PAYLOADS
// INTELLIGENCE

A new paper from arXiv describes an AI agent that rewrites its own source code when it fails — not its prompts, not its memory schema, its actual code. Combined with Fujitsu's production self-evolution data, this changes everything about how we think about agent maintenance.

Google I/O 2026 delivered the most practically significant LLM announcement in months: Gemini 3.5 Flash ships at half the cost of comparable models with competitive reasoning benchmarks. This isn't about benchmarks — it's about economics.

Every AI team eventually discovers that their models are the easy part. The hard part is everything around them: data validation, model serving, monitoring, retraining triggers. Apache Airflow has been solving this problem for years, and it's still the best option for complex AI pipeline orchestration.

Every study published in the last two years showing AI coding tools improve productivity is measuring the wrong thing. Productivity metrics don't capture what happens to engineers who stop thinking for themselves. I'm watching this happen in real time and it's exactly as bad as you think.

After two years of watching teams struggle with getting LLMs to output consistent structured data, I've found the combination that works. It's not a fancy prompt technique. It's just being explicit about what you want in a way the model can't misunderstand.

Yesterday Fujitsu announced self-evolving multi-agent technology that learns from its own failures — and achieves 28-point accuracy gains without human intervention. This is the missing piece that enterprise AI has been waiting for.

llama.cpp merged Multi-Token Prediction support — and if you're running local LLMs, this is the upgrade you've been waiting for. Here's what it does and why it matters.

Stop paying per-token fees for development work. Here's how to get a production-quality LLM running on your own machine in under 20 minutes, with the exact setup I use every day.

The biggest infrastructure decision your AI team will make this year isn't which model to use. It's whether your agents work together through orchestration or through auction. Only one of those scales.

Google I/O 2026 just shipped something the industry has been pretending to want for two years: a frontier-quality model that's genuinely cheap and genuinely fast. Gemini 3.5 Flash isn't a lighter model. It's a redefinition of what a production LLM should be.

Every AI agent framework eventually runs into the same wall: the model knows the tools exist, but it doesn't know how to use them reliably. Here's the engineering discipline that actually makes tool calling work.

Every team is racing to adopt AI pair programmers. The data from places that have used them longest tells a darker story: the tools that were supposed to make us sharper are making us duller.

Stop paying for API calls when you are iterating on prompts. Here is how I run Llama 3 and friends locally in under 10 minutes.

Bigger context windows won't save you from bad architecture. They'll just let you delay the reckoning longer.

Cursor's .cursorrules file is the most powerful config you're probably ignoring. Most teams write one paragraph. Here's how to write one that actually changes how the AI behaves on your codebase.

Mira Murati's Thinking Machines Lab shipped its first model on May 11, 2026 — and nobody in the press is writing about the right part. This isn't a faster voice model. It's a fundamentally different architecture for how humans and AI collaborate.

Every AI startup offers fine-tuning as a premium service tier. Most of them are selling you expensive prompt engineering and calling it machine learning. Here's why the math never works out the way the vendors promise.

F5-TTS is an open-source zero-shot voice cloning system that produces results competitive with commercial offerings — without their price tags or their terms of service. If you're building anything that needs speech synthesis, you need to know about this.

Every week a new model tops a leaderboard. Every week teams make decisions based on those numbers. The whole system is theater. Here is why.

Most teams are still hand-crafting prompts like it's 2023. DSPy, the Stanford framework that's been quietly rewriting the playbook for building with LLMs, treats prompts as compiled artifacts — not handwritten guesses. Here's why that distinction matters.

JSON Schema isn't just for APIs. Learn how to use it as a first-class pattern in your AI agent tool pipelines to catch bad data before it breaks your workflow.

Subquadratic's SubQ 1M-Preview claims to be the first commercially available LLM where compute scales linearly with context length, not quadratically. Here's what that actually means and why the benchmark numbers are the least interesting thing about this release.

Most agent frameworks treat memory as a feature request. Mastra, the TypeScript-first agent framework that hit its 1.0 stable in January 2026, built observational memory as a first-class primitive — and that tells you everything about what kind of project this is.

If you're indexing thousands of records and reaching for LIKE queries, you're leaving performance on the table. Here's how to wire up PostgreSQL's full-text search with GIN indexes to cut query times from seconds to single-digit milliseconds.