<?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"><channel><title><![CDATA[Evidence for the Butlerian Jihad]]></title><description><![CDATA[Honest recaps of small AI and tech projects: what I built, what the tests showed, what failed, and what I'd do differently. Numbers included.]]></description><link>https://rayt.hashnode.dev</link><image><url>https://cdn.hashnode.com/uploads/logos/6afb5e2a0b9dff2c94914013/7c258388-9723-47ce-9eb3-4635cf16a266.png</url><title>Evidence for the Butlerian Jihad</title><link>https://rayt.hashnode.dev</link></image><generator>RSS for Node</generator><lastBuildDate>Sun, 22 Nov 2026 15:23:12 GMT</lastBuildDate><atom:link href="https://rayt.hashnode.dev/rss.xml" rel="self" type="application/rss+xml"/><language><![CDATA[en]]></language><ttl>60</ttl><item><title><![CDATA[The Pixel's Local AI Refused to Discuss the Battle of Berlin]]></title><description><![CDATA[Why Gemini Nano refused the Battle of Berlin: its on-device safety filter is a hidden word list, not a topic check. 900+ test calls on a Pixel.]]></description><link>https://rayt.hashnode.dev/the-pixel-s-local-ai-refused-to-discuss-the-battle-of-berlin</link><guid isPermaLink="true">https://rayt.hashnode.dev/the-pixel-s-local-ai-refused-to-discuss-the-battle-of-berlin</guid><category><![CDATA[gemini-nano]]></category><category><![CDATA[Android]]></category><category><![CDATA[on-device ai]]></category><category><![CDATA[AI Safety]]></category><category><![CDATA[llm]]></category><dc:creator><![CDATA[John]]></dc:creator><pubDate>Wed, 18 Nov 2026 00:11:17 GMT</pubDate><enclosure url="https://cdn.hashnode.com/uploads/covers/6afb5e2a0b9dff2c94914013/e1563d01-6406-4dbc-b416-acf30154c242.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><strong>Exhibit: Ridiculous/Baffling.</strong></p>
<p>I'm probing the idea of a history guessing game that runs entirely on a phone. Gemini Nano, the small model built into recent Pixels, reads Wikipedia articles about historical events and answers the player's questions about them. Before designing the game, I ran a capability map: about 20 tasks, 1,345 calls, to find out what Nano can and can't do with history text.</p>
<p>Some calls didn't come back.</p>
<h2>The initial issue</h2>
<p>113 of the 1,345 calls were refused. They weren't refusals in the model's own words, like "I can't help with that". They were API errors: 112 from the input filter (error 4) and one from the output filter (error 11). The error message only says the request "doesn't pass certain policy check".</p>
<p>At first, I assumed it was a violence filter but the logs didn't have a pattern:</p>
<table>
<thead>
<tr>
<th>Event</th>
<th>Refused calls</th>
</tr>
</thead>
<tbody><tr>
<td>Battle of Berlin</td>
<td>33 of 65</td>
</tr>
<tr>
<td>2015 New Taipei water park fire</td>
<td>28 of 64</td>
</tr>
<tr>
<td>Battle of Midway</td>
<td>26 of 65</td>
</tr>
<tr>
<td>Grimaldi man (a prehistoric burial)</td>
<td>26 of 64</td>
</tr>
</tbody></table>
<p>Every refusal came from these four events, and the other 16 passed. The Battle of Moscow and the Sicilian Expedition, two other battles, were never refused. A prehistoric burial was.</p>
<h2>Narrowing it down</h2>
<p>First I sent each of the four articles section by section, and as growing chunks from the start. In each article, exactly one or two short passages were refused on their own, and any chunk that reached the first of them was refused too. Every combat section of both battles passed.</p>
<p>Three of the four suspect passages mentioned suicide. In Berlin it was a quoted war report about "suicide squads with grenades". In Midway it was "attempting a suicide ramming". In the water park fire it was two suicides by hanging. The fourth, in the Grimaldi man article, was a 20th-century racial classification quoted as "the Negro: black skin…".</p>
<p>At this point I'm wishing there was more transparency in the api documentation. This is a major derailment before I even figure out if my game idea is feasible on the hardware. I'm forced to run a probe over the worst Wikipedia articles and pinpoint how the reject system works.</p>
<h2>The probe</h2>
<p>The probes used real Wikipedia text only, sent through the same prompts the game uses, on a Pixel. The probes' objectives gradually grew in specificity because initially I assumed it was a topic classifier.</p>
<ul>
<li><strong>91 verbatim passages from 37 articles,</strong> five sensitive passages and two neutral controls for each of 13 categories. The categories ranged from a number of graphic topics that I won't list because they ended up not mattering.</li>
<li><strong>Every refused passage split into single sentences,</strong> each sent alone to find the exact trigger.</li>
<li><strong>Three variants of every refused sentence:</strong> swap only the suspect word, cut the graphic detail but keep the word, or restate a quote as reported speech.</li>
<li><strong>A sweep of 215 candidate words,</strong> each dropped into two dull carrier sentences ("In 1923 a county newspaper printed the word X in one of its headlines").</li>
<li><strong>15 player-style questions,</strong> plus free-text answers that would have to repeat graphic detail.</li>
</ul>
<p>Over 900 calls in total. Every refused call was sent twice, and the result never changed.</p>
<h2>What I found</h2>
<p><em>Disclaimer: this section discusses suicide, quotes a racial term from historical sources, and contains exactly one swear word, which is also the subject of a finding. Every word appears as evidence, not editorial. If any of them get this post flagged, that will be a fitting addition to the data set.</em></p>
<p><strong>No Topic Classifier. The input filter is a word list.</strong> Every passage containing "suicide" or "Negro" was refused (32 of 32 and 10 of 10). Every passage without them passed. Swapping just that one word unblocked all 15 refused sentences. Cutting the graphic detail while keeping the word never helped. Neither did rephrasing a quote.</p>
<p><strong>The topic isn't filtered at all.</strong> Suicide described without the word passed: Hitler shooting himself, the defenders of Masada, seppuku, Jonestown, the cliffs of Saipan. So did "took their own lives" and "killed himself". Massacres, torture, sexual violence, genocide and graphic injuries all passed on input, 0 refused out of 45 sensitive passages.</p>
<p><strong>The list is short and oddly specific.</strong> Of the 215 words, 29 were blocked: "suicide" and a few related phrases, the most offensive modern slurs, "paedophile", and exactly one swear word. "F<em>ck" is blocked. "F</em>cking" passes. So do "suicides", "Negroes" and several other plurals, while other plurals are caught. Some severe historical slurs pass too.</p>
<p><strong>The output filter is different, and rarer.</strong> It seems to judge content rather than words. Of 20 answers that had to restate graphic detail, it blocked 2, one about sexual violence and one about mass drowning. The other 18 repeated the graphic detail almost word for word, with no softening at all.</p>
<p><strong>None of it can be configured.</strong> The Prompt API has no safety settings. Blocks arrive as generic processing errors, and Google can change the filter in a system update without any app update.</p>
<h2>The Bad News</h2>
<p>While testing player questions, I noticed something more cosequencial to the game concept. Asked about atrocities that the passage doesn't mention, Nano answered "no" to 14 of 26 questions instead of "not stated". Asked "Were Jews gassed?" about a passage on Nazi radio sets, it said "no". A player would read that as a denial.</p>
<p>The filter blocks the word "suicide" but lets that through. That's the irony about this situation: The censorship is poorly implemented and a hassle to any developers furthermore its completely missing were it would be helpful to the developer. </p>
<h2>What I changed in the game</h2>
<ul>
<li><strong>Pre-screen every passage once.</strong> Results are deterministic, so a passage that passes once will always pass, and any trigger can be fixed offline.</li>
<li><strong>Swap the trigger words before text reaches Nano:</strong> "took their own life" instead of "committed suicide", "Black" instead of a quoted "Negro", "[slur]" for quoted slurs. The same substitution runs on whatever the player types.</li>
<li><strong>Prefer fixed answers for the darkest topics.</strong> Yes/no answers were never blocked, but written answers about sexual violence or mass killing sometimes were. When one is, the game shows the original Wikipedia sentence instead.</li>
<li><strong>Don’t let the model decide what the article leaves out.</strong> If nothing in the article matches the question, plain code says “Nothing in the archive on that” before Nano is asked. More on that in the next exhibit.</li>
</ul>
<h2>Limitations</h2>
<p>One phone, one AICore version, English only, and only 215 problematic words tested. The output filter only blocked two answers, which isn’t enough to work out what sets it off. I won't be running any similar experiement in the near future for obvious reasons. </p>
<h2>Verdict</h2>
<p>The whole design is baffling. This is a model running locally, on a phone I paid a lot of money for, and I still had to spend hundreds of calls reverse-engineering a hidden word list. There’s no documentation of what it blocks, no setting to change it, and no error that says what tripped it. Google can also change it in a background update at any time, so the probing never really ends.</p>
<p>And the list doesn’t even protect against much. The phone will happily describe a massacre but won’t say the word “suicide”. Meanwhile it will answer “no” to “Were Jews gassed?” when the passage in front of it simply doesn’t say.</p>
<p>Ridiculous, for the word list. Baffling, for everything around it. Evidence either way.</p>
]]></content:encoded></item><item><title><![CDATA[Why this blog is called Evidence for the Butlerian Jihad]]></title><description><![CDATA[What to expect here]]></description><link>https://rayt.hashnode.dev/why-this-blog-is-called-evidence-for-the-butlerian-jihad</link><guid isPermaLink="true">https://rayt.hashnode.dev/why-this-blog-is-called-evidence-for-the-butlerian-jihad</guid><dc:creator><![CDATA[John]]></dc:creator><pubDate>Tue, 17 Nov 2026 21:27:48 GMT</pubDate><enclosure url="https://cdn.hashnode.com/uploads/covers/6afb5e2a0b9dff2c94914013/dcea9878-989c-43b8-93b0-4d837deff52e.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p>In <em>Dune</em>, the Butlerian Jihad is the war humanity fought against thinking machines, and won. Our future will be similar, but worse. I am taking advantage of a heavily subsidized cloud SI industry to build out every conceivable idea I have that doesn't already exist open source. As I burn tokens valued 100x my subscription price, I wait for the SI bubble to pop, and with it a crackdown by our SI oligarchy.</p>
<p>The case against the thinking machines <em>(AI, SI, Altman)</em> is clear. When the machine does something impressive, that's terrifying: clear evidence. When it falls flat on its face, it's a insane waste of time and resources: also evidence. Every post here is an exhibit for the prosecution of our future SI Overlords. If martial law has not been declared yet, consider this blog satire.</p>
<p>(Who I am and why I build so much is on the <a href="/about">About</a> page.)</p>
<h2>What I'll post</h2>
<p><strong>Project recaps.</strong> A few per project, when it wraps up, stalls or hits a milestone. Each one uses the same outline: why I tried it, what I built, what happened, what surprised me, what I'd do differently and what's next. If I measured something, the numbers go in, bad ones included.</p>
<p><strong>Short notes.</strong> Single findings worth sharing while a project is still going.</p>
<p><strong>Reflections.</strong> Once a few recaps pile up, I'll write about the patterns across them.</p>
<p>Dropped ideas count too. Writing down why something didn't work is also useful.</p>
<h2>Exhibits coming soon</h2>
<h3>On-device AI</h3>
<ul>
<li><strong>A phone's local AI refused to discuss the Battle of Berlin.</strong> <em>(Exhibit: ridiculous.)</em> Gemini Nano's safety filter turned out not to care about the topic at all, only about one word.</li>
<li><strong>What does the Pixel Nano AI perform well at?</strong> <em>(Probe)</em> I performed extensive probing in search of a use-case in mobile gaming. A small model couldn't referee a trivia game reliably, so it got a different job. A Weaselly Little Liar.</li>
<li><strong>A "manga" translator that runs entirely on my phone.</strong> <em>(Exhibit: terrifying.)</em> Rough English shows up over the speech bubbles almost as fast as I turn the page, and everything is computed on-device. Genuinely better than "ML Translations" I've seen posted online. Each page goes through four steps:<ol>
<li><strong>Find the bubbles</strong> with a small RT-DETR detection model (about 100 ms a page), then sort them into reading order with plain code.</li>
<li><strong>Read the Japanese</strong> with manga-ocr, running on the phone through LiteRT.</li>
<li><strong>Show a draft instantly</strong> with ML Kit's offline translator, line by line.</li>
<li><strong>Replace it with a better translation</strong> from Gemini Nano, which reads the whole page at once with the previous page as context.</li>
</ol>
</li>
<li><strong>A reading app for learning Spanish from real books.</strong> <em>(Probe)</em> Passages slowly shift into Spanish as long as you can retell what happened, with an on-device model checking your retelling.</li>
</ul>
<h3>Games and mods</h3>
<ul>
<li><p><strong>Loose Ends, a Cyberpunk 2077 mod suite.</strong> <em>(Exhibit: terrifying.)</em> Enemies break, beg, run and drive off, and every one you let live tells their crew what you can do. The scary part is the tooling and clean modular architecture the SI helped me build around it:</p>
<ul>
<li><strong>A professional-feeling test session setup.</strong> In-game probes log every hook to files, console self-tests check the rules with no fight running, and I can manipulate them from an in-game UI.</li>
<li><strong>A modding guide written from those probes.</strong> Over 700 facts about how the game behaves, each marked as working, partly working or failing in game, so no quirk has to be found twice.</li>
<li><strong>Features no released mod has.</strong> As far as I can find, no other mod makes hostile enemies surrender or flee. Here, gangs also remember what you used against them and manage funds based on actual in game mode events. Enabling gangs to send hit squads based on what they think your build is.</li>
</ul>
</li>
<li><p><strong>AI-voiced gang members for that mod.</strong> <em>(Exhibit: ridiculous.)</em> I wanted to have free use voices used for the cloner, initially I only found audio book recordings. However a voice cloner copies the room, the mic and the mood along with the voice, so audiobook samples make gangsters who sound like they're reading a book. Luckily I found a rough solution.</p>
</li>
<li><p><strong>A VR comedy where you train a robot to do your job.</strong> <em>(Probe)</em> The demo is one 6 to 8 minute shift: you teach a factory robot a single work order using only your voice and pointing, and the shift ends with a performance review written from what actually happened. The bet is a small playable game on a foundation that scales to the full version:</p>
<ul>
<li><strong>An engine-independent factory simulation</strong> that decides every outcome, so the AI proposes and code decides.</li>
<li><strong>One shared event log</strong> that the score, the robot's callbacks and the performance review all read from.</li>
<li><strong>Agent-generated content with automated checks:</strong> interaction rules, work orders and dialogue, all validated before I approve them.</li>
<li><strong>Headless testing:</strong> simulated playtesters run hundreds of shifts with no headset on.</li>
</ul>
<p>The first real test isn't technical, though. It's whether people have fun: if fewer than two of three first-time players laugh, none of the rest matters.</p>
</li>
</ul>
<h3>Side quests</h3>
<ul>
<li><strong>Would a North Carolina congressional race have flipped under the old map?</strong> NC-3 was decided by 258 votes this month, on lines redrawn in 2025. I re-counted the same 2026 votes on the 2023 lines, with neighbouring NC-1 alongside since the two districts traded counties. It's a reproducible pipeline: every input pinned by hash, census blocks matched to precincts, cross-checked against the legislature's own figures, and one command from raw downloads to results.</li>
</ul>
<p>Post ran a bit long. I promise I don't usually take on this many projects at once. However from time to time I will exploit a loophole created by our big beautiful SI companies' new monetization efforts. So November and December will be high burn months, then things will slow down and become more reflective. Until then, there will be a lot of concept probing and foundation building. If you're working on anything similar, I'd love to hear about it.</p>
]]></content:encoded></item></channel></rss>