AI News
AI community posts, news and release updates
Latest updates
A tool like brew doctor to check the health of skills and MCP loadouts.
- MCP
- 工具
- 技能
- brew
- Tutorial
Run Gemma 4b inside ChatGPT's VM: bypass network restrictions and deploy LLM in 14 minutes
Published Jul 22, 2026A user discovered a workaround to run smaller LLMs like Gemma 4b inside ChatGPT's virtual machine by setting up a proxy to bypass network restrictions, successfully downloading and running the model in 14 minutes.
- ChatGPT
- LLM
- workaround
- Gemma
- hack
With AI-generated fake resumes flooding the hiring process, a developer built a detection tool to help recruiters filter out bot-submitted resumes and restore order.
- tool
- Hiring
- HR
- AI resume spam
- anti-bot
- Launch
Korean AI Firm Upstage Releases 250B MoE Model, Rivals DeepSeek-V4-Flash
Published Jul 22, 2026Upstage, a Korean AI company, has released a new open-weight model with 250B total parameters and 15B active MoE architecture, showing performance similar to Deepseek-V4-Flash. It features strong Korean language capabilities and is suitable for local deployment on DGX setups, highlighting the growing sovereign AI trend.
- MoE
- open-weight
- sovereign AI
- Upstage
- Korean AI
- Idea
The Intelligence Explosion May Come Sooner: Longer Reasoning Makes Models Smarter, and the Cost of Thought Is Crashing
Published Jul 22, 2026This essay argues that the intelligence explosion may happen before superclusters are finished. Three reasons: 1) Model capability scales with test-time compute; Noam Brown says models can improve for weeks. 2) The cost of reasoning is collapsing—OpenAI halved costs, Vera Rubin achieves 10x tokens per megawatt. 3) Breakthroughs are already here: an OpenAI model disproved an Erdős conjecture, Fable 5 helped find a counterexample to the Jacobian conjecture, and GPT-5.6 SOL used massive inference to find a zero-day. Persistence is key. Sam Altman talks about directing enormous compute to the hardest problems and getting results beyond entire teams.
- OpenAI
- Reasoning
- agents
- Efficiency
- test-time compute
- intelligence explosion
Yorishiro is a macOS terminal built for AI agents to live and interact directly in your shell environment.
- AI
- terminal
- agents
- macOS
- News
US science advisor accuses Moonshot AI of distilling Claude Fable 5 to train Kimi K3
Published Jul 22, 2026Amid reports of Trump admin considering restrictions on open-source AI, US science advisor Kratsios accuses Moonshot AI of distilling Claude Fable 5 at scale to train Kimi K3, and using Nvidia GB300 servers in Thailand. Commenters question the timeline: Fable 5 was released July 1, Kimi K3 July 16, only 15 days apart.
- 蒸馏
- 月之暗面
- Claude Fable 5
- Kimi K3
- 美国制裁
- Launch
Grok Build getting major upgrade: App Deployer lets you build and publish apps instantly from descriptions
Published Jul 22, 2026Grok Build is introducing an App Deployer that allows users to describe an app, have Grok build it, and then publish it instantly with a public URL, including one-click deployment, templates, and custom domains.
- Grok
- Grok Build
- AI platform
- app deployer
- News
Kimi K3 + Tinker enables auto-research: one-shot reproduction of Self-Distilled RLVR paper
Published Jul 22, 2026A user shares that Kimi K3 combined with Tinker creates a powerful auto-research loop. Reproducing the Self-Distilled RLVR paper, Kimi one-shot implemented the baseline, ran 19 experiments across 6 configs, and wrote the final report (plus a bonus Chinese version).
- AutoResearch
- Kimi K3
- Tinker
- RLVR
- reproducibility
This article shows how to reverse-engineer and recreate a 1980s version of MATLAB using Claude, bringing extinct software back to life.
- AI
- Claude
- reverse engineering
- retro
- Matlab