12-21-Daily AI News Daily

Alright, time to get this done. The user wants me to translate and edit this AI news digest, following a strict process. First, the Zero-Tolerance Rule: gotta scan the whole thing for those [text](URL) and ![text](URL) patterns and translate all the Chinese text inside the square brackets to English, while leaving the URLs and file paths in the parentheses untouched. Then, a quick self-check to make sure I didn’t miss anything.

Next, the Primary Editorial Task: Each paragraph gets a once-over to find the core idea, then I rewrite the paragraph putting the core subject upfront, and reorganize the rest of the information around it.

After that comes the Style Guide: Gotta keep things conversational, informal, and energetic, you know, everyday English, maybe with a little slang sprinkled in. Emojis? Gotta ditch the old ones and add fresh ones that make sense in English. And of course, the Markdown format has to stay exactly as it is. Code blocks? No translation of the code itself, but the comments within get the English treatment. Every bit of content has to be translated, and the paragraph structure must be kept.

Finally, the Final Output: Just the finished English text, no preamble, no explanations, no showing my work. Let’s do this!

Aivora AI Daily 2025/12/21

AI Daily

Today’s Summary

OpenAI now lets you manually adjust AI “enthusiasm” without needing incantations, Sora has officially partnered with Disney. Exo, a cool tool, can link idle phones into a GPU cluster, letting you run 70B large models at home. A must-read deep reflection from a top-tier expert (Redis Creator). And for those wanting to freeload Google Gemini, hurry and grab the plugin before it’s gone!

⚡ Quick Navigation

• 📰 Today’s AI News - Latest Updates at a Glance

💡 Tip: Want to be the first to try the latest AI models (Claude, GPT, Gemini) mentioned in this article? No account? Head over to Aivora to grab an account, get started in a minute, and enjoy worry-free after-sales support.

Today’s AI News

👀 Just One Sentence

Redis Creator declares the “stochastic parrot” theory dead, while OpenAI finally lets you manually adjust AI’s “enthusiasm”—no more long prompts begging it to talk like a normal human.

🔑 3 Key Takeaways

#RedisCreatorReflects #HomeComputeCluster #OpenAIPersonalization

🔥 Must-Read TOP 10 (In-Depth Visuals)

1. Redis Creator’s 2025 Year-End Reflection: Don’t Believe the “Stochastic Parrot” Theory Anymore

Salvatore (Redis Creator), though not an AI insider, offers an incredibly sharp perspective as a top-tier programmer. He bluntly states that the “stochastic parrot” theory (claiming AI merely probabilistically stitches words together) is no longer believed in 2025. This is absolutely today’s most worthwhile deep read.

Salvatore’s understanding of “Chain of Thought” (CoT) is a key highlight: he believes CoT essentially has the model lay out the context on a scratchpad first, then combine it with reinforcement learning to find the optimal solution. This explains why LLMs can solve math problems. He also mentions that the future bottleneck won’t be data volume, but “verifiable rewards” (like whether code runs or math proofs are correct), which will enable AI self-evolution.

The significance of a database expert analyzing AI is huge; when a big shot who writes databases starts deeply dissecting AI epistemology, you know things are fundamentally changing.

2. ChatGPT Adds “Personality Tuner”: Customize Tone Without Prompts

OpenAI’s new Personalization settings for ChatGPT are a game-changer; you no longer need to write a bunch of “Please act as a gentle…” prompts to make ChatGPT sound nicer. Now, you can directly drag sliders to adjust the AI’s “enthusiasm,” “title style,” and even “emoji usage frequency.”

Ethan Mollick confirms that changing the tone will not affect output accuracy. This is super interesting because it means we can train the AI to be a cool, detached coding assistant or an enthusiastic copywriter without worrying about it getting dumber.

This is what a product should look like. Turning prompt engineering into a UI interaction significantly lowers the barrier for ordinary users.

3. Exo: Connect Your Old Phones and Computers to Run Your Own AI Cluster

Exo is a godsend for hardware hoarders! Can’t afford an H100 GPU? This project lets you link up idle iPhones, iPads, Macs, and even old computers at home to form a unified GPU cluster for running large models.

Exo’s technical highlights are impressive: it’s not just simple distributed computing; it supports dynamic device discovery and automatically allocates loads based on device performance. You don’t need expensive specialized hardware to run Llama-3 70B-level models at home.

For privacy-conscious or budget-limited geeks, this is incredibly cool, even if the latency definitely can’t beat a single dedicated card.

4. OpenAI Codex “Skill Pack” Update: AI Finally Learns to Organize Files

GPT-5.2 (Codex) seems to be gray-testing new features, so coders, pay attention! Previously, AI coding felt like a one-off freelance job, but now it’s starting to understand “skill encapsulation.” Developers can package instructions and resources into folders (Skills), and the AI will automatically call them based on context.

The implication of this update is huge: the AI is starting to think about project structure like a real engineer, rather than just focusing on a few lines of code. After testing, some exclaimed, “This is the first model that feels like it’s writing code for itself, even leaving room for future work.”

The potential impact on Copilot is significant; if this feature rolls out widely, Copilot in IDEs might even start to look outdated. 😲

5. Gemini Nexus: Not Only Free Pro Models Without a Key, But Also Immersive Conversations

Gemini Nexus, a Chrome plugin developed by an L-site big shot, focuses on being “both useful and money-saving.” Google’s Gemini 3 Pro is great, but the web version can sometimes be clunky. This plugin integrates Gemini directly into your sidebar, supporting OCR text extraction, selection-based translation, and most importantly—no API Key needed!

Gemini Nexus’s core features include model selection and text replacement, updated to v3.0. For those who don’t want to bother with environment configurations and just want to quickly use top-tier AI capabilities in their browser, this is currently one of the most elegant solutions.

This kind of tool belongs to the “use it while you can” series; Google might block the API any day, so grab it before it’s gone! ⏳

6. Sora Teams Up with Disney: Mickey Mouse to Be AI-Generated?

Disney’s announcement could be a bombshell for the film and television industry. They’ve reached a three-year exclusive partnership with OpenAI, allowing Sora to use Disney’s IP (like Mickey Mouse, Marvel characters, etc.) to generate videos.

The commercial impact is significant: previously, everyone worried that copyright issues would kill AI video, but now it seems the giants have chosen the “if you can’t beat ’em, join ’em” approach. Fan-created content might even get a chance to appear on Disney+.

With official involvement, the “fan-fiction outshines official content” trope will likely play out even more frequently in the future. 🤯

7. Ranke-4B: A Time Capsule LLM That Doesn’t Know WWI Happened

The Ranke-4B project is quite poetic: what if an AI’s memory stopped in 1913? Researchers trained a model using only data from before 1913. It understands language structure but knows nothing of world wars, pandemics, or the internet.

The purpose of this project isn’t practicality, but research. How did people a century ago view the future? Confronting a century-old “brain” with modern topics could create highly literary conversations due to this sense of temporal displacement.

This is an excellent sociological experiment. Chatting with it feels like genuinely communicating with someone from the Titanic era.🚢

8. Anthropic Releases Bloom: A Tool Specifically for Testing if AI “Goes Bad”

Anthropic’s Bloom is an open-source tool for generating “behavioral misalignment assessments.” When AI capabilities grow stronger, how do you know if it’s a “two-faced” agent? Simply put, it automatically creates various诱导性 scenarios to test if the AI violates safety rules.

Bloom’s technical details allow researchers to specify a behavior and then quantify the frequency of AI “misbehavior” in massively generated scenarios. This is far more efficient than manually writing red-teaming prompts one by one.

Claude’s solid safety research is evident; this kind of tool is an essential quality inspector before AI enters critical domains. ✅

9. Jim Fan: Atari Games Are Key to Robot Evolution

Jim Fan’s NitroGen model from Nvidia can play over 1000 games, and don’t think AI playing games is just for fun. He points out that while large models can write poetry, their adaptability to the physical world is still inferior to early OpenAI Five.

Jim Fan’s core argument is to view 1000 games as 1000 simulated physical worlds. If AI can adapt to so many different rules and physics engines, it’s not far from controlling general-purpose robots.

The path to embodied AI (robotics) might just be paved with countless Marios and Zeldas. 🕹️

10. Paul Graham’s Advice to 18-Year-Olds: Curiosity > Aura of Success

Paul Graham’s (PG) advice, though not purely about AI tech, is a particularly refreshing dose of wisdom in this age of AI anxiety. The Silicon Valley startup godfather suggests: don’t blindly chase the so-called “aura of success” (like joining big tech or earning a high salary); what’s truly worth pursuing is curiosity.

The golden quote summarizes it best: “Effort determines how far you can go, curiosity determines where you will go.” In a world where AI can replace most repetitive labor, humanity’s remaining core competitive edge might truly just be curiosity.

The relevance of this advice today is profound; in an era where code can be automatically generated, these words sound more like a survival guide. 💯

📂 Categorized Overview

🛠️ Productivity Tools

• DeepAudit : China’s first open-source multi-agent for code vulnerability mining, allowing even beginners to deploy it with one click for security auditing.
• Claude Code : Anthropic’s official terminal programming tool, letting you control codebases directly from the command line using natural language.
• Leapcell : A new way to deploy Hugo blogs, managing static sites like a Python service, even supporting Astro.
• NotebookLM for PPT : Netizens discovered that Google’s NotebookLM can optimize PPT visuals, automatically generating styled pages by uploading PDFs.

🔬 Research & Observation

• Karpathy’s 2025 Review : The AI guru speaks: large model training is shifting towards RLVR (Reinforcement Learning with Verifiable Rewards), making programming as cheap as sketching.
• Zhipu AI to IPO in Hong Kong : Zhipu AI has officially submitted its prospectus, bringing the first large model stock to market, with Tencent and Alibaba both on the shareholder list. 🚀
• Moore Threads New GPU : Is the domestic GPU finally flexing? The new architecture claims to support a hundred-thousand-card cluster, with impressive specs on paper.

🧩 Open Source & Development

• Free-Certifications : A 47k-starred GitHub project that collects free certified courses from across the web—a must-have for ambitious learners.
• PayloadsAllTheThings : A legendary library for web security penetration testing, a comprehensive collection of payloads and bypass techniques.

🧾 Full Index

Aivora ⬆️ aivora.cn

Aivora AI Daily 2025/12/21

AI Daily

Today’s Summary

OpenAI now lets you manually adjust AI “enthusiasm” without needing incantations, and Sora has officially partnered with Disney. The awesome tool Exo can link idle phones into a GPU cluster, letting you run 70B large models at home. A must-read deep reflection from a top-tier expert (Redis Creator) is out, and for those wanting to freeload Google Gemini, hurry and grab the plugin before it’s gone!

⚡ Quick Navigation

• 📰 Today’s AI News - Latest Updates at a Glance

💡 Tip: Want to be the first to try the latest AI models (Claude, GPT, Gemini) mentioned in this article? No account? Head over to Aivora to grab an account, get started in a minute, and enjoy worry-free after-sales support.

Today’s AI News

📰 Just One Sentence

Redis Creator declares the “stochastic parrot” theory dead, while OpenAI finally lets you manually adjust AI’s “enthusiasm”—no more long prompts begging it to talk like a normal human.

🔑 3 Key Takeaways

#RedisCreatorReflects #HomeComputeCluster #OpenAIPersonalization

🔥 Must-Read TOP 10 (In-Depth Visuals)

1. Redis Creator’s 2025 Year-End Reflection: Don’t Believe the “Stochastic Parrot” Theory Anymore

Salvatore (Redis Creator), though not an AI insider, offers an incredibly sharp perspective as a top-tier programmer. He bluntly states that the “stochastic parrot” theory (claiming AI merely probabilistically stitches words together) is no longer believed in 2025. This is absolutely today’s most worthwhile deep read.

Salvatore’s understanding of “Chain of Thought” (CoT) is a key highlight: he believes CoT essentially has the model lay out the context on a scratchpad first, then combine it with reinforcement learning to find the optimal solution. This explains why LLMs can solve math problems. He also mentions that the future bottleneck won’t be data volume, but “verifiable rewards” (like whether code runs or math proofs are correct), which will enable AI self-evolution.

The significance of a database expert analyzing AI is huge; when a big shot who writes databases starts deeply dissecting AI epistemology, you know things are fundamentally changing. ![AI News Image]( https://pbs.twimg.com/media/G8ovcV