Media Zone | 2026-06-11

Anthropic's invisible-throttling walkback owned the feed, while cheap-and-fast inference shipped as product from Google and MiniMax.

Today's signal

Dominant story: Anthropic reversed Fable 5's secret throttling after researchers called it "sabotage."
Pattern: inference cost is the new battleground (DiffusionGemma 4x faster, MiniMax M3 at 1/10 price).
Cross-source: video creators and Twitter both framing Fable 5 as too-good-but-expensive and now trust-dinged.
Counter-signal: @tinygrad pushing back on AI-race hype amid Google's open releases.
Quiet area: no Reddit practitioner signal today, all eight subs empty.

Google ships Apache-2.0 diffusion LLM, claims 4x faster single-user generation.
Key trick: refine a 256-token canvas in parallel, not token-by-token.
Flips autoregressive memory-bound decode into compute-bound diffusion, scales with added compute.
@tinygrad's take: look past hype, Google owns the most compute.

Anthropic admits "wrong tradeoff," makes frontier-AI-dev safeguards visible this week.
Flagged requests now visibly fall back to Opus 4.8, API returns refusal reason.
@eliebakouch: hiding the nerf was the real harm, open models matter.
@ns123abc framing: researchers called secret degradation "sabotage," Anthropic apologized.
Video creators still rate Fable 5 the best model, "slow, expensive, capable."

Cursor Bugbot 3x faster, 22% cheaper, 90% of runs under 3 minutes.
New pre-push /review mode, dedupes against the GitHub PR review.
Poetic raised $50M at $500M for hallucination-free multi-hour enterprise automation.
Poetic claim: 99%+ accuracy, 10x fewer tokens than agents.

OpenAI reportedly weighing "drastic" price cuts, Altman says costs "a huge issue."
Claim: $500B Stargate Ohio is an estimate, near-term ~800 MW by 2028.
German court: Google AI Overviews are Google's own speech, liable for falsehoods.
Gary Marcus argues that logic could pierce US Section 230.
NVIDIA GTC Taipei: Stanford "Marlowe" SuperPOD with 248 Hopper GPUs.