https://bertobi.github.io/blog/Recent content on Tobias's BlogHugoen-usSat, 13 Jun 2026 00:00:00 +0000https://bertobi.github.io/blog/posts/2026-06-13-arena-hackathon-nlas-all-the-way-down/Sat, 13 Jun 2026 00:00:00 +0000https://bertobi.github.io/blog/posts/2026-06-13-arena-hackathon-nlas-all-the-way-down/<h1 id="nlas-all-the-way-down-a-verbalizers-self-readout-doesnt-reveal-when-its-making-things-up">NLAs all the way down: a verbalizer&rsquo;s self-readout doesn&rsquo;t reveal when it&rsquo;s making things up</h1> <p><em>ARENA hackathon, June 2026.</em></p> <h2 id="tldr">TL;DR</h2> <p>A <a href="https://transformer-circuits.pub/2026/nla/index.html">natural language autoencoder (NLA)</a> is a tool that translates a model&rsquo;s internal activations into English. We (Me and Claude), used the NLA to explain its own internal state. When the verbalizer made up a fact, we read its own activations and passed them through the same NLA to see whether the readout would give signs of it being a fabrication. It didn&rsquo;t. A confabulating state and a faithful one produce the same confident description, with no hedging and no sign that anything was invented. Our reading: confabulation is confident, the model isn&rsquo;t internally flagging doubt, so a self-readout has nothing to surface.</p>https://bertobi.github.io/blog/about/Mon, 01 Jan 0001 00:00:00 +0000https://bertobi.github.io/blog/about/<p>Hi, I&rsquo;m <strong>Tobias</strong> — a junior AI Safety researcher based in Buenos Aires.</p> <p>I started this blog to share my work in AI Safety and ocasional showcases of personal projects.</p> <p>You can find me on <a href="https://github.com/BerTobi">GitHub</a>. Say hello in the comments on any post.</p>