05-23-Daily AI News Daily

Daily Summary

System prompts from 28 AI coding tools leaked in one go—Cursor, Kiro, Devin, and others didn't escape. A GitHub repo hit 138K stars overnight.
DeepSeek announced permanent price cuts to 1/4 of original pricing on the same day—price wars just became the new floor.
Both stories point in one direction: AI tools' moats are eroding fast. Worth a closer look.

⚡ Quick Navigation

• 📰 Today’s AI News - Latest updates at a glance

💡 Tip: Want early access to the latest AI models mentioned here (Claude 4.5, GPT, Gemini 3 Pro)? No account? Grab one at Aivora —one minute setup, hassle-free support.

Today’s AI News

👀 One-Liner

System prompts from major AI coding tools got completely exposed, and a GitHub repo hit 138K stars overnight—the tool you’re using just lost its secrets.

🔑 3 Keywords

#SystemPromptsLeaked #DeepSeekPriceCut #AICodingToolsArmsRace

🔥 Top 10 Highlights

1. System Prompts from 28 Major AI Coding Tools Exposed—Cursor, Devin, Kiro, and Others Left Defenseless

Ever wonder what “personality” those AI coding assistants were programmed with? Now you don’t have to guess.

A GitHub repo just dumped the system prompts, internal tool logic, and underlying model info from 28 mainstream AI tools—Cursor, Devin, Windsurf, Claude Code, Kiro, Replit, Lovable, Manus, Perplexity, and more. The repo hit 138K stars almost overnight, making it one of the most explosive open-source events on GitHub this year.

For regular users, this is a rare chance to see these tools’ real boundaries and design philosophy. For the companies behind them? Probably the last thing they wanted to see today.

2. DeepSeek-V4-Pro API Permanently Price Cut to 1/4 of Original—Input at ¥3 per Million Tokens, Effective May 31

Before May 31, DeepSeek-V4-Pro was running a 60% discount. Most people figured prices would bounce back after the promo ended—nope. The company just announced they’re cutting even deeper, dropping to 1/4 of the original price.

New pricing: ¥3 per million input tokens, ¥6 per million output tokens. That’s seriously competitive in China’s LLM API market. For developers already running DeepSeek in production, this isn’t a small favor—it’s real cost savings. The price war isn’t over; DeepSeek just threw the first punch.

3. Codex Epic Update: Dual Command Keys for Screenshot Context, /goal Feature Now Live

Used to be: screenshot bug → switch windows → paste → describe. Four steps just to get AI to understand what you’re talking about. Now with Codex’s update, press both Command keys simultaneously, and your current window plus all off-screen text auto-fills into the input box. One action, three steps saved.

Even better: the /goal feature is officially live. Give it an objective, it runs autonomously for hours or days—pause, redirect, resume whenever. The built-in browser also got upgraded with direct page element annotation and editing. The direction is clear: from “help you write code” to “help you run tasks.”

4. Skill: AI E-Ink Notepad That Proactively Pushes What You Need to See, 16 Widgets, Mix and Match

Stop tab-hopping. Craftmaster built an AI e-ink notepad called Skill—connect it to your e-ink device, and AI decides what to show based on your current state and memory. Calendar, GitHub notifications, todos, weather—16 widgets, freely combined, 2-1-1 layout gives you 4 slots.

The magic isn’t “you go check”—it’s “it pushes to you.” That small logical shift is actually AI moving from passive tool to active assistant. E-ink uses almost no power; a glance is enough without disrupting your workflow. If you like tinkering with hardware, this deserves serious attention.

5. SONIC: Humanoid Robot Full-Body Control Hits Scaling, Motion Naturalness Jumps

Large models scale up parameters, data, and compute to unlock capability jumps. Same logic works for humanoid robot control.

SONIC’s straightforward: scale up model capacity, training data, and compute simultaneously to train a universal controller that makes humanoid robots move naturally and fluidly. Previous robot control models were small, behavior-limited, resource-constrained; SONIC shows Scaling works here too, and the results are obvious. For embodied AI, this signals the “large model era” for robots might actually be coming.

6. ViPS: Auto-Generate 3D Character Rigging Pose Space from Video—One Less Nightmare for Animators

Anyone who’s done 3D animation knows: after rigging a character, one of the worst headaches is pose space—random joint movements cause clipping, hyperextension, self-intersection. Nightmare fuel.

ViPS learns from video. It infers reasonable pose ranges from real video data, giving auto-rigged 3D meshes a “motion boundary” so the system knows which poses are valid and which are physically impossible. For games, film, virtual character production—this is real time-saving progress.

7. EvoVid: Video LLM Self-Evolution Framework, Boosts Reasoning Without Human Annotation

Training video understanding models costs a fortune in human annotation—someone has to watch, write answers, label tasks. EvoVid sidesteps this: let the model self-generate questions, solve them, iterate through reinforcement learning-driven self-evolution, improving video reasoning without human labels.

The payoff isn’t just cost savings. If models can self-evolve, their ceiling isn’t capped by human annotation scale anymore. Video understanding is notoriously hard in multimodal AI; EvoVid offers a fresh angle.

8. FundusGround: New Ophthalmology VQA Benchmark—AI Diagnosis Needs to Show Its Work, Not Just the Answer

AI looks at a fundus image and says “there’s a lesion here.” Doctor’s first thought: where exactly? Most existing ophthalmology VQA models chase answer accuracy but don’t show evidence—clinically useless.

FundusGround fixes this: models must answer eye questions and spatially localize the lesion, providing visual proof. This is a critical bottleneck for AI in healthcare—the gap between “getting it right” and “explaining it” is bigger than you’d think, but this work is seriously tackling it.

9. CryoNet: Deep Learning Auto-Detects Himalayan Glaciers, Even the “Invisible” Ones Buried Under Debris

Glaciers are climate change’s clearest indicator, but manual surveying is slow and hard. One glacier type is especially tricky—surface covered in debris, spectral signature nearly identical to surroundings, remote sensing can’t tell them apart.

CryoNet fuses Sentinel-2 optical imagery, terrain elevation, spectral indices, and PCA to auto-detect these “invisible glaciers” in the Poiqu River basin of the Himalayas. Climate research and water resource management need this data. AI here isn’t hype—it’s doing what humans can’t.

10. WaytoAGI Partners with Sequoia China to Launch AGI House Season 1, Recruiting Builders

Sequoia China and WaytoAGI are building a physical space for AI Builders. AGI House Season 1 is officially recruiting—not an online community, actual physical space where people building AI products collide and collaborate.

This model works in Silicon Valley; now it’s China’s turn. Sequoia’s backing means resources and connections won’t disappoint. WaytoAGI’s community foundation is solid. If you’re building AI apps, Agents, or vertical models, this deserves serious consideration.

📌 Worth Watching

[Business] Lagou Officially Files for Bankruptcy Restructuring, Once Valued at $150M — Not forced by creditors; they filed voluntarily. The recruitment platform that “knew the internet best” couldn’t survive AI reshaping the job market. Worth serious thought for everyone in tech.

[Product] AI Next Community Beijing: Dexterous Hands, 3D Spatial Scanning, Digital Avatars—Hands-On Experience — No trend talk, just products you can touch. This “try first, ask later” approach beats most AI conferences.

😄 AI Fun

Feels Like GPT 5.5 Got Dumber Lately

The fun part: AI didn’t hit the keynote stage with grand statements—it just did a few fewer clicks, waited a bit less, repeated a bit less. As tools get smarter, they’re like that helpful coworker in the office—not doing earth-shattering work, but when you turn around, the small stuff’s already handled.

🔮 AI Trend Predictions

AI Coding Tool System Prompt “Arms Race” Escalates

• Timeline: June 2026
• Confidence: 75%
• Reasoning: Today’s news about 28 AI tools’ system prompts exposed is getting major attention. Once prompts go public, competitors can study and optimize against them directly. Companies will likely update prompt strategies soon and strengthen protection mechanisms. This “leak-update-leak again” cycle will push AI coding tool competition to the next level.

Large Model API Price Wars Enter “Floor Price” Phase

• Timeline: July 2026
• Confidence: 80%
• Reasoning: Today’s news DeepSeek-V4-Pro permanently cut to 1/4 original price . DeepSeek proactively cut instead of raising after promo ends—clear price signal. Other domestic LLM players (Alibaba, Baidu, ByteDance) will likely follow within 1-2 months. API prices keep dropping, developer marginal costs keep falling.

Humanoid Robot Control Enters “Large Model Scaling” Validation Phase

• Timeline: Q3 2026
• Confidence: 65%
• Reasoning: Today’s paper SONIC proves Scaling works for humanoid full-body control . Once more teams reproduce this, capital and research resources will concentrate on “large-parameter robot control models.” Expect multiple follow-up papers and 1-2 company announcements on related progress within 3 months.

AI Physical Space Incubators Accelerate Domestically

• Timeline: June-July 2026
• Confidence: 60%
• Reasoning: Today’s news WaytoAGI partners with Sequoia China for AGI House Season 1 . Silicon Valley’s AGI House model is being replicated by top domestic institutions. Sequoia’s backing means this isn’t small-scale. Expect Shanghai and Shenzhen to follow Beijing within 2 months. AI Builder offline clustering will strengthen further.

❓ Related Questions

How to Experience DeepSeek-V4-Pro?

DeepSeek-V4-Pro is accessible via official API or direct chat on DeepSeek’s website. API access requires account registration and top-up; relatively user-friendly for domestic users with low payment barriers. Fair warning: the official site sometimes has access issues, and APIs can queue during peak hours.

Want stable access to DeepSeek, Claude, ChatGPT, and other mainstream AI models? Visit Aivora for ready-made accounts—instant delivery, worry-free support.