Social Media Briefing — June 17, 2026

swyx reports that Cursor/Graphite's Tomas Reimers announced Origin, a Git competitor scalable for agent workloads with API/MCP extensibility and built-in merge conflict and CI failure agent resolution.

Midjourney announces it will reveal its first hardware project at an in-person San Francisco launch event with a livestream.

Mollick argues that if AGI is achievable and labs can only run a model internally by also releasing it publicly, top labs may instead capture all value via expansion and acquisition rather than share access.

NVIDIA Research introduces SpatialClaw, a training-free agent that uses Python code as its action interface for spatial reasoning and visual tasks, reporting an 11.2-point gain over a prior agent across 20 benchmarks.

Nathan Lambert announces a new podcast surveying 2026 post-training recipes (GLM 5.1, Kimi K2.6, DeepSeek V4, Xiaomi MiMo V2.5, Nemotron Ultra), discussing the industry shift to multi-teacher on-policy distillation, Olmo recipe needs, and career advice.

Anthropic introduces a framework for tracking Claude Code as it scales, examining who uses it, what for, how task value changes, and how domain expertise affects success.

Following yesterday's News coverage of the GLM-5.2 launch, Newsletter reports Chinese lab Z AI released GLM-5.2, an open-weights MIT-licensed model with a 1M-token context window, claiming benchmark wins over GPT-5.5 and Opus 4.8 on coding and math.

Argues that Microsoft, simply by saying it is seeking alternatives, has damaged the potential IPOs of both OpenAI and Anthropic.

Argues that since open models lag closed ones by 8-12 months, IT systems have roughly 4-8 months to harden against Mythos-class models, making safe defensive Mythos-class models important now.

Teases a new Mistral model arriving this summer that starts a new family described as fat but sparse, with an early access program in July for research, government, and industry partners.

Chollet argues open and universally available AI will come from radical efficiency gains in inference compute and especially training data, which he attributes to symbolic learning.

OpenAI announces new research on anticipating real-world model behavior by simulating deployment with recent de-identified user requests and studying candidate responses.