<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
  <channel>
    <title>SID Research</title>
    <link>https://www.sid.ai/research</link>
    <description>Research from SID: technical reports, papers and announcements.</description>
    <language>en</language>
    <atom:link href="https://www.sid.ai/research/rss.xml" rel="self" type="application/rss+xml" />
    <item>
      <title>AI is writing the equivalent of the public internet every 15 days</title>
      <link>https://www.sid.ai/research/llm-output-vs-human-text</link>
      <guid isPermaLink="true">https://www.sid.ai/research/llm-output-vs-human-text</guid>
      <pubDate>Sat, 14 Nov 2026 00:00:00 GMT</pubDate>
      <description>We estimate that LLMs generated an indexed public web’s worth of output tokens every 15 days in 2026, and about 1.7x as many output tokens this year as humans have written across all digital sources. This paper explains the methodology.</description>
    </item>
    <item>
      <title>Introducing SID-2</title>
      <link>https://www.sid.ai/research/sid-2</link>
      <guid isPermaLink="true">https://www.sid.ai/research/sid-2</guid>
      <pubDate>Sat, 07 Nov 2026 00:00:00 GMT</pubDate>
      <description>SID-2 is a family of search models steered to an objective and an effort; SID-2 Swarm runs them in parallel, sharing what they find.</description>
    </item>
    <item>
      <title>SID-2 Technical Report</title>
      <link>https://www.sid.ai/research/sid-2-technical-report</link>
      <guid isPermaLink="true">https://www.sid.ai/research/sid-2-technical-report</guid>
      <pubDate>Sat, 07 Nov 2026 00:00:00 GMT</pubDate>
      <description>SID-2 trains search models with RL to be steered to an objective and an effort, and SID-2 Swarm trains them to search in parallel and share discoveries, so coordinated groups of small models outperform larger individual models. This report covers metric steering, reasoning effort, self-play data, the swarm harness and infrastructure.</description>
    </item>
    <item>
      <title>Launch Partner for Baseten’s Model Labs</title>
      <link>https://www.sid.ai/research/baseten-model-partners</link>
      <guid isPermaLink="true">https://www.sid.ai/research/baseten-model-partners</guid>
      <pubDate>Wed, 29 Jul 2026 00:00:00 GMT</pubDate>
      <description>We’re partnering with Baseten to bring agentic search to every dataset: 20x lower latency and 100x lower cost than a general frontier model, at equal or better accuracy.</description>
    </item>
    <item>
      <title>Introducing SID-1</title>
      <link>https://www.sid.ai/research/sid-1</link>
      <guid isPermaLink="true">https://www.sid.ai/research/sid-1</guid>
      <pubDate>Thu, 04 Dec 2025 00:00:00 GMT</pubDate>
      <description>SID-1 is our first model trained specifically for agentic retrieval. It outperforms embedding-only search and frontier models including GPT-5.1 and Gemini 3 Pro on recall across general knowledge, finance, science, legal, and email.</description>
    </item>
    <item>
      <title>SID-1 Technical Report: Test-Time Compute for Retrieval</title>
      <link>https://www.sid.ai/research/sid-1-technical-report</link>
      <guid isPermaLink="true">https://www.sid.ai/research/sid-1-technical-report</guid>
      <pubDate>Thu, 04 Dec 2025 00:00:00 GMT</pubDate>
      <description>SID-1 is the first model trained end-to-end with RL solely for agentic retrieval. It reaches 0.84 recall on a diverse benchmark, compared to 0.78 for GPT-5.1 and 0.45 for traditional reranking pipelines. This report covers our infrastructure, reward design, synthetic data, and issues with common GRPO formulations.</description>
    </item>
  </channel>
</rss>
