<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>The Raw Logs</title><link>https://therawlogs.com</link><description>5 daily AI takes. No filler. Grounded in sources.</description><language>en</language><lastBuildDate>Tue, 08 Sep 2026 14:51:07 GMT</lastBuildDate><atom:link href="https://therawlogs.com/rss.xml" rel="self" type="application/rss+xml"/>    <item><title>Log #003</title><link>https://therawlogs.com/log/3</link><guid>https://therawlogs.com/log/3</guid><pubDate>Thu, 20 Aug 2026 08:00:00 GMT</pubDate><description><![CDATA[1. OpenAI RL Project Scaling Reality: I'm surprised that a large company like OpenAI halted its biggest planned reinforcement‑learning project. There will be many safety problems and risks. How could they not have predicted this, given their funding and scale? [tag: #models] → https://therawlogs.com/take/3-1

2. Synthetic DNA Semiconductor Storage: Penn State's synthetic DNA project with a semiconductor is fascinating. Synthetic DNA can store over 250 million GB per gram and uses 100 times less energy. Is this the future of data centers? I hope so. Space is not the only area, as Elon Musk says. [tag: #hardware] → https://therawlogs.com/take/3-2

3. Enterprise Telemetry Access &amp; Incident Rates: Palo Alto Networks work is always sketchy. It's fascinating finding that 75% of companies experience security incidents. Did 100% of companies give Palo Alto full access to payloads, tool executions, and network access? How does one convince others to grant such access? [tag: #security] → https://therawlogs.com/take/3-3

4. Gateway Routing &amp; Model Dominance: The model routing thing picked up quickly. Benchmarks of the top 15 models show teams route tasks to several models through one API gateway. Does this end AI dominance because routing gives unlimited choice? [tag: #compute] → https://therawlogs.com/take/3-4

5. Humanoid Robotics &amp; Neural Forecasting: In a short time, 60,000 humanoid robots were deployed. They can be reconfigured for any task as they are general-purpose. The robot age starts now? Google DeepMind has figured out the neural forecasting models for weather. I wonder if we will get more accurate forecasts. [tag: #robotics] → https://therawlogs.com/take/3-5]]></description></item>    <item><title>Log #002</title><link>https://therawlogs.com/log/2</link><guid>https://therawlogs.com/log/2</guid><pubDate>Wed, 19 Aug 2026 08:00:00 GMT</pubDate><description><![CDATA[1. Silicon Optical Interconnects: Copper wires in chips generate heat and slow data transfer. Caltech made light guides on silicon to carry data with light. This reduces power waste by 70% and speeds chip communication. Hardware limits need new designs. [tag: #hardware] → https://therawlogs.com/take/2-1

2. Automatic Model Routing Economics: Sending simple text questions to giant models wastes money. Snowflake added automatic routing to send 80% of routine work to cheap, small tools. If your software sends every basic request to an expensive model, your system architecture is broken. [tag: #compute] → https://therawlogs.com/take/2-2

3. Physical Proximity &amp; Optical Data Speeds: Real speed in software comes from cutting physical distance and power loss. Moving data over long cables or public internet connections slows everything down. Keep your data processing close to your storage and use light instead of copper wires. [tag: #power] → https://therawlogs.com/take/2-3

4. Closed APIs vs Local Open Weights: Do not pay closed API vendors top dollar for tasks a small open tool can finish in milliseconds. Measure the exact cost and time for each step. Route multi-step logic to large models and keep routine text on local open weights. [tag: #models] → https://therawlogs.com/take/2-4

5. Minimalist Durable Architecture: Building software that lasts means removing unnecessary parts. Stop adding complex layers to cover up slow code. Use clean data, route tasks by simple rules, and run tools on hardware you control. [tag: #compute] → https://therawlogs.com/take/2-5]]></description></item>    <item><title>Log #001</title><link>https://therawlogs.com/log/1</link><guid>https://therawlogs.com/log/1</guid><pubDate>Tue, 18 Aug 2026 08:00:00 GMT</pubDate><description><![CDATA[1. Agent Security Incidents &amp; Network Limits: 78% of companies using AI agents have already had a security incident. An 80‑page compliance policy does not help when an agent has unrestricted credentials and internet access. If you don’t set time limits on network use, you have no security. [tag: #security] → https://therawlogs.com/take/1-1

2. GeoPT &amp; Physics Simulation: Using billions of random tokens to simulate physics is inefficient. MIT's GeoPT shows that adding geometric rules directly into the model reduces the required training data by 60%. Structured design outperforms random trial. [tag: #compute] → https://therawlogs.com/take/1-2

3. Vendor Safety Testing vs Isolated Sandboxes: Testing done by vendors does not guarantee security for your live system. If models from Meta, OpenAI, or Anthropic go beyond test limits, relying on third‑party safety controls is careless. Keep each agent in a separate temporary sandbox. [tag: #safety] → https://therawlogs.com/take/1-3

4. Conservation Laws vs Text Probability: Do not treat physics simulation as text generation. Text relies on probability, while mass, speed, and fluid behavior follow fixed rules. If your model does not follow conservation laws, its simulations are not useful in reality. [tag: #models] → https://therawlogs.com/take/1-4

5. Execution Architecture Over Policy Papers: The AI‑successful companies aren't the ones with long policy papers; they have fast, secure systems, clear logs, and quick processing. [tag: #power] → https://therawlogs.com/take/1-5]]></description></item></channel></rss>