Tokenmaxxing

<p>Two days ago, I burnt 250 million tokens in a single day.</p> <p>That&rsquo;s up 20x in six weeks. This idea, called tokenmaxxing, is the deliberate practice of maximizing token consumption. The question : how much electricity can we turn into useful work?</p> <p>The secret is parallelization. Structure a plan at the start of the day that allows multiple agents to work simultaneously. <a href="https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/">METR research</a> shows the latest models can now work autonomously for 12 hours, up from 1 hour a year ago. Here&rsquo;s the ramp once I started implementing a <a href="https://tomtunguz.com/filling-the-queue-for-ai/">daily plan</a> :</p> <p><a href="https://res.cloudinary.com/dzawgnnlr/image/upload/q_auto/f_auto/tokenmaxxing_ggplot.png"><!--[if mso | IE]> <table role="presentation" border="0" cellpadding="0" cellspacing="0" width="560" align="center" style="margin:0 auto;"><tr><td style="padding:0;"> <![endif]--> <img alt="The Path to 250M Tokens per Day" src="https://res.cloudinary.com/dzawgnnlr/image/upload/w_840,c_limit,f_auto,q_auto/tokenmaxxin.png" style="display: block; width: 100%; height: auto; border: 0; text-decoration: none;" width="560" /> <!--[if mso | IE]> </td></tr></table> <![endif]--></a></p> <p>So, what did I do two days ago? Here&rsquo;s one example. I prepared a presentation for the <a href="https://luma.com/1xn4wqtm">AI Engineers Tech Talk</a> on the infrastructure for building with agents that I&rsquo;m delivering tonight.</p> <p>One agent pulled git commit history from the code repository &amp; generated a lines-of-code chart. Another queried the agent error logs &amp; built a time series of agent failures by root cause. A third fact-checked the METR research citations. A fourth built the presentation using a JavaScript library. A fifth critiqued the overall flow &amp; content. All of this happened in the background.</p> <p>This was just one of the parallel flows in a day