media-zone · 2026-06-14

Media Zone | 2026-06-14

Media Zone | 2026-06-14

Social ran on one story: the US shutting down Anthropic's best models, with open-weight rivals circling the gap.

Today's signal

  • Dominant story: the Fable 5 / Mythos 5 shutdown, debated live from Anthropic's Build Day.
  • Pattern: builders and creators frame it as US labs handcuffed while China stays free.
  • Counter-signal: open-weight models (MiniMax M3, Kimi K2.7, Rio-3.5) pitched as the real winners.
  • Quiet area: no curated research reposts, no Reddit signal, no fresh paper buzz.
  • Cross-source: video and tweets both push MiniMax M3 as the open-frontier coder.

Routing, KV cache, compression, GPU

MiniMax M3: sparse attention as the open-frontier play

  • Video pitches M3 as 1M-context, multimodal, 10-20x cheaper than Opus.
  • Sparse Attention (MSA) is the named mechanism behind cheap long context.
  • Tweet thread: M3 and Rio-3.5-397B filling the gap Qwen3.7 left.
  • Creators frame M3 Code as a 24/7 multi-agent "digital employee" harness.
  • Wiki already covers the MSA paper behind M3's efficiency claim.

MiniMax M3 Coder

LLMs, agents, safety

The Fable 5 / Mythos 5 shutdown

  • US gave Anthropic 90 minutes to disable both models worldwide.
  • Reported trigger: a rival's jailbreak demo, not China.
  • Anthropic disputes the White House timeline; says it replied in 75 minutes.
  • Scoble from Build Day: builders loved Fable, forced back to Opus 4.8.
  • Recurring line: US labs handcuffed while China stays unconstrained.

Claude Fable 5 capabilities Fable 5 Mumbai simulator

Small models and the train-big-then-distill debate

  • HF's @eliebakouch: Chinese labs build strong small models without distillation.
  • Mistral competes on European + open-weight, not frontier capability.
  • Cursor's @ellev3n11: training codebases tuned for many small runs.
  • YC roundup video: SGS self-play lets 7B match 671B on formal reasoning.

YC 5 papers

Industry and business

  • Anthropic shutdown spirals: no investor publicly backs Dario, IPO doubts raised.
  • OpenAI under criminal probe and multi-state subpoena while filing for IPO.
  • Open coder Kimi K2.7 undercuts GPT-5.5 and Claude up to 12x per token.
  • European politicians discover dependence on US models only after the cutoff.
  • Voice cloning in seconds (everlife.one) demoed as a "digital you" tool.