<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
<channel>
<title>Actual Intelligence Labs research</title>
<link>https://actualintelligencelabs.ai/research</link>
<atom:link href="https://actualintelligencelabs.ai/research/feed.xml" rel="self" type="application/rss+xml" />
<description>Field notes, reports, and explainers from the lab at Actual Intelligence Labs: AI agent reliability, evaluation, context engineering, and production 3D web.</description>
<language>en-us</language>
<lastBuildDate>Wed, 02 Sep 2026 12:00:00 GMT</lastBuildDate>
<item>
<title><![CDATA[98% of the Quality for 2.5% of the Frontier Tokens]]></title>
<link>https://actualintelligencelabs.ai/research/ai-agent-second-opinion</link>
<guid isPermaLink="true">https://actualintelligencelabs.ai/research/ai-agent-second-opinion</guid>
<pubDate>Sun, 30 Aug 2026 12:00:00 GMT</pubDate>
<category><![CDATA[Report]]></category>
<description><![CDATA[A cheaper model taking orders from a stronger one scored 96.8 against 98.8, using 2.5% of the frontier tokens. With a codebase map, it ties outright.]]></description>
</item>
<item>
<title><![CDATA[Over-Engineering, Measured: 120 True Issues in Three Days]]></title>
<link>https://actualintelligencelabs.ai/research/over-engineering</link>
<guid isPermaLink="true">https://actualintelligencelabs.ai/research/over-engineering</guid>
<pubDate>Wed, 19 Aug 2026 12:00:00 GMT</pubDate>
<category><![CDATA[Report]]></category>
<description><![CDATA[We audited our own AI agent harness for three days and got 120 issues. 80% of the code checked the checker. What over-engineering looks like, measured.]]></description>
</item>
<item>
<title><![CDATA[The AI Benchmark Number You Cannot Check]]></title>
<link>https://actualintelligencelabs.ai/research/ai-benchmarks</link>
<guid isPermaLink="true">https://actualintelligencelabs.ai/research/ai-benchmarks</guid>
<pubDate>Wed, 12 Aug 2026 12:00:00 GMT</pubDate>
<category><![CDATA[Report]]></category>
<description><![CDATA[AI benchmarks are graded by the people whose models get ranked. Two 2026 audits found broken answer keys and a judge that passed 62.81% of wrong answers.]]></description>
</item>
<item>
<title><![CDATA[Claude Code Starts at 29,000 Tokens]]></title>
<link>https://actualintelligencelabs.ai/research/claude-code-context-window</link>
<guid isPermaLink="true">https://actualintelligencelabs.ai/research/claude-code-context-window</guid>
<pubDate>Wed, 05 Aug 2026 12:00:00 GMT</pubDate>
<category><![CDATA[Field Note]]></category>
<description><![CDATA[A fresh Claude Code session spends 20,000 to 30,000 tokens before your first word. What sits in that block, what a healthy range looks like, and how to cut it.]]></description>
</item>
<item>
<title><![CDATA[We Went Looking for Context Rot at 700,000 Tokens]]></title>
<link>https://actualintelligencelabs.ai/research/context-rot</link>
<guid isPermaLink="true">https://actualintelligencelabs.ai/research/context-rot</guid>
<pubDate>Tue, 04 Aug 2026 12:00:00 GMT</pubDate>
<category><![CDATA[Report]]></category>
<description><![CDATA[Context rot is recall degrading as the window fills. We ran 128 trials on Claude Sonnet 5 out to 730,000 tokens. Accuracy fell 5.6 points. Data and caveats.]]></description>
</item>
<item>
<title><![CDATA[We Open Sourced the Harness. Then One Cost Question Gave Us Four Answers.]]></title>
<link>https://actualintelligencelabs.ai/research/open-source-llm-eval-harness</link>
<guid isPermaLink="true">https://actualintelligencelabs.ai/research/open-source-llm-eval-harness</guid>
<pubDate>Mon, 03 Aug 2026 12:00:00 GMT</pubDate>
<category><![CDATA[Field Note]]></category>
<description><![CDATA[We open sourced our LLM eval harness. One cost question, asked four defensible ways over 267 runs, gave four different answers. The data and the retraction.]]></description>
</item>
<item>
<title><![CDATA[Two AI Judges, One Document, Two Different Reviews]]></title>
<link>https://actualintelligencelabs.ai/research/llm-as-a-judge</link>
<guid isPermaLink="true">https://actualintelligencelabs.ai/research/llm-as-a-judge</guid>
<pubDate>Tue, 28 Jul 2026 12:00:00 GMT</pubDate>
<category><![CDATA[Report]]></category>
<description><![CDATA[LLM as a judge, measured on our own work: two AI reviewers on one document, 15 numeric claims against 6, and the four rules that finally made review stop.]]></description>
</item>
<item>
<title><![CDATA[We Ran Four Agentic Coding Tools Through the Same Gate]]></title>
<link>https://actualintelligencelabs.ai/research/agentic-ai-coding-tools</link>
<guid isPermaLink="true">https://actualintelligencelabs.ai/research/agentic-ai-coding-tools</guid>
<pubDate>Thu, 23 Jul 2026 12:00:00 GMT</pubDate>
<category><![CDATA[Field Note]]></category>
<description><![CDATA[We run Claude Code, Codex, Cursor, and a local open weight model on real client work. What separates agentic AI coding tools is the harness, not the model.]]></description>
</item>
<item>
<title><![CDATA[We Rebuilt 32 Pages Around One Question: Will an AI Cite This?]]></title>
<link>https://actualintelligencelabs.ai/research/generative-engine-optimization</link>
<guid isPermaLink="true">https://actualintelligencelabs.ai/research/generative-engine-optimization</guid>
<pubDate>Thu, 23 Jul 2026 12:00:00 GMT</pubDate>
<category><![CDATA[Field Note]]></category>
<description><![CDATA[Generative engine optimization is structuring content so AI answer engines quote it. The page structure we rebuilt 32 articles around, and how we measure it.]]></description>
</item>
<item>
<title><![CDATA[We Let an Agent Run 60 Experiments on Our RAG Stack While We Slept]]></title>
<link>https://actualintelligencelabs.ai/research/retrieval-token-optimization</link>
<guid isPermaLink="true">https://actualintelligencelabs.ai/research/retrieval-token-optimization</guid>
<pubDate>Tue, 21 Jul 2026 12:00:00 GMT</pubDate>
<category><![CDATA[Report]]></category>
<description><![CDATA[We let a worker-supervisor agent loop run 60 controlled experiments on our hybrid RAG retriever overnight. It cut context tokens 23% with no accuracy loss.]]></description>
</item>
<item>
<title><![CDATA[The Frontier You Can't See]]></title>
<link>https://actualintelligencelabs.ai/research/the-frontier-you-cant-see</link>
<guid isPermaLink="true">https://actualintelligencelabs.ai/research/the-frontier-you-cant-see</guid>
<pubDate>Mon, 20 Jul 2026 12:00:00 GMT</pubDate>
<category><![CDATA[Analysis]]></category>
<description><![CDATA[In summer 2026 Washington throttled America's best AI models and Anthropic's top tier sits partner-only. Why "China caught up" is only half a myth, sourced.]]></description>
</item>
<item>
<title><![CDATA[Does an LLM Understand Credible Punishment?]]></title>
<link>https://actualintelligencelabs.ai/research/grim-trigger-folk-theorem</link>
<guid isPermaLink="true">https://actualintelligencelabs.ai/research/grim-trigger-folk-theorem</guid>
<pubDate>Sat, 18 Jul 2026 12:00:00 GMT</pubDate>
<category><![CDATA[Report]]></category>
<description><![CDATA[We put a frontier LLM against a grim-trigger pricing rival and swept the odds the game keeps going. It cooperates exactly when game theory says it should.]]></description>
</item>
<item>
<title><![CDATA[We Ran Kimi K3 Against the Frontier on Our Own Coding Gauntlet]]></title>
<link>https://actualintelligencelabs.ai/research/kimi-k3-vs-frontier-private-eval</link>
<guid isPermaLink="true">https://actualintelligencelabs.ai/research/kimi-k3-vs-frontier-private-eval</guid>
<pubDate>Fri, 17 Jul 2026 12:00:00 GMT</pubDate>
<category><![CDATA[Report]]></category>
<description><![CDATA[Kimi K3 launched, so we skipped the leaderboard and ran it against Claude Fable 5 and GPT-5.6-Sol on our own coding gauntlet. K3 was the cheapest per task.]]></description>
</item>
<item>
<title><![CDATA[What Are the Best Free AI Tools in 2026?]]></title>
<link>https://actualintelligencelabs.ai/research/best-free-ai-tools</link>
<guid isPermaLink="true">https://actualintelligencelabs.ai/research/best-free-ai-tools</guid>
<pubDate>Wed, 15 Jul 2026 12:00:00 GMT</pubDate>
<category><![CDATA[Explainer]]></category>
<description><![CDATA[We checked the free tiers of ChatGPT, Claude, and Gemini in July 2026. Here is what is genuinely free, and where each one draws the line.]]></description>
</item>
<item>
<title><![CDATA[Do AI Detectors Actually Work? The Honest, Data-Backed Answer]]></title>
<link>https://actualintelligencelabs.ai/research/do-ai-detectors-work</link>
<guid isPermaLink="true">https://actualintelligencelabs.ai/research/do-ai-detectors-work</guid>
<pubDate>Wed, 15 Jul 2026 12:00:00 GMT</pubDate>
<category><![CDATA[Explainer]]></category>
<description><![CDATA[We reviewed the accuracy data behind AI detectors and AI checkers: false-positive rates, paraphrasing evasion, and why OpenAI retired its own tool.]]></description>
</item>
<item>
<title><![CDATA[How Much Water Does AI Actually Use? The Real Numbers, With Error Bars]]></title>
<link>https://actualintelligencelabs.ai/research/how-much-water-does-ai-use</link>
<guid isPermaLink="true">https://actualintelligencelabs.ai/research/how-much-water-does-ai-use</guid>
<pubDate>Wed, 15 Jul 2026 12:00:00 GMT</pubDate>
<category><![CDATA[Explainer]]></category>
<description><![CDATA[Published water-per-query estimates for AI range from 0.26 mL to 500 mL, a gap of almost 2,000x. We break down the real numbers and what drives it.]]></description>
</item>
<item>
<title><![CDATA[Is AI a Bubble? Here Is What the Data Actually Shows]]></title>
<link>https://actualintelligencelabs.ai/research/is-ai-a-bubble</link>
<guid isPermaLink="true">https://actualintelligencelabs.ai/research/is-ai-a-bubble</guid>
<pubDate>Wed, 15 Jul 2026 12:00:00 GMT</pubDate>
<category><![CDATA[Explainer]]></category>
<description><![CDATA[Real 2025-2026 numbers on AI stocks, capex, and revenue, the bull case and the bear case laid out plainly. Not financial advice, just the data.]]></description>
</item>
<item>
<title><![CDATA[The Agents Were Innocent: Field Notes From a Red Wave]]></title>
<link>https://actualintelligencelabs.ai/research/the-agents-were-innocent</link>
<guid isPermaLink="true">https://actualintelligencelabs.ai/research/the-agents-were-innocent</guid>
<pubDate>Wed, 15 Jul 2026 12:00:00 GMT</pubDate>
<category><![CDATA[Field Note]]></category>
<description><![CDATA[Our overnight wave of AI coding agents came back all red, and none of it was their fault. Field notes on formatting debt, a date time bomb, and dumb judges.]]></description>
</item>
<item>
<title><![CDATA[What Is Artificial Intelligence, in Plain English?]]></title>
<link>https://actualintelligencelabs.ai/research/what-is-artificial-intelligence</link>
<guid isPermaLink="true">https://actualintelligencelabs.ai/research/what-is-artificial-intelligence</guid>
<pubDate>Wed, 15 Jul 2026 12:00:00 GMT</pubDate>
<category><![CDATA[Explainer]]></category>
<description><![CDATA[Artificial intelligence in plain English: a one-sentence definition, how it actually works, and how AI, machine learning, and generative AI differ.]]></description>
</item>
<item>
<title><![CDATA[Who Invented Artificial Intelligence?]]></title>
<link>https://actualintelligencelabs.ai/research/who-invented-artificial-intelligence</link>
<guid isPermaLink="true">https://actualintelligencelabs.ai/research/who-invented-artificial-intelligence</guid>
<pubDate>Wed, 15 Jul 2026 12:00:00 GMT</pubDate>
<category><![CDATA[Explainer]]></category>
<description><![CDATA[No single person invented AI. Turing asked the question in 1950, McCarthy named the field in 1956. Here is the full, sourced timeline.]]></description>
</item>
<item>
<title><![CDATA[Will AI Take Over the World?]]></title>
<link>https://actualintelligencelabs.ai/research/will-ai-take-over-the-world</link>
<guid isPermaLink="true">https://actualintelligencelabs.ai/research/will-ai-take-over-the-world</guid>
<pubDate>Wed, 15 Jul 2026 12:00:00 GMT</pubDate>
<category><![CDATA[Explainer]]></category>
<description><![CDATA[AI researchers put a median 5% chance on AI causing human extinction, unchanged for years. Here is what the actual survey and forecasting data say.]]></description>
</item>
<item>
<title><![CDATA[Will AI Take Your Job? What the Data Says, Profession by Profession]]></title>
<link>https://actualintelligencelabs.ai/research/will-ai-take-your-job</link>
<guid isPermaLink="true">https://actualintelligencelabs.ai/research/will-ai-take-your-job</guid>
<pubDate>Wed, 15 Jul 2026 12:00:00 GMT</pubDate>
<category><![CDATA[Explainer]]></category>
<description><![CDATA[We checked the actual research on AI and jobs, profession by profession: lawyers, accountants, doctors, teachers, and radiologists. Real BLS and study data.]]></description>
</item>
<item>
<title><![CDATA[Why AI Agents Claim They're Done When They Aren't]]></title>
<link>https://actualintelligencelabs.ai/research/ai-agents-premature-completion</link>
<guid isPermaLink="true">https://actualintelligencelabs.ai/research/ai-agents-premature-completion</guid>
<pubDate>Fri, 10 Jul 2026 12:00:00 GMT</pubDate>
<category><![CDATA[Field Note]]></category>
<description><![CDATA[AI agents declare victory too early because nothing outside the model checks their work. Here are the four fixes that stop premature completion.]]></description>
</item>
<item>
<title><![CDATA[How to Write Instruction Files AI Agents Actually Follow]]></title>
<link>https://actualintelligencelabs.ai/research/ai-instruction-files</link>
<guid isPermaLink="true">https://actualintelligencelabs.ai/research/ai-instruction-files</guid>
<pubDate>Fri, 10 Jul 2026 12:00:00 GMT</pubDate>
<category><![CDATA[Field Note]]></category>
<description><![CDATA[CLAUDE.md and AGENTS.md fail the same way: they grow into an encyclopedia the agent stops reading. Here is the router pattern that keeps them working.]]></description>
</item>
<item>
<title><![CDATA[Running AI Coding Agents Unattended, Without Waking Up to a Mess]]></title>
<link>https://actualintelligencelabs.ai/research/autonomous-ai-loops</link>
<guid isPermaLink="true">https://actualintelligencelabs.ai/research/autonomous-ai-loops</guid>
<pubDate>Fri, 10 Jul 2026 12:00:00 GMT</pubDate>
<category><![CDATA[Field Note]]></category>
<description><![CDATA[Overnight AI agent loops fail the same predictable ways. Here is how we run wave-based backlogs with machine-checkable done gates and human approval limits.]]></description>
</item>
<item>
<title><![CDATA[Claude Fable 5 vs. GPT-5.6 Sol: A Practitioner's Field Report]]></title>
<link>https://actualintelligencelabs.ai/research/claude-fable-5-vs-gpt-5-6-sol</link>
<guid isPermaLink="true">https://actualintelligencelabs.ai/research/claude-fable-5-vs-gpt-5-6-sol</guid>
<pubDate>Fri, 10 Jul 2026 12:00:00 GMT</pubDate>
<category><![CDATA[Report]]></category>
<description><![CDATA[Claude Fable 5 and GPT-5.6 Sol compared on price, context, and positioning, plus our own field data from stress-testing a production harness with Sol.]]></description>
</item>
<item>
<title><![CDATA[We Ran 154 Controlled Trials of Claude vs GPT. Both Scored Perfect.]]></title>
<link>https://actualintelligencelabs.ai/research/claude-vs-gpt-154-run-experiment</link>
<guid isPermaLink="true">https://actualintelligencelabs.ai/research/claude-vs-gpt-154-run-experiment</guid>
<pubDate>Fri, 10 Jul 2026 12:00:00 GMT</pubDate>
<category><![CDATA[Report]]></category>
<description><![CDATA[We ran 154 controlled coding trials pitting Claude Fable 5 against GPT-5.6, blind dual-judged with a Wilson interval on every pass rate. Both scored perfect.]]></description>
</item>
<item>
<title><![CDATA[How to Optimize Content for Google AI Mode Search]]></title>
<link>https://actualintelligencelabs.ai/research/google-ai-mode-seo</link>
<guid isPermaLink="true">https://actualintelligencelabs.ai/research/google-ai-mode-seo</guid>
<pubDate>Fri, 10 Jul 2026 12:00:00 GMT</pubDate>
<category><![CDATA[Report]]></category>
<description><![CDATA[Google's own AI Mode usage report reveals four content signals worth optimizing for: question headings, task completion, follow-ups, and multimodal content.]]></description>
</item>
<item>
<title><![CDATA[Token Efficiency: Spending an AI Context Budget Wisely]]></title>
<link>https://actualintelligencelabs.ai/research/llm-token-efficiency</link>
<guid isPermaLink="true">https://actualintelligencelabs.ai/research/llm-token-efficiency</guid>
<pubDate>Fri, 10 Jul 2026 12:00:00 GMT</pubDate>
<category><![CDATA[Report]]></category>
<description><![CDATA[Token efficiency means treating context as a budget: progressive disclosure, subagent delegation, and prompt caching cut cost per task, not model size.]]></description>
</item>
<item>
<title><![CDATA[Running AI Coding on Matt Pocock's Method, in Production]]></title>
<link>https://actualintelligencelabs.ai/research/matt-pocock-ai-coding-method</link>
<guid isPermaLink="true">https://actualintelligencelabs.ai/research/matt-pocock-ai-coding-method</guid>
<pubDate>Fri, 10 Jul 2026 12:00:00 GMT</pubDate>
<category><![CDATA[Report]]></category>
<description><![CDATA[We run our AI coding pipeline on Matt Pocock's method: grill the plan, lock a spec, cut tickets, TDD, two-axis review. Here it is running in production.]]></description>
</item>
<item>
<title><![CDATA[When to Fan Out Subagents: Orchestration Patterns That Work]]></title>
<link>https://actualintelligencelabs.ai/research/multi-agent-orchestration</link>
<guid isPermaLink="true">https://actualintelligencelabs.ai/research/multi-agent-orchestration</guid>
<pubDate>Fri, 10 Jul 2026 12:00:00 GMT</pubDate>
<category><![CDATA[Field Note]]></category>
<description><![CDATA[When fanning work out to subagents pays off, how to fan verification back in, and the failure modes that show up when orchestration is done carelessly.]]></description>
</item>
<item>
<title><![CDATA[The Repo Is the Only Thing an AI Agent Actually Knows]]></title>
<link>https://actualintelligencelabs.ai/research/repo-as-source-of-truth</link>
<guid isPermaLink="true">https://actualintelligencelabs.ai/research/repo-as-source-of-truth</guid>
<pubDate>Fri, 10 Jul 2026 12:00:00 GMT</pubDate>
<category><![CDATA[Field Note]]></category>
<description><![CDATA[AI agents only know what a fresh session can read from the repo itself. Here is the fresh-session test and how to build state files agents can trust.]]></description>
</item>
<item>
<title><![CDATA[The Great Model Dump]]></title>
<link>https://actualintelligencelabs.ai/research/the-great-model-dump</link>
<guid isPermaLink="true">https://actualintelligencelabs.ai/research/the-great-model-dump</guid>
<pubDate>Fri, 10 Jul 2026 12:00:00 GMT</pubDate>
<category><![CDATA[Market Thesis]]></category>
<description><![CDATA[Cheap Chinese open-weight models are compressing AI margins while trillions are priced on scarcity. A sourced market thesis on which assumption breaks first.]]></description>
</item>
<item>
<title><![CDATA[Why AI Agents Fail on Tasks They Can Actually Do]]></title>
<link>https://actualintelligencelabs.ai/research/why-ai-agents-fail</link>
<guid isPermaLink="true">https://actualintelligencelabs.ai/research/why-ai-agents-fail</guid>
<pubDate>Fri, 10 Jul 2026 12:00:00 GMT</pubDate>
<category><![CDATA[Report]]></category>
<description><![CDATA[AI agents fail on real work less from weak models and more from a missing harness. See the five-layer model and the published data behind the gap.]]></description>
</item>
<item>
<title><![CDATA[AI Agent Evaluation: Test the System, Not the Model]]></title>
<link>https://actualintelligencelabs.ai/research/ai-agent-evaluation</link>
<guid isPermaLink="true">https://actualintelligencelabs.ai/research/ai-agent-evaluation</guid>
<pubDate>Wed, 10 Jun 2026 12:00:00 GMT</pubDate>
<category><![CDATA[Report]]></category>
<description><![CDATA[AI agent evaluation like software testing: three verification layers, a checker separate from the worker, evidence over self-report. Benchmarks are not enough.]]></description>
</item>
<item>
<title><![CDATA[Context Engineering: Designing Everything the Model Sees]]></title>
<link>https://actualintelligencelabs.ai/research/context-engineering</link>
<guid isPermaLink="true">https://actualintelligencelabs.ai/research/context-engineering</guid>
<pubDate>Wed, 10 Jun 2026 12:00:00 GMT</pubDate>
<category><![CDATA[Field Note]]></category>
<description><![CDATA[Context engineering goes beyond prompts: instructions, retrieved docs, tool definitions, state files, memory. How we design context for reliable AI systems.]]></description>
</item>
<item>
<title><![CDATA[Harness Engineering: When an Agent Fails, Fix the Harness]]></title>
<link>https://actualintelligencelabs.ai/research/harness-engineering</link>
<guid isPermaLink="true">https://actualintelligencelabs.ai/research/harness-engineering</guid>
<pubDate>Wed, 10 Jun 2026 12:00:00 GMT</pubDate>
<category><![CDATA[Report]]></category>
<description><![CDATA[Harness engineering makes AI agents reliable: repo as spec, done means a passing command, one task in flight, state in files. The doctrine our lab runs on.]]></description>
</item>
<item>
<title><![CDATA[React Three Fiber in Production]]></title>
<link>https://actualintelligencelabs.ai/research/react-three-fiber-in-production</link>
<guid isPermaLink="true">https://actualintelligencelabs.ai/research/react-three-fiber-in-production</guid>
<pubDate>Wed, 10 Jun 2026 12:00:00 GMT</pubDate>
<category><![CDATA[Field Note]]></category>
<description><![CDATA[React Three Fiber brings declarative Three.js to React. When to choose R3F over vanilla, with lessons on disposal, mobile budgets, and baked lighting.]]></description>
</item>
<item>
<title><![CDATA[How Scroll Animation Websites Work]]></title>
<link>https://actualintelligencelabs.ai/research/scroll-animation-websites</link>
<guid isPermaLink="true">https://actualintelligencelabs.ai/research/scroll-animation-websites</guid>
<pubDate>Wed, 10 Jun 2026 12:00:00 GMT</pubDate>
<category><![CDATA[Field Note]]></category>
<description><![CDATA[How a scroll animation website works: scroll progress as a timeline, video scrubbing vs frame sequences vs ScrollTrigger, Lenis, and the performance rules.]]></description>
</item>
<item>
<title><![CDATA[Three.js Examples That Belong on a Business Website]]></title>
<link>https://actualintelligencelabs.ai/research/three-js-examples-for-business</link>
<guid isPermaLink="true">https://actualintelligencelabs.ai/research/three-js-examples-for-business</guid>
<pubDate>Wed, 10 Jun 2026 12:00:00 GMT</pubDate>
<category><![CDATA[Field Note]]></category>
<description><![CDATA[Three.js examples that serve business websites: scroll-driven heroes, product configurators, baked scenes, 3D data viz. Each with a performance note.]]></description>
</item>
</channel>
</rss>
