Media Zone | 2026-06-09

Cyber-capable Mythos and the IPO money race dominate; the only real architecture talk is Apple's per-prompt-routing on-device MoE.

Today's signal

Dominant story: Anthropic publishes measured N-day exploit uplift for unreleased Mythos, days before rumored public launch.
Money race: OpenAI files confidential IPO S-1, Cursor hits $4B ARR, US government weighs an OpenAI equity stake.
Architecture signal: Apple AFM 3 ships a 20B on-device MoE that picks experts + compute budget once per prompt.
Tooling pattern: parallel coding agents converge on git-worktree isolation (Kilo Agent Manager).
Price war: MiniMax M3 at half price with 1M-token context; Xiaomi and DeepSeek also cutting.
Quiet area: no practitioner Reddit signal today; all eight tracked subs returned nothing past filters.

Apple's on-device 20B MoE activates only 1-4B params, chosen once per prompt.
Early router fixes experts and compute budget for the full model depth.
Server model uses Parallel Track parallelism (sync every few layers), not expert parallelism.
HuggingFace's eliebakouch unpacked the official Apple blog into the routing details.
The product proof of "route compute by request difficulty" the research side keeps chasing.

Anthropic Red: Mythos Preview turns public patches into N-day exploits in hours, for a few thousand dollars.
Disclosed pre-release deliberately, to give defenders lead time before the capability spreads.
X chatter: a "public Mythos tomorrow" and a guardrailed "Claude 5 fable" variant in 48h.
YouTube leak-coverage channels already running Mythos 5 / GPT-5.6 checkpoint rumors.

Kilo Agent Manager runs each agent in its own git worktree: no file collisions, no context bleed.
Same prompt fans out to Claude Opus, GPT-5.5, and Claude Sonnet at once; keep the best.
bcherny: a year of Claude Code, now coding mostly from his phone via auto mode + routines.
Echoes the multi-agent-orchestration video thread on systems that coordinate without a shared LLM.

OpenAI filed a confidential IPO S-1, announcing early because it expected a leak.
Cursor crossed $4B annualized revenue, doubling in four months; SpaceXAI holds a $60B buy option.
US government reportedly weighing a direct equity stake in OpenAI, an idea attributed to Altman.
Anthropic poached OpenAI's second-ever chip engineer, signaling a custom-silicon push pre-IPO.

MiniMax M3 now runs at half its standard input price with a 1M-token context window.
Part of a broad race to the bottom; Xiaomi and DeepSeek also cut rates.
Apple expands Private Cloud Compute to Google Cloud running on NVIDIA GPUs.