AI News
AI community posts, news and release updates
Latest updates
Grok 4.5 is shockingly fast, efficient, and competent—a true workhorse. The only missing piece is GIGA GROK for the 1% of tasks that require a 4,000 IQ AI (token-hungry and inefficient though it may be).
- coding agent
- Efficiency
- Grok 4.5
- Launch
Kimi K3 open weights go live tomorrow! Run a frontier model privately on a $12,000 setup, 5-10x cheaper than predecessors
Published Jul 27, 2026Kimi K3's open weights go live tomorrow. You can run a frontier model privately on a $12,000 setup, costing 5-10x less than previous models while being more intelligent.
- 开源模型
- AI模型
- Kimi K3
- 低成本部署
- News
DeepSeek funding halted, leaked recording reveals chip smuggling and Huawei performance gap
Published Jul 27, 2026Reportedly, DeepSeek founder Liang Wenfeng halted Series B funding after a leaked recording of a closed-door investor meeting in July. The leak revealed that DeepSeek hoarded ~20,000 H100 GPUs via gray channels, bought large quantities of Huawei Ascend chips to bypass Nvidia CUDA, and admitted Huawei chips lag behind Nvidia by at least two years. The incident also involves internal factional struggles in China's tech sector.
- DeepSeek
- 融资
- 华为
- 地缘政治
- 梁文锋
- 芯片走私
Both offer fast mode and up to 1M token context windows, but monetize differently: Claude Code includes large context in subscription but charges for fast mode, while Codex includes fast mode but charges for larger context.
- Claude Code
- codex
- 定价
- 上下文窗口
Moonshot AI has released the Kimi K3 model weights today, available for download and deployment.
- Moonshot AI
- Kimi K3
- Discussion
Convo between current and ex-Anthropic researchers: Kimi K3 is Mythos-tier in cyber, but hidden by eval limits
Published Jul 27, 2026A user summarizes a conversation between current and former Anthropic researchers. Key points: Kimi K3's cyber capabilities may be Mythos-tier, but not shown due to inefficient reasoning and UK AISIS eval limited to 100M total tokens (seems too low); Anthropic's internal safety intent is genuine, but externally the system appears to optimize for revenue and control, does not diffuse capabilities to defenders, and lacks adequate safeguards against misuse.
- Anthropic
- safety
- eval
- Kimi K3
- Cyber
- News
Kimi K3 arrives on Together tomorrow with Provisioned Throughput, 65% cheaper than Fable
Published Jul 27, 2026Kimi K3 lands on Together tomorrow with Provisioned Throughput, offering guaranteed tokens per minute, 99% uptime SLA, and 65% lower cost than Fable.
- 成本优化
- Kimi K3
- Together
- API服务
New Baseten users can get $25 in free API credits to test Kimi K3 by joining the waitlist, no credit card required. Credits expire one month after redemption.
- 薅羊毛
- 免费API
- Baseten
- Kimi K3
A lightweight hook that intercepts before Claude Code uses tools, preventing it from editing .env files.
- Claude Code
- security
- hook
- .env
- PreToolUse
- News
Grok Build crushes efficiency: ~3.5M tokens per task with top-tier coding performance
Published Jul 27, 2026Grok Build stands alone on the efficiency frontier, delivering top-tier coding-agent performance using only ~3.5 million tokens per task. Competitors consume 7 to 18 million tokens—up to 5x more—for similar scores. More intelligence per token.
- coding agent
- Grok Build
- token efficiency