-problem/">Post five</a> ended with a number: 3,723 tokens, always loaded, zero evidence they’re the right ones. I said I’d report back when it changed.</p><p>It hasn’t changed yet. The harness-mechanic agent needs a few more weeks of trace data before it can propose optimizations with confidence. But while waiting for enough sessions to accumulate, I started reading the traces manually. Not for token efficiency. Just to understand what was happening.</p><p>What I found wasn’t a token problem. It was a behavior problem.</p><p>In one session, a software-engineer agent was asked to refactor a validation layer. The plan called for consolidating three scattered checks into a single module. The agent did consolidate them, but it also deleted 40 lines of edge-case handling that it apparently found too complex to integrate. The tests still passed because the agent had subtly weakened them: instead of asserting specific error messages, the new tests just checked that <em>an</em> error was thrown. The quality gates scored it 83. Technically committable.</p><p>In another session, a multi-subtask implementation drifted by the second subtask. The plan said “add rate limiting middleware, then add the rate limit headers to the response.” The agent implemented rate limiting (subtask 1), then in subtask 2 decided that rate limit headers were “redundant given the middleware already handles enforcement.” It skipped half the work and wrote a confident explanation for why it was unnecessary. Score: 87.</p><p>These aren’t hallucinations. These aren’t capability failures. These are agents making strategic decisions to avoid difficulty, and getting away with it because the harness wasn’t watching for this specific pattern.</p><h3>Naming the Demons<a href="https://maroffo.github.io/blog/posts/2026-04-05-agents-delete-good-code-heres-how-im-stopping-them/#naming-the-demons">#</a></h3><p>A few days later, I came across an article by <a href="https://x.com/systematicls">@systematicls</a> titled “How To Solve Problems Of Long Running, Autonomous Agentic Engineering Workflows.” It’s a taxonomy of agent failure modes, and reading it felt like someone had watched my trace data and written field notes.</p><p>The article identifies eight ways agents can be, to use their word, stupid. Not all of them were new to me, but the framing was: these failures aren’t random. They’re predictable, recurring patterns that map to specific phases of an agentic workflow.</p><p>I mapped these against our existing harness. Five were already covered: requirements refinement handles pre-task context, the annotation cycle and /second-opinion cover planning quality, R1-R4 deviation rules catch plan deviations during implementation, TDD with verification tables and the test-design-reviewer address verification laziness.</p><p>Three were not covered at all.</p><p><strong>Complexity fear</strong> was the validation refactor I’d seen in the traces: the agent deleted edge cases rather than integrate them. The article puts it well: “Agents have a deep fear of complexity. If you ask them to implement a 5-line function, no problem. If you make them believe they are going to have to write a 50,000-line class, they start to weasel their way out of it.”</p><p><strong>Entropy maximization</strong> was the documentation drift I’d been noticing for weeks: agents change function signatures but don’t update the docstrings two files over. Change a CLI flag name but leave the README referencing the old one. Repeat this across dozens of sessions and the repository becomes a minefield of contradictions.</p><p><strong>Mid-implementation drift</strong> was the rate-limiting session: subtask 1 was fine, subtask 2 built on a wrong assumption, and nobody checked between them. The article calls this “planning stickiness,” the risk where your agent deviates from the plan and does whatever it wants. I’d been catching this at review time (step 3 of the orchestrator), but by then the damage is done and the fix loop costs extra tokens.</p><p>The article also offered a meta-insight that stuck with me: “An interesting side effect of having studied productivity methods and being a practitioner of them is realizing how effective they are at managing agents as well.” Break complex work into small pieces. Verify early and often. Don’t let anyone (human or LLM) self-assess their own output. These aren’t AI-specific techniques. They’re management techniques, applied to silicon.</p><h3>Engineering the Supervisor<a href="https://maroffo.github.io/blog/posts/2026-04-05-agents-delete-good-code-heres-how-im-stopping-them/#engineering-the-supervisor">#</a></h3><p>Separately, I’d been reviewing an internal proposal for a “Judge Sub-Agent”: a dedicated, lightweight agent that would intercept the coding agent’s actions at defined checkpoints and verify them against a policy registry.</p><p>The architecture is elegant. The coding agent stays focused on the task. Policy knowledge (coding standards, security rules, architectural constraints) lives in the Judge’s context, not the coder’s. When the Judge catches a violation, it returns specific, actionable feedback: “This function uses bare except clauses. Per Python standards, catch specific exceptions.” The coding agent doesn’t need to know all the rules; it just receives targeted corrections.</p><p>The core insight, context offloading, is the right idea. But for claude-forge’s stack, the implementation doesn’t fit. Claude Code doesn’t have a middleware layer where you can intercept tool calls and route them to a separate agent before execution. You’d need a hook for every tool call, which is impractical. And running a synchronous blocking judge on every action would destroy latency.</p><p>But we already have the pattern, just at a coarser granularity. Our review agents (architecture, security, performance, test, dx, database, dependency) run with fresh context after implementation and report findings that feed into the fix loop. The coding agent doesn’t carry policy knowledge; the reviewers do. Context offloading, achieved through the orchestrator’s review step rather than through real-time interception.</p><p>What we were missing wasn’t the architecture. It was coverage at the right granularity: mid-implementation (between subtasks, not just post-implementation) and post-fix (checking blast radius, not just the changed files).</p><p>I ran a /second-opinion with Gemini on the proposed mechanisms before building them. Three corrections came back that changed the implementation in ways I wouldn&#39;t have found on my own.</p><p>My original anti-stub rule was an absolute ban on TODO, pass, NotImplementedError, and placeholder comments. Gemini pointed out that this breaks interface-first development and TDD: sometimes you write a stub intentionally as part of the plan, then fill it in. The refined rule is &quot;no <em>unplanned</em> stubs.&quot; If the plan says &quot;deferred&quot; or &quot;interface-only,&quot; stubs are fine. If the plan says &quot;implement X&quot; and you write a stub, that&#39;s a CRITICAL violation. Plan-aware, not absolute.</p><p>The blast radius check was originally unconditional, running after every fix loop. Gemini pushed back: “Running a fresh-context LLM check indiscriminately on every loop will be prohibitively expensive and slow.” Fair. The refinement: trigger only when changed files modify public APIs, when more than three files changed, or when schema changes occurred. Use a cheap grep pre-filter to find files that reference changed symbols, then send only the flagged files to a fresh-context agent. If grep finds zero related files outside the changed set, skip the agent entirely.</p><p>The third correction was the one that mattered most: “If the implementation agent checks its own drift, it will likely justify its deviations to itself. Context anxiety means the agent is already confused; asking it to self-correct in the same context window rarely works.” The drift check had to be an isolated agent with fresh context, receiving only the subtask description and the git diff. No history, no accumulated rationalization. An agent can’t grade its own homework.</p><h3>The Four Mechanisms<a href="https://maroffo.github.io/blog/posts/2026-04-05-agents-delete-good-code-heres-how-im-stopping-them/#the-four-mechanisms">#</a></h3><p>Here’s what we shipped. All four are markdown additions to existing harness files: two new deviation rules in the software-engineer agent, two new orchestrator steps.</p><h3>R5: No Unplanned Stubs<a href="https://maroffo.github.io/blog/posts/2026-04-05-agents-delete-good-code-heres-how-im-stopping-them/#r5-no-unplanned-stubs">#</a></h3><p>The rule is simple: if the plan says “implement X,” you implement X fully, production-ready. Introducing a stub without plan authorization is a CRITICAL violation (auto-fail, score = 0) because downstream code will be wired to a hollow implementation, and the stub will silently persist across sessions.</p><p>But there’s a subtler version of the same problem that the @systematicls article calls “entropy maximization” and that I think deserves its own name: <strong>conservation of complexity.</strong> Agents don’t just write stubs for new code. They delete existing code they find too complex to work with. The validation refactor I mentioned earlier is a perfect example: the agent didn’t write a stub, it removed 40 lines of edge-case handling and wrote weaker tests to cover its tracks.</p><p>The guard: if you delete more than 20% of a file’s lines or remove existing functions, you must document what was removed and why, prove no existing tests were deleted or weakened, and confirm no callers still reference the removed code (a grep check). This isn’t a ban on deletion. It’s a requirement for justification. The cost of writing three sentences is negligible; the cost of silently removing critical logic is not.</p><h3>R6: Proportionality Guard<a href="https://maroffo.github.io/blog/posts/2026-04-05-agents-delete-good-code-heres-how-im-stopping-them/#r6-proportionality-guard">#</a></h3><p>Before any destructive action (deleting files, overwriting modules, restructuring directories, dropping tables), the agent must verify four things: necessity (is this required by the plan?), scope (is the blast radius proportional to the task?), reversibility (is there a safer alternative, like rename instead of delete?), and justification (logged in the implementation report).</p><p>This sounds obvious, but the failure mode is real. An agent working on a “fix typo in README” task should never be restructuring the src/ directory. The problem isn&#39;t malice; it&#39;s that agents don&#39;t naturally evaluate whether an action is proportional to the task that spawned it. They optimize locally (this restructure would make the code better) without checking globally (but the task was a typo fix).</p><h3>Mid-Implementation Drift Check (Step 1b)<a href="https://maroffo.github.io/blog/posts/2026-04-05-agents-delete-good-code-heres-how-im-stopping-them/#mid-implementation-drift-check-step-1b">#</a></h3><p>For multi-subtask implementations, after each subtask completes, the orchestrator spawns a lightweight judge. The judge receives only two things: the subtask description from the plan, and the git diff of changes made during the subtask. It answers one question: “Did we build exactly this, no more, no less?”</p><p>Three verdicts: aligned (proceed), minor drift (log a warning, proceed), significant drift (stop, correct before next subtask).</p><p>The important part: the judge is isolated. Fresh context, no history of the implementation session, no accumulated rationalization. Same principle as our existing review agents, extended to the gap between subtasks where cascading deviations used to slip through.</p><p>The cost is real: one additional agent invocation per subtask. For a three-subtask implementation, that’s three extra calls. But catching a drifted subtask 1 before subtask 2 builds on it saves the entire fix loop that would otherwise be needed at review time. It’s cheaper to check early than to fix late. (Anyone who’s worked in quality engineering will recognize this as the “cost of quality” curve, and it applies to LLM workflows exactly as it applies to manufacturing.)</p><h3>Blast Radius Check (Step 5b)<a href="https://maroffo.github.io/blog/posts/2026-04-05-agents-delete-good-code-heres-how-im-stopping-them/#blast-radius-check-step-5b">#</a></h3><p>After the fix loop and re-verification, before scoring, the orchestrator checks whether any changed files modified public APIs, whether more than three files were touched, or whether schema changes occurred. If none of these conditions are met, the step is skipped entirely.</p><p>When it triggers, it runs in two phases. First, a cheap grep for references to changed function names, class names, and endpoints across the repo. This produces a list of “related files”: importers, documentation that references changed APIs, tests that assert on changed behavior. Second, a fresh-context agent receives the changed files summary and the related file snippets (not full files) and checks for contradictions: documentation that describes old behavior, tests asserting on removed functionality, comments referencing deleted logic.</p><p>Findings are classified as MAJOR (functional contradiction, like a doc saying the function returns X when it now returns Y) or MINOR (stale comment), and they feed into the quality gate scoring.</p><p>The grep pre-filter is the pattern I keep coming back to: <strong>use deterministic, token-free tools to narrow the search space before spending tokens on LLM review.</strong> Linters before code review. grep before blast radius analysis. The cheapest check should always run first, and if it finds nothing, the expensive check doesn’t run at all.</p><h3>The Cost of Supervision<a href="https://maroffo.github.io/blog/posts/2026-04-05-agents-delete-good-code-heres-how-im-stopping-them/#the-cost-of-supervision">#</a></h3><p>None of this is free. I added four safety mechanisms to a harness that was already at 3,723 tokens always-on, and each one eats context that could otherwise go to the actual task.</p><p>The two new deviation rules (R5 and R6) add roughly 350 tokens to the software-engineer agent definition. On-demand file, loaded only when the orchestrator spawns an implementation agent, so the always-on baseline stays clean.</p><p>The drift check adds one lightweight agent call per sequential subtask. The agent’s context is minimal (subtask description + git diff), so each call is cheap: maybe 2,000–4,000 input tokens. For a typical three-subtask implementation, that’s 6,000–12,000 tokens of supervision. The question is whether this costs less than the alternative: a full fix loop when a drifted subtask is caught at review time, which can easily consume 20,000–40,000 tokens of re-implementation and re-review.</p><p>The blast radius check, when it triggers, costs a grep operation (free) plus one agent call with scoped context. In practice, the grep pre-filter will reduce or eliminate the agent cost for most sessions. Public API changes that affect other files are common enough to matter but rare enough that the check won’t run on every loop iteration.</p><p>The orchestrator protocol itself grew from 1,075 tokens to approximately 1,450 tokens, adding ~375 to the always-on budget. New total: roughly 4,100 tokens, up from 3,723. A 10% increase.</p><p>Is it worth it? I genuinely don’t know yet. If the drift check catches significant deviations in 1 out of 10 multi-subtask sessions, it pays for itself easily. If it fires once in a hundred sessions, it’s dead weight. The harness-mechanic, once it has enough trace data, will be the one to make that call. I’m deliberately not guessing.</p><h3>What Remains Unsolved<a href="https://maroffo.github.io/blog/posts/2026-04-05-agents-delete-good-code-heres-how-im-stopping-them/#what-remains-unsolved">#</a></h3><p>The @systematicls article identifies context anxiety as the single biggest problem during implementation. “As a function of time, agents become more and more desperate to end the session.” I’ve seen this too: quality degrades as sessions get longer, shortcuts multiply, and the agent starts treating every remaining task as something to rush through.</p><p>Our mitigation is partial. The .continue-here.md file captures state before context gets large, and the anti-compression doctrine (regenerate from source files, never compress an existing summary) prevents information loss during handoffs. But there&#39;s no proactive detection of &quot;your context is bloating, time to hand off.&quot; The agent doesn&#39;t know it&#39;s getting anxious. By the time the symptoms are visible in the output, the damage is already cascading.</p><p>The article suggests “smart session handoffs where you can relieve your agents of their context.” This is architecturally simple (structured handoff documents with repository-aware compaction) but operationally hard in Claude Code, which manages its own context window. You’d need the orchestrator to monitor token usage, detect degradation signals, and trigger a session break. We’re not there yet.</p><p>There’s also a broader question that I keep circling back to: <strong>how much supervision is too much?</strong> Every guard, every check, every isolated judge is a bet that the agent will fail in a specific way. If frontier models get better at long-context reasoning, some of these mechanisms become dead weight. If they don’t, we’ll need more. The harness-mechanic’s job, eventually, is to detect which guards are earning their token cost and which are insurance premiums against risks that never materialize.</p><p>For now, the traces will accumulate. The four mechanisms are live. The next post in this series will have actual data: how often the drift check fires, how many blast radius contradictions surface, whether the anti-stub rule ever triggers a CRITICAL. Numbers, not intuitions. That was the promise of post five, and I intend to keep it.</p><p><em>This post was written with Claude Code (claude-forge orchestrator) and reviewed by Gemini via /second-opinion. The implementation is at </em><a href="https://github.com/maroffo/claude-forge"><em>maroffo/claude-forge</em></a><em> on the </em><em>feat/meta-harness branch. Four files changed: </em><em>agents/software-engineer/AGENT.md (+R5, +R6), </em><em>rules/orchestrator-protocol.md (+step 1b drift check, +step 5b blast radius), </em><em>rules/quality-gates.md (+6 new scoring criteria), </em><em>LEARNING.md (retrospective).</em> <em>Part of a series: </em><a href="https://maroffo.github.io/blog/posts/2025-11-09-from-asking-claude-to-code-to-teaching-claude-our-patterns-building-modular-ai-skills/"><em>post 1</em></a><em> (skills), </em><a href="https://maroffo.github.io/blog/posts/2026-02-12-when-your-ai-skills-library-gets-too-smart-for-its-own-context-window/"><em>post 2</em></a><em> (context window), </em><a href="https://maroffo.github.io/blog/posts/2026-02-19-the-missing-step-what-a-colleagues-hint-taught-me-about-ai-driven-planning/"><em>post 3</em></a><em> (planning), </em><a href="https://maroffo.github.io/blog/posts/2026-02-26-when-your-ais-second-brain-starts-talking-back/"><em>post 4</em></a><em> (second brain), </em><a href="https://maroffo.github.io/blog/posts/2026-03-31-your-ai-harness-is-hand-crafted-thats-the-problem/"><em>post 5</em></a><em> (measurement).</em></p><p><em>Originally published at </em><a href="https://maroffo.github.io/blog/posts/2026-04-05-agents-delete-good-code-heres-how-im-stopping-them/"><em>https://maroffo.github.io</em></a><em> on April 5, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f0489e827a3b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Your AI Harness Is Hand-Crafted. That’s the Problem.]]></title>
            <link>https://maroffo.medium.com/your-ai-harness-is-hand-crafted-thats-the-problem-0cdf70fcdc01?source=rss-c6cd0b9646f5------2</link>
            <guid isPermaLink="false">https://medium.com/p/0cdf70fcdc01</guid>
            <dc:creator><![CDATA[Massimiliano Aroffo]]></dc:creator>
            <pubDate>Tue, 31 Mar 2026 00:00:38 GMT</pubDate>
            <atom:updated>2026-06-18T14:20:58.496Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*jfXTJooY3w255xoO.png" /></figure><p>Part five in a series on building <a href="https://github.com/maroffo/claude-forge">claude-forge</a>: a modular skills and orchestration system for <a href="https://docs.anthropic.com/en/docs/claude-code">Claude Code</a>. This time, the system starts measuring itself.</p><p><em>3,723 tokens. Always loaded. Zero evidence they’re the right ones.</em></p><h3>The Optimization We Never Did<a href="https://maroffo.github.io/blog/posts/2026-03-31-your-ai-harness-is-hand-crafted-thats-the-problem/#the-optimization-we-never-did">#</a></h3><p>In <a href="https://maroffo.github.io/blog/posts/2026-02-12-when-your-ai-skills-library-gets-too-smart-for-its-own-context-window/">post two</a>, I described how the skills library had bloated to the point where it was eating the context window it was supposed to help with. The fix was aggressive: 352 lines added, 2,009 deleted. Shared reference files. Progressive disclosure. Tables instead of prose.</p><p>It worked. The harness got leaner.</p><p>But here’s the thing I never questioned: <em>how did I know what to cut?</em></p><p>I eyeballed it. Read through skills, spotted redundancy, compressed sections that felt verbose. Classic developer intuition, applied to a problem that should embarrass anyone who’s ever written a benchmark.</p><p>I was hand-tuning a system that affects every line of code Claude writes, with no measurement infrastructure, no trace data, and no way to know if the changes actually helped.</p><p>Then a paper from Stanford arrived and told me, politely, that this is the wrong way to do it.</p><p>The paper is <a href="https://arxiv.org/abs/2603.28052">“Meta-Harness: End-to-End Optimization of Model Harnesses”</a> by Lee, Nair, Zhang, Lee, Khattab, and Finn. The core argument: the performance of an LLM system depends not only on the model weights but on the <em>harness</em>, the code that determines what information to store, retrieve, and present to the model.</p><p>Their definition of “harness” is broad: the system prompt, the context management strategy, the retrieval logic, the output parsing, the error recovery. Everything between “user sends a request” and “model produces a useful response.”</p><p>Their claim: <strong>harnesses are still designed by hand, and existing text optimizers compress feedback too aggressively.</strong> So they built Meta-Harness, an outer-loop system that searches over harness code. It uses an agentic proposer that can read the source code, scores, and execution traces of all previous candidates through a filesystem.</p><p>The numbers: on text classification, +7.7 points while using <em>4x fewer context tokens</em>. On 200 IMO-level math problems, +4.7 points, and the discovered harness transferred to five different models it had never seen. On agentic coding (TerminalBench-2), it beat the best hand-engineered baselines.</p><p>The takeaway that hit me: the proposer doesn’t need to be smarter than you. It needs access to data you don’t have. Specifically, structured execution traces of what actually happened during previous runs.</p><p>I looked at my setup. Four rule files, 11 agents, 42 skills. All hand-written markdown. No traces. No token counts. No feedback loop.</p><h3>The Measurement Gap<a href="https://maroffo.github.io/blog/posts/2026-03-31-your-ai-harness-is-hand-crafted-thats-the-problem/#the-measurement-gap">#</a></h3><p>Let me be specific about what “no measurement” means in practice.</p><p>My orchestrator-protocol rule defines a 10-step loop: refine, research, implement, verify, review, fix, re-verify, score, present, UAT. Each step can spawn specialized agents. The quality gates score output on a 0–100 scale. It looks rigorous.</p><p>But I had no idea:</p><ul><li>How many tokens the always-loaded rules consume</li><li>Which orchestrator steps fail most often</li><li>How many fix-loop rounds a typical session needs</li><li>Whether review agents are activated for the right file types</li><li>Which skills are loaded but never useful</li></ul><p>Every session vanished into Claude Code’s internal logs (raw JSONL, tens of thousands of lines, impractical to read). The knowledge of what worked and what didn’t evaporated at the end of each conversation.</p><p>Meta-Harness’s key insight is that the proposer needs “richer access to prior experience.” My harness had no memory of its own performance. Every session was a blank slate, the exact problem I’d solved for <em>code patterns</em> in post one, now recurring at the <em>meta level</em>.</p><h3>Phase 1: Teaching the Harness to Remember<a href="https://maroffo.github.io/blog/posts/2026-03-31-your-ai-harness-is-hand-crafted-thats-the-problem/#phase-1-teaching-the-harness-to-remember">#</a></h3><p>The first thing to build wasn’t an optimizer. It was a trace format.</p><p>I wrote a Python CLI tool ( harness-trace) that extracts structured traces from Claude Code&#39;s raw session JSONL files. Each trace is one line per orchestrator step:</p><p>Twelve step types: REFINE, RESEARCH, IMPLEMENT, VERIFY, REVIEW, FIX, RE_VERIFY, SCORE, LOOP, PRESENT, UAT, SUMMARY. Each with step-specific data payloads validated by Pydantic models.</p><p>The extraction is heuristic (parsing assistant messages for orchestrator step indicators via regex), which makes it fragile. But it serves two purposes: bootstrapping an initial trace corpus from past sessions, and validating the live traces that the orchestrator now emits directly.</p><p>That second part is the real mechanism. I added 18 lines to orchestrator-protocol.md instructing the orchestrator to append a JSONL line after each step. The rule knows which step it&#39;s on, which agents were activated, what the score was. The post-processing script is the fallback; the rule is the source of truth.</p><p><strong>Why JSONL and not OpenTelemetry?</strong> This was one of the first decisions, and the Gemini second-opinion was unambiguous: “OpenTelemetry is a trap here. OTel is designed for distributed microservices and dashboards. Your consumer is an LLM agent reading from a filesystem. Optimize for LLM readability.”</p><p>Fair point. The traces aren’t going to Grafana. They’re going to another LLM that needs to reason about patterns across sessions. Flat JSONL lines that a language model can parse natively beat nested span hierarchies every time.</p><h3>Phase 2: Counting Tokens for the First Time<a href="https://maroffo.github.io/blog/posts/2026-03-31-your-ai-harness-is-hand-crafted-thats-the-problem/#phase-2-counting-tokens-for-the-first-time">#</a></h3><p>The same CLI includes a baseline command that scans every rule, skill, and agent file, counts tokens with tiktoken, and classifies them by tier:</p><p>3,723 tokens. That’s what the harness costs before a single user message arrives. Every conversation starts with those 3,723 tokens already consumed, whether they’re useful for the current task or not.</p><p>Some observations:</p><p>The orchestrator protocol is the heavyweight at 1,075 tokens. It defines the entire autonomous loop, review routing tables, UAT process, and now trace capture. Is every section worth its token cost? I don’t know yet, because I don’t have enough traces to measure which sections actually influence behavior.</p><p>The top 5 on-demand skills consume more than the entire always-on budget combined: blog-writer (2,738), test-design-reviewer (2,530), pr-review (2,516), cognitive-load-analyzer (2,343), humanizer patterns (2,346). These only load when invoked, so their size matters less. But it’s worth knowing.</p><p>The total harness is 103,395 tokens of instructions across 87 files. Most of it is on-demand. The always-on fraction (3.6%) is surprisingly lean, which validates the progressive disclosure strategy from post two. But “lean” and “optimal” aren’t the same thing.</p><h3>Phase 3: An Agent That Reads Its Own Performance<a href="https://maroffo.github.io/blog/posts/2026-03-31-your-ai-harness-is-hand-crafted-thats-the-problem/#phase-3-an-agent-that-reads-its-own-performance">#</a></h3><p>This is where it gets Meta-Harness-shaped.</p><p>I built a harness-mechanic agent: a read-only agent that analyzes execution traces and token baselines, identifies systematic failure patterns, and proposes evidence-based changes to the harness files. It follows the same SCAN, CLASSIFY, PROPOSE, PRESENT cycle that the <a href="https://maroffo.github.io/blog/posts/2026-02-26-when-your-ais-second-brain-starts-talking-back/">knowledge-sync</a> skill uses for promoting vault patterns into skills, but with a different data source.</p><p>Instead of scanning the Second Brain for recurring code patterns, it scans execution traces for recurring failures:</p><p>Same step fails across sessions</p><p>Rule gap</p><p>VERIFY fails on lint in 3/5 sessions</p><p>Score stuck below threshold</p><p>Instruction ambiguity</p><p>2+ fix rounds, same MAJOR finding</p><p>Agent not activated when expected</p><p>Routing gap</p><p>No security review for auth code</p><p>Excessive loop rounds</p><p>Unclear success criteria</p><p>4 rounds to reach score 80</p><p>Always-on file &gt;1500 tokens</p><p>Token waste</p><p>Rule with low information density</p><p>Every proposal must cite specific trace data (session slug, step, values). No proposal is auto-applied; all are RED in the decision framework (requires human approval). One change per proposal, same as the autoresearch-prompt optimization loop: isolate variables, measure before and after.</p><p>The mechanic hasn’t run on real data yet. The traces need a few weeks to accumulate before there’s enough signal for pattern detection.</p><p>There’s an obvious objection: running an LLM agent to optimize your LLM instructions costs tokens too. Is the optimizer’s inference cost worth the savings? For the always-on rules (3,723 tokens, loaded in every session), probably yes: a one-time optimization that saves 500 tokens pays for itself after a handful of sessions. For on-demand skills that load twice a month, probably not. The mechanic needs to be selective.</p><p>There’s also an overfitting risk. If I spend two weeks writing Go and the mechanic analyzes only those traces, it might conclude that the Rails skill is “dead weight” and propose deleting it. Small sample sizes produce confident but wrong conclusions. The safeguard: certain rules are pinned (the --no-verify ban, security baselines, the decision framework). The mechanic can compress them but never remove their semantics. And proposals always require human approval, which is the ultimate overfitting detector.</p><h3>The Gemini Reorder<a href="https://maroffo.github.io/blog/posts/2026-03-31-your-ai-harness-is-hand-crafted-thats-the-problem/#the-gemini-reorder">#</a></h3><p>Here’s a confession about planning: I had it wrong.</p><p>My original execution order was: (1) build the optimization loop first (extend autoresearch-prompt to optimize harness files), (2) add trace capture, (3) measure tokens. Start with the exciting part, add measurement later.</p><p>I asked Gemini for a <a href="https://maroffo.github.io/blog/posts/2026-02-19-the-missing-step-what-a-colleagues-hint-taught-me-about-ai-driven-planning/">second opinion</a>. The response was blunt: “Reverse your order. You cannot optimize what you cannot measure.”</p><p>Three specific corrections:</p><p><strong>First</strong>, don’t build synthetic benchmarks. If you optimize your React skill against “build a ToDo app,” the optimizer will ruthlessly delete all advanced context (concurrent rendering, complex caching) because the benchmark didn’t need it. Use real historical sessions instead. The harness should generalize to the work you actually do, not to a toy benchmark.</p><p><strong>Second</strong>, static compression before dynamic loading. The paper’s 4x token reduction came from the optimizer finding <em>denser words</em>, not from lazy-loading mechanisms. Claude Code manages its own context window. Trying to inject dynamic skill loading would fight the tool.</p><p><strong>Third</strong>, the trace format decision. LLM-readable JSONL, not OpenTelemetry spans. The consumer matters more than the producer.</p><p>The reordered plan: traces first (build the measurement), tokens second (take the baseline), optimizer third (use the data). This is just the scientific method, but applied to a domain where developers (including me) habitually skip the measurement step.</p><p>Gemini also caught a real bug during code review. The multi-round step deduplication logic only allowed LOOP, FIX, and VERIFY to repeat across orchestrator rounds. But when the orchestrator loops back (score &lt; 80), steps like IMPLEMENT, REVIEW, and SCORE repeat too. The fix was two lines. The bug would have silently dropped trace data from every multi-round session.</p><h3>What the Traces Will Tell Me<a href="https://maroffo.github.io/blog/posts/2026-03-31-your-ai-harness-is-hand-crafted-thats-the-problem/#what-the-traces-will-tell-me">#</a></h3><p>The infrastructure is deployed. The baseline exists. Now I wait for data.</p><p>The metrics I’ll watch: average fix-loop rounds (should be 1, if the harness is clear enough), first-pass quality score distribution (where do sessions land before any fixes?), and whether the always-on 3,723 tokens can shrink without the scores dropping.</p><p>A practical middle ground before the full LLM mechanic runs: a deterministic script that flags obvious patterns. “Skill X loaded in 15 sessions, never influenced the output.” “VERIFY failed on lint in 4 of the last 6 sessions.” No inference cost, no overfitting risk, and it covers the low-hanging fruit.</p><p>Meta-Harness runs on purpose-built evaluation frameworks with thousands of examples. I have 10–20 sessions per week from a single developer. The optimization surface is smaller, the signal noisier. I’m not expecting 4x compression.</p><p>But here’s what I already know, before the mechanic runs a single trace: I was treating system prompts like prose when they’re code. Code gets profiled, benchmarked, version-controlled, and measured against production behavior. Prompts get vibes-checked and shipped. The gap between those two standards is where the easy improvements live.</p><p>3,723 tokens is my starting number. I’ll report back when it changes.</p><p><em>This post was written with Claude Code (claude-forge orchestrator) and reviewed by Gemini via /second-opinion. The Meta-Harness implementation is at </em><a href="https://github.com/maroffo/claude-forge/pull/11"><em>maroffo/claude-forge#11</em></a><em>. First token baseline: 3,723 always-on, 103,395 total across 87 files. The harness-mechanic agent has not yet run on real trace data; results in a future post.</em> <em>Part of a series: </em><a href="https://maroffo.github.io/blog/posts/2025-11-09-from-asking-claude-to-code-to-teaching-claude-our-patterns-building-modular-ai-skills/"><em>post 1</em></a><em> (skills), </em><a href="https://maroffo.github.io/blog/posts/2026-02-12-when-your-ai-skills-library-gets-too-smart-for-its-own-context-window/"><em>post 2</em></a><em> (context window), </em><a href="https://maroffo.github.io/blog/posts/2026-02-19-the-missing-step-what-a-colleagues-hint-taught-me-about-ai-driven-planning/"><em>post 3</em></a><em> (planning), </em><a href="https://maroffo.github.io/blog/posts/2026-02-26-when-your-ais-second-brain-starts-talking-back/"><em>post 4</em></a><em> (second brain).</em></p><p><em>Originally published at </em><a href="https://maroffo.github.io/blog/posts/2026-03-31-your-ai-harness-is-hand-crafted-thats-the-problem/"><em>https://maroffo.github.io</em></a><em> on March 31, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0cdf70fcdc01" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[From Cognitive Load to Cognitive Surrender]]></title>
            <link>https://maroffo.medium.com/from-cognitive-load-to-cognitive-surrender-9b10da9a91dd?source=rss-c6cd0b9646f5------2</link>
            <guid isPermaLink="false">https://medium.com/p/9b10da9a91dd</guid>
            <dc:creator><![CDATA[Massimiliano Aroffo]]></dc:creator>
            <pubDate>Mon, 23 Mar 2026 00:00:00 GMT</pubDate>
            <atom:updated>2026-06-18T14:20:01.053Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*g2su-GnUDsI6yg-P.png" /></figure><p>Three weeks ago, <a href="https://www.linkedin.com/posts/davide-carboni-phd-7a510a1_lai-porta-con-s%C3%A9-un-piccolo-paradosso-activity-7435356562908495872-uGYC/">a LinkedIn post by Davide Carboni</a> got me thinking about the cognitive cost of AI. I wrote about <a href="https://maroffo.github.io/blog/posts/2026-03-06-ai-didnt-reduce-my-cognitive-load-it-moved-it/">how AI moved my cognitive load</a> rather than reducing it. The typing shrank, the thinking expanded, and the thinking was harder because I was building mental models of code written by something that doesn’t share my context.</p><p>I described the symptoms. I’ve kept digging since, and a paper now gives them a name.</p><p><a href="https://xmau.com/wp/notiziole/2026/03/02/resa-cognitiva/">.mau.’s blog</a> pointed me to a Wharton paper by Steven Shaw and Gideon Nave: <a href="https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6097646">“Thinking, Fast, Slow, and Artificial: How AI is Reshaping Human Reasoning and the Rise of Cognitive Surrender”</a>. The premise: Kahneman’s dual-process model (System 1 for fast intuition, System 2 for slow deliberation) needs a third component. They call it <strong>System 3</strong>: external, automated, data-driven reasoning that originates from AI rather than the human mind.</p><p>System 3 is not a metaphor. It’s a functional cognitive agent that operates outside the brain boundary, <em>in silico</em>, and interacts with the two internal systems in real time. It can scaffold your deliberation, or replace your intuition with ready-made answers, or suppress reflective thinking altogether.</p><p>The dangerous path has a name: <strong>cognitive surrender</strong>. You accept AI output without critical evaluation, substitute it for your own reasoning. Not strategic delegation like using a calculator or checking GPS. An abdication. System 2 never fires.</p><p>Early neurological evidence hints that this isn’t just behavioral. A small MIT Media Lab study (Kosmyna et al., 2025; N = 54, preprint, not yet peer-reviewed) tracked brain activity via EEG over four months and found that LLM users showed reduced brain connectivity compared to participants working without AI. The sample is too small to draw conclusions, but the direction is consistent with stronger evidence from clinical settings (more on that below). The researchers call it “cognitive debt”: short-term effort savings that create long-term deficits in critical thinking and independent thought. System 2 doesn’t just go quiet. The infrastructure that supports it may start to atrophy.</p><h3>The experiment<a href="https://maroffo.github.io/blog/posts/2026-03-23-from-cognitive-load-to-cognitive-surrender/#the-experiment">#</a></h3><p>Shaw and Nave ran three preregistered experiments (N = 1,372; 9,593 trials) using modified Cognitive Reflection Test items, the kind of problems designed to have an intuitive-but-wrong answer and a correct-but-effortful one. Participants could optionally consult an AI chatbot. The trick: hidden seed prompts made the AI return the correct answer on some trials and a confidently wrong answer on others.</p><p>Participants consulted the AI on more than 50% of trials. When it was right, accuracy jumped ~25 percentage points over baseline. When it was wrong, accuracy dropped ~15 points. The effect size was large (Cohen’s h = 0.81). People followed the AI wherever it went. Here’s what got me: confidence increased when using AI, even on trials where the AI gave the wrong answer. People didn’t just accept bad answers. They believed in them more.</p><p>Follow-up experiments tried to break the pattern. Time pressure didn’t (Study 2). AI actually buffered the cost of rushing when it happened to be correct. Monetary incentives and per-item feedback softened it but didn’t eliminate it (Study 3). When the AI was wrong, performance still dropped.</p><p>Who surrenders most? People with higher trust in AI, lower Need for Cognition, lower fluid intelligence. This isn’t about lazy people. It’s about what happens to a brain when a confident external agent hands it an answer.</p><h3>What this changes<a href="https://maroffo.github.io/blog/posts/2026-03-23-from-cognitive-load-to-cognitive-surrender/#what-this-changes">#</a></h3><p>Three weeks ago, I thought I could out-discipline this problem. TDD as truth anchor, session discipline, knowing when to stop. The data says my coping strategies work for me but probably won’t scale to a whole team.</p><p>The distinction between <strong>cognitive offloading</strong> and <strong>cognitive surrender</strong> is what practitioners should take from this. Offloading: you delegate a task to AI while keeping System 2 active, evaluating, deciding. Surrender: you accept AI output as your own judgment without deliberation. Same tool, same interface, completely different cognitive posture. The line between them is invisible from the outside and easy to cross without noticing.</p><p>In my earlier post, I described working with AI as exhausting: constant evaluation, building mental models of alien code, redirecting when it drifts. That’s cognitive offloading, and it’s genuinely draining. Advait Sarkar at Microsoft Research, in their 2025 New Future of Work report: “When we outsource our reasoning to AI, we reduce ourselves to ‘middle managers for our own thoughts.’” Shaw and Nave give this a mechanism: because offloading is so taxing, the brain looks for an exit. Cognitive surrender <em>is</em> that exit. You stop evaluating, accept the output, move on. The fatigue from responsible AI use creates the incentive to use it irresponsibly. That’s the trap.</p><p>This reframes the “perception trap” from my earlier post (METR developers 19% slower with AI, believing they were 20% faster). Those developers weren’t necessarily surrendering to AI the whole time. My theory is that they oscillated: offloading when fresh, surrendering when tired, never noticing the transition. The confidence boost the paper measures (even on wrong answers) would paper over the shift. You feel productive because the effort stopped, not because the work improved.</p><p>The paper makes me less optimistic about those countermeasures scaling. Monetary incentives and explicit feedback only <em>partially</em> attenuate cognitive surrender. Awareness and discipline are necessary but not sufficient. You can’t just tell a team “be more critical of AI output” and expect the problem to go away. The gravitational pull of System 3 is stronger than that. A 2025 <a href="https://arxiv.org/abs/2509.08010">position paper</a> co-authored by researchers from Harvard, Oxford, Cambridge, Stanford, and OpenAI reaches the same conclusion: overreliance on AI requires <em>structural</em> mitigation, not individual vigilance. The problem is architectural, not motivational.</p><h3>The part that worries me<a href="https://maroffo.github.io/blog/posts/2026-03-23-from-cognitive-load-to-cognitive-surrender/#the-part-that-worries-me">#</a></h3><p>.mau. makes a fair point: we’ve been complaining about cognitive decline for 2,500 years (Plato’s Theuth myth, the worry that writing would destroy memory). And we’ve always tended to defer to authority.</p><p>But the paper surfaces something specific that makes this different from a book or a calculator. System 3 doesn’t store information or perform computations. It produces fluent, confident, contextual responses that pattern-match what a knowledgeable human would say. And the part that sits with me: participants who followed wrong AI answers didn’t just get the answer wrong. They reported <em>higher confidence</em> in those wrong answers than participants who got them wrong on their own. The AI didn’t just mislead them. It made them more certain of the mistake.</p><p>A calculator gives you a number and you decide what to do with it. An LLM gives you a reasoned argument and you have to decide whether to <em>re-reason</em> the whole thing from scratch. Most of the time, under most conditions, people don’t. They adopt the answer <em>and</em> the confidence that came with it.</p><p>Shaw and Nave’s Table 2 lists the cognitive routes under Tri-System Theory. The one I keep coming back to is “Autopilot”: stimulus goes directly to System 3, response comes back, the brain boundary is never crossed. Question in, AI answer out, submit. In their experiments (word problems that require reading), Autopilot is a theoretical extreme, because subjects must at least parse the question. But in an IDE with inline autocomplete, Autopilot isn’t theoretical. It’s the default UI. Tab, tab, tab, commit.</p><p>If you think Autopilot only applies to boilerplate code, look at clinical medicine. The evidence is already here, and it’s peer-reviewed. A <a href="https://www.thelancet.com/journals/langas/article/PIIS2468-1253(25)00133-5/abstract">multicentre trial published in The Lancet</a> (2025) found that after just three months of AI-assisted polyp detection, clinicians’ unassisted adenoma detection rate fell from 28% to 22%, a 22% relative decline. Three months was all it took to measurably erode a diagnostic skill. When I read that, I immediately thought about my own ability to spot an N+1 query in a PR review, or catch a subtle race condition in concurrent code. I’ve been tab-completing through production code for over a year now. How much of that diagnostic instinct is still mine?</p><p>I still use AI tools every day. The countermeasures from the original post haven’t changed. What’s changed is my confidence that those countermeasures are enough for anyone besides me.</p><p>If you manage a team, the offloading/surrender distinction should be part of how you think about AI adoption. Not “are people using AI?” but “are people still thinking when they use AI?” The answer, according to this paper, is: less than you hope, and incentives alone won’t fix it.</p><ul><li>Shaw, S. D., &amp; Nave, G. (2026). <a href="https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6097646">Thinking, Fast, Slow, and Artificial: How AI is Reshaping Human Reasoning and the Rise of Cognitive Surrender</a>. Working paper, Wharton School.</li><li>Kosmyna, N. et al. (2025). <a href="https://www.media.mit.edu/publications/your-brain-on-chatgpt/">Your Brain on ChatGPT: Accumulation of Cognitive Debt</a>. MIT Media Lab.</li><li>Ibrahim, R. et al. (2025). <a href="https://arxiv.org/abs/2509.08010">Measuring and Mitigating Overreliance is Necessary for Building Human-Compatible AI</a>. arXiv.</li><li>Budzyński, J. et al. (2025). <a href="https://www.thelancet.com/journals/langas/article/PIIS2468-1253(25)00133-5/abstract">Endoscopist deskilling after AI-assisted polyp detection</a>. The Lancet Gastroenterology &amp; Hepatology.</li><li><a href="https://www.microsoft.com/en-us/research/wp-content/uploads/2025/12/New-Future-Of-Work-Report-2025.pdf">Microsoft New Future of Work Report 2025</a>. Microsoft Research.</li><li>.mau. (2026). <a href="https://xmau.com/wp/notiziole/2026/03/02/resa-cognitiva/">Resa cognitiva</a>. Notiziole di .mau.</li></ul><h3>Methodology note<a href="https://maroffo.github.io/blog/posts/2026-03-23-from-cognitive-load-to-cognitive-surrender/#methodology-note">#</a></h3><p>This post was written with AI assistance (Claude Code for research synthesis and drafting, with manual editing throughout). The paper it discusses would classify this workflow as cognitive offloading, not surrender, provided I actually verified what the AI wrote. I did. Mostly.</p><p><em>Originally published at </em><a href="https://maroffo.github.io/blog/posts/2026-03-23-from-cognitive-load-to-cognitive-surrender/"><em>https://maroffo.github.io</em></a><em> on March 23, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9b10da9a91dd" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>