Newsletter Archive

A lead engineer describes how AI agents fundamentally changed their development process, moving beyond incremental speedups to reshaping how work gets done — a candid look at what actually works versus hype.

DZone AI · 1 min read Community

OpenAI's Frontier team built a 1M+ line-of-code internal product (Symphony) with zero human-written or human-reviewed code, consuming 1B+ tokens daily. Ryan Lopopolo describes this as 'harness engineering'—the next frontier beyond prompt engineering—where teams focus on structuring context and capabilities for AI agents rather than writing code manually. The five-month experiment demonstrates AI agents can handle full software lifecycle decisions when given proper system prompts and reference implementations.

Latent Space · 57 min read Tools

Anthropic is limiting Mythos Preview to a handpicked group of tech and cybersecurity firms because the model autonomously discovers tens of thousands of vulnerabilities—83% of which it can exploit on first attempt—including decades-old flaws in Linux, Windows, and major browsers that humans missed. This is the first time Anthropic has withheld a model from public release due to safety concerns about its core capabilities.

Axios · 4 min read Industry

Anthropic is previewing Claude Mythos—its most powerful model yet—to selected enterprises including Amazon, Apple, and Microsoft through Project Glasswing, a controlled-access cybersecurity initiative designed to harden defenses before the model's general release. The company believes Mythos's advanced agentic and reasoning capabilities pose unprecedented cyberattack risks and wants industry partners to stress-test security protections.

Fortune · 4 min read Industry

An Indianapolis councilmember who supported a $500M data center project was shot 13 times at his home with a note reading "No Data Centers" left at his door, with his 8-year-old son present. The incident marks an escalation in nationwide opposition to AI infrastructure expansion as hyperscalers race to build capacity.

Fortune · 3 min read Policy

Anthropic jumped from $19B to $30B ARR and unveiled Claude Mythos, a model so capable at finding zero-day vulnerabilities (including in Linux, OpenBSD, FFmpeg) that it's too dangerous for public release—restricting it to 40 enterprise partners under Project Glasswing. The 244-page system card details breakthrough capabilities with notable training efficiency improvements.

Latent Space · 3 min read Industry

Google's March 2026 Gemini API update lets developers combine Google Search, Google Maps, and custom functions in a single API call with context circulation and parallel tool execution. This tutorial walks through five progressive demos building from basic tool combination to full multi-step agentic chains, with working code examples using gemini-3-flash-preview and gemini-2.5-flash.

MarkTechPost · 12 min read Tools

MegaTrain trains 100B+ parameter LLMs at full precision on a single GPU by storing weights in CPU memory and streaming them to the GPU during computation. The system achieves 1.84× throughput over DeepSpeed ZeRO-3 on 14B models and can train 120B parameter models on an H200 with 1.5TB host RAM.

arXiv NLP · 3 min (abstract) Research