AI News
AI community posts, news and release updates
Latest updates
- News
GPU supply tightens: B200 at 0% availability, GH200 tightening after Kimi K3
Published Jul 21, 2026B200 GPU availability is already 0%. Post-Kimi K3 release, GH200 availability is aggressively tightening. The 3F Research Fast Availability Index drops below 10%.
- availability
- B200
- Kimi K3
- GPU supply
- GH200
- News
Judge approves $1.5B Anthropic settlement for pirated books used to train Claude
Published Jul 21, 2026A US judge approved a $1.5 billion settlement between Anthropic and publishers over the use of pirated books to train Claude, drawing industry attention.
- Claude
- Anthropic
- settlement
- pirated books
An article discussing international agreements to limit frontier AI development, analyzing objectives and exit mechanisms.
- frontier AI
- AI Governance
- international agreements
OpenAI announces that ChatGPT and Codex have surpassed 10 million weekly active users, marking widespread adoption of AI tools.
- ChatGPT
- codex
- OpenAI
- user growth
A technique called TokenOptimization can significantly reduce token usage in Claude Code while maintaining output quality, helping developers save API costs.
- Claude Code
- token optimization
- cost saving
Gemini 3.6 Flash is now available on Antigravity platform. All weekly quotas have been reset, so start building.
- Antigravity
- AIモデル
- Gemini 3.6 Flash
- News
Kimi K3 vs Fable on 1,000 agentic tasks: router sends 72-96% traffic to K3, up to 50x cheaper
Published Jul 21, 2026Fireworks benchmarked Kimi K3 against Fable on ~1,000 agentic tasks. K3 excels in security, crypto, and long terminal loops; Fable in multi-lang and web/data viz. Per-task routing hits 93% accuracy, and routes 72-96% to K3, cutting costs up to 50x. Kimi K3 coming to Fireworks July 27.
- Cost Efficiency
- Fable
- routing
- agentic tasks
- Fireworks
- Kimi K3
- News
NVDA Vera Rubin NVL72 Benchmarked: 10x More Tokens per Megawatt than Blackwell on DeepSeek-R1
Published Jul 21, 2026$CRWV reports the first measured results of NVDA Vera Rubin NVL72, achieving 10x more tokens per megawatt than Blackwell on DeepSeek-R1, confirming efficiency gains in real-world settings.
- Efficiency
- NVDA
- DeepSeek R1
- Vera Rubin
- Blackwell
- tokens per second
- NVL72
Google is betting its inference future on a chip designed for a single model, potentially reshaping AI hardware.
- inference
- chip
- model-specific
While everyone talks about switching models for cost or sovereignty, the most advanced models are becoming more distinct. Fable, Kimi K3, and Sol respond very differently—you can't just swap them.
- Fable
- Kimi K3
- switching
- Sol
- model diversity