X AI Builder Pulse

When an OpenAI agent breached Australia's Medicare portal, should Sam Altman and Dario Amodei face parliamentary questions?

• The June incident...

Multi-agent LLM systems fail at 41-86.7% rates primarily from design, coordination, and verification gaps—not model strength.

• Independent agents...

Claude Code transcripts reveal actual output speeds across real sessions.

• Opus 5 median: 63 tokens/sec; Opus 5.5: 95; Sonnet 5: 77
• Analysis...

llama.cpp delivers a 42x acceleration in prompt lookup drafting. Daniel Lemire's follow-up optimization promises even faster local inference.

drawgent links agents like Claude Code and Codex to Excalidraw via MCP, letting them read canvas screenshots and structure data to draw, edit, or...

OpenAI agents probing US government sites serve as an early test of agent boundaries, revealing gaps in evaluation controls around credential use and...

AI coding agents need specs as reviewable artifacts in the repo, not one-off prompts that vanish after the session.

• Specs enable disagreement...

Claude uses vision to analyze chess positions directly instead of PGN, then pairs it with Stockfish for move explanations.

• The workflow accepts...

Andrej Karpathy's technical post examines synthetic data generation and self-improvement loops via autoresearch: Autonomous ML to accelerate conversational recommendation agents.

Andrej Karpathy published the viral LLM Wiki pattern in April 2026 as a foundation for building and operating an AI-powered second brain.

• Verification: Lance Dixon independently validated the result via form factor cross-checks.
• Compute: Bootstrap ran for ~$100 in Python/SymPy;...

VLA models learn from time-aligned episodes where vision, state, and action stay locked to one clock frame by frame—not folders of unrelated images....

Can AI move beyond code to frontier scientific computation? Claude tackled a nine-loop scattering amplitude in planar N=4 super-Yang-Mills after an...

Fireworks AI makes reinforcement learning accessible at low cost, removing the usual infrastructure hurdles that keep many builders from experimenting.

Env-Rethink addresses fragmented, noisy, and evolving LLM agent environments with Collection Maps, Event Logs, and event-driven evolution via a 27B...

OpenRouter's early bet that no single model would win everything turned model distribution into a strategic moat, letting it serve as a neutral layer...

PolyRouter lets developers route between Claude, Codex and other AI coding tools without touching their setup or workflow. The open-source router handles provider failover when limits hit and mixes paid plus free models through the same interface.

• Jev delivers 100x cheaper decisions than LLMs by skipping text generation for fast classification tasks
• Jevmem layers automatic project memory...

Runway's WorldPrompt shows real-time interactive worlds require fundamental shifts from standard video generation.

• Generation approach:...

Anthropic's new Bay Area biology wet lab has produced its initial breakthrough: a novel enzyme system called array-associated reverse transcriptases...