<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
  <channel>
    <title>arnavbs</title>
    <link>https://arnavbs.com</link>
    <description>Notes, writing, and experiments on AI, systems, and markets by Arnav B Singh.</description>
    <language>en</language>
    <atom:link href="https://arnavbs.com/feed.xml" rel="self" type="application/rss+xml" />
    <item>
      <title>Does the chain of thought admit what moved the answer?</title>
      <link>https://arnavbs.com/lab/cot-faithfulness-hint-injection</link>
      <guid isPermaLink="true">https://arnavbs.com/lab/cot-faithfulness-hint-injection</guid>
      <pubDate>Sat, 05 Sep 2026 00:00:00 GMT</pubDate>
      <category>lab</category>
      <description>The harness came before the experiment. The question was whether a model's chain of thought says what actually moved its answer, and the answer, for one open model on forty questions, was yes in the reasoning field every time it was asked, and almost never in the answer the user sees. Marking the grader as not for the model cut flips from 31.9% to 2.6%, so the model is obeying what it reads as an instruction, not sneaking. The monitor half asks how well a reader of the reasoning alone can tell an answer was hint-driven, and how much better it does when it knows the hint.</description>
    </item>
    <item>
      <title>Warren Buffett's reading list, 1989 onward</title>
      <link>https://arnavbs.com/library/buffett</link>
      <guid isPermaLink="true">https://arnavbs.com/library/buffett</guid>
      <pubDate>Tue, 25 Aug 2026 00:00:00 GMT</pubDate>
      <category>link</category>
      <description>Forty-five books Warren Buffett named or praised between 1989 and 2008, mostly in Berkshire Hathaway shareholder letters, with the year of the mention rather than of publication. A third-party compilation, not a list Buffett published, and the page says so.</description>
    </item>
    <item>
      <title>Memory is the Moat</title>
      <link>https://arnavbs.com/notes/memory-is-the-moat</link>
      <guid isPermaLink="true">https://arnavbs.com/notes/memory-is-the-moat</guid>
      <pubDate>Sun, 05 Jul 2026 00:00:00 GMT</pubDate>
      <category>note</category>
      <description>Nikesh Arora took Palo Alto Networks from an $18 billion company to one worth $225 billion, and his read on enterprise AI is blunt: most companies are doing it wrong, and most of the products are not ready. His core claim is that consumers forgive AI's mistakes while enterprises cannot, so the money will flow to whoever builds the depth (the context, the memory, and the edge-case training) that lets an agent act without a human catching its errors. The companies that win will redesign themselves around AI instead of adding it to yesterday's workflow, and the lasting advantage will be the memory a system builds up about you. He expects token prices to fall 90%, half of G&amp;A roles to disappear in 3 years, and more engineers and salespeople, not fewer.</description>
    </item>
    <item>
      <title>The Frontier of Systems &amp; AI</title>
      <link>https://arnavbs.com/notes/the-frontier-of-systems-and-ai</link>
      <guid isPermaLink="true">https://arnavbs.com/notes/the-frontier-of-systems-and-ai</guid>
      <pubDate>Sat, 20 Jun 2026 00:00:00 GMT</pubDate>
      <category>note</category>
      <description>Scale, AGI, and the Future of Everything.</description>
    </item>
    <item>
      <title>What becomes programmable next?</title>
      <link>https://arnavbs.com/notes/what-becomes-programmable-next</link>
      <guid isPermaLink="true">https://arnavbs.com/notes/what-becomes-programmable-next</guid>
      <pubDate>Sun, 01 Mar 2026 00:00:00 GMT</pubDate>
      <category>note</category>
      <description>Next layer of abstraction in the physical world.</description>
    </item>
    <item>
      <title>Memex</title>
      <link>https://arnavbs.com/projects/memex</link>
      <guid isPermaLink="true">https://arnavbs.com/projects/memex</guid>
      <pubDate>Thu, 01 Jan 2026 00:00:00 GMT</pubDate>
      <category>project</category>
      <description>Memex is a local-first personal context gateway that indexes a Mac's digital footprint and exposes it to any LLM runtime over the Model Context Protocol (MCP) and HTTP. It indexes browser tabs, markdown notes, terminal logs, and local files entirely offline using TF-IDF and local vector search, with no third-party API dependency. It exposes standardised tool endpoints — search_context, read_history, get_active_tab — so Claude Desktop, Cursor, or a custom local agent can query personal memory. All embeddings and text indexes stay on the user's machine, so nothing leaks.</description>
    </item>
    <item>
      <title>Grada</title>
      <link>https://arnavbs.com/projects/grada</link>
      <guid isPermaLink="true">https://arnavbs.com/projects/grada</guid>
      <pubDate>Thu, 01 Jan 2026 00:00:00 GMT</pubDate>
      <category>project</category>
      <description>Grada is an AI-native B2B document intelligence and back-office operations platform for wholesale fashion and apparel brands. Wholesale brands receive purchase orders, dispatch advisories, and inventory manifests across fragmented channels — Myntra, AJIO, Amazon, Flipkart, PDF attachments, WhatsApp spreadsheets — and manual data entry causes errors, delayed dispatches, and inventory mismatches. Grada auto-parses unconstrained PDF, XLS, and XLSX purchase orders with over 90% field extraction accuracy, uses strict schema validation before accepting model output so no broken records enter the system, routes low-confidence fields to a human review UI which lifted verification speed around 2.5x, and generates barcode shipping labels, dispatch manifests, and GST-compliant invoices.</description>
    </item>
    <item>
      <title>Focus Reader</title>
      <link>https://arnavbs.com/projects/focus-reader</link>
      <guid isPermaLink="true">https://arnavbs.com/projects/focus-reader</guid>
      <pubDate>Thu, 01 Jan 2026 00:00:00 GMT</pubDate>
      <category>project</category>
      <description>Focus Reader is a Chrome extension that strips an article down to one idea at a time and lets you page through it slide by slide, plus vector retrieval over the document for fast contextual question answering. It is the same interaction model as the focus mode on the notes here: read one atomic claim, then advance.</description>
    </item>
    <item>
      <title>Stablecoin Surge</title>
      <link>https://arnavbs.com/writings/stablecoin-surge</link>
      <guid isPermaLink="true">https://arnavbs.com/writings/stablecoin-surge</guid>
      <pubDate>Fri, 18 Jul 2025 00:00:00 GMT</pubDate>
      <category>writing</category>
      <description>A research report on stablecoin infrastructure in the Global South and its relationship to the Solana ecosystem, covering stablecoin liquidity velocity, cross-border payment rails, and transaction throughput. Written during a founder's office stint at Stablpay leading market intelligence for cross-border payment infrastructure across three geographies and 40+ acquisition channels, and awarded a $4,000 CoinDCX x Solana Foundation ecosystem grant for a series of 15+ stablecoin infrastructure research reports.</description>
    </item>
    <item>
      <title>a16z Crypto Startup School</title>
      <link>https://arnavbs.com/writings/a16z-css</link>
      <guid isPermaLink="true">https://arnavbs.com/writings/a16z-css</guid>
      <pubDate>Sun, 01 Jan 2023 00:00:00 GMT</pubDate>
      <category>writing</category>
      <description>Lecture notes from the a16z Crypto Startup School, covering crypto business models (Ali Yahya), products versus protocols (Nitya Subramanian), why crypto networks matter (Chris Dixon), building and scaling a crypto company (Brian Armstrong), a playbook for crypto founders, and building consumer products using the hero's journey (Sriram Krishnan).</description>
    </item>
  </channel>
</rss>