- Sept 8 joint advisory (AA26-251A) names DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Z.AI for systematically distilling Claude, GPT, Gemini, and Grok since late 2024.
- Alleges billions of tokens across millions of queries were pulled via fraudulent accounts, bulk premium subs, and proxy 'transfer stations' — likely with Chinese government knowledge.
- Specifics cited: DeepSeek R1/V3 trained on Claude, GPT, and Gemini output; Moonshot's Kimi K3 trained on Claude Fable, Kimi K2 on GPT-4o.
- Unusual mitigation ask: US labs 'quietly degrade' suspected accounts rather than block them, to preserve the counter-distillation signal.
- GA today (Sept 10, Beijing): new architecture at 284B total MoE / ~13B active, top-6 of 256 routed experts, 1M-token context, native multimodal in one endpoint.
- Every call to the deepseek-v4-pro API now silently serves V4.1 Flash at Flash prices until an unspecified V4.1 Pro lands.
- New Flash pricing effective today: $0.003/M cached-in, $0.15/M input, $0.60/M output off-peak; peak-hour rates double.
- HN reaction focused on the silent reroute — production users unhappy that a workflow validated on V4 Pro is now served by a different model with no opt-out.
- Sept 8: Muse leaves closed alpha for public US availability on iOS, Android, muse.ai, and WhatsApp; AI-glasses support 'coming soon.'
- Free tier caps at 100M tokens/week; Power ($20/mo) and Maximum ($100/mo) lift the cap. Meta stock closed +6% on the news.
- Each user's agent runs inside its own Muse Secure VM in Meta's cloud; purchases route through Stripe Link one-time card numbers so Meta never sees a real card.
- HN thread: 362 points, 379 comments — dominant sentiment 'I want this, just not from Meta.'
- Sept 2 X post targets Sept 12 for Grok 4.7 at 2.1T total parameters (40% up from 4.6's 1.5T), partly trained on internal SpaceX engineering and Starlink telemetry.
- As of Sept 10, xAI dev docs still list 4.6 as newest; no model card, pricing, context window, benchmark, or architecture disclosure has been published.
- Musk's own timeline has slipped twice already: Aug 22 → 'three to four weeks' → 'ten days.'
- Sept 9 Fireship video runs an identical game brief through GPT-6 Astra and Claude Fable 5.1, calls Fable's version the only playable output.
- Echoes a week of cost-versus-taste chatter: Astra is priced 2.5× above OpenAI Sol; Fable 5.1 cut prompt-cache reads 75% on Sept 1.
- Signals the community's shift from raw benchmarks to end-artifact quality when picking a coding model.
- jingyaogong/minimind climbs GitHub trending: full pretraining, SFT, DPO, PPO/GRPO, tool-calling, and agentic RL — all raw PyTorch, no framework wrappers.
- The 64M MiniMind-3 model — roughly 1/2700th the size of GPT-3 — trains on a single RTX 3090 in about 2 hours; cloud-GPU cost around $0.40.
- 56.8K+ stars and rising, as GitHub trending continues to pivot from 'download a model' to 'build one yourself.'
01
NSA, CISA, and FBI Name Six Chinese AI Firms in 'Industrial-Scale' Distillation Advisory
industry cisa.gov
02
DeepSeek V4.1 Flash Ships and Silently Reroutes V4 Pro Traffic
models deepseek.com
03
Meta Ships Muse to the US Public With Free, $20, and $100 Tiers
industry about.fb.com
04
Musk Puts Grok 4.7 on a Sept 12 Countdown — 2.1T Params, No Model Card
models x.ai
05
Fireship: 'I Built the Same Game With Astra and Fable 5.1… Only One Was Fun'
community youtube.com
06
MiniMind Trends: Train a 64M-Param LLM From Scratch in 2 Hours for $0.40
open-source github.com