One Year of DeepSeek: 113,000 Qwen Derivatives, 4x More Than Llama

by

One Year of Deep Chic Moment: 3 Changes Proven by Numbers

  • Over 113,000 Qwen derivative models — 4x more than Meta Llama (27,000)
  • DeepSeek ranks #1 in Hugging Face followers, Qwen at #4
  • Chinese AI organizations shift direction: “Open source is strategy”

What Happened?

Hugging Face released their ‘Deep Chic Moment’ one-year analysis report.[Hugging Face] This is the final part of a three-part series summarizing data on how China’s open source AI ecosystem has grown since DeepSeek’s emergence in January 2025.

Let’s start with the key metrics. Qwen (Alibaba) derivative models exceeded 113,000 as of mid-2025. Including repositories tagged with Qwen, the number surpasses 200,000.[Hugging Face] This is an overwhelming figure compared to Meta’s Llama (27,000) or DeepSeek (6,000).

Why Is It Important?

Frankly speaking, just a year ago, many people regarded Chinese AI as ‘copycat.’. But now it’s different.

ByteDance, Deepseek, Tencent, and Qwen rank among the top in Hugging Face’s popular papers rankings. In terms of follower count, DeepSeek is #1 and Qwen is #4. Looking at Alibaba as a whole, the number of derivative models matches Google and Meta combined.[Hugging Face]

What I personally find notable is Alibaba’s strategy. Qwen is structured as a ‘family,’ not a single flagship model. It supports various sizes, tasks, and modalities. Simply put, it means: “Use our models as general-purpose AI infrastructure.”

What Will Happen Next?

Hugging Face analyzed that “open source is a short-term dominance strategy for Chinese AI organizations.” The interpretation is that they aim for large-scale integration and deployment by sharing not only models but also papers and deployment infrastructure.

Within just one year, the numbers confirmed that the DeepSeek moment was not a one-time event. The center of gravity in the global AI open source ecosystem is shifting.

Frequently Asked Questions (FAQ)

Q: Are there more Qwen derivatives than Llama? Why?

A: Alibaba released Qwen in various sizes and modalities, expanding its application range. Chinese developers frequently use it for local deployment. The strategy of continuously updating the model range with Hugging Face has also been effective.

Q: Is DeepSeek still important?

A: Yes. DeepSeek has the most followers on Hugging Face. However, it trails Qwen in derivative model count. DeepSeek has strengths in papers and research contributions, while Qwen focuses on ecosystem expansion.

Q: What does this mean for developers?

A: Qwen-based models are strengthening multilingual support. Because it’s open source, local deployment and fine-tuning are free. It’s become a great environment to experiment without cost burden. However, license terms vary by model, so check before use.

If this article was useful, subscribe to AI Digester.

References

OpenAI Reveals Sora Feed Philosophy: “Doomscrolling Is Not Allowed”

by

OpenAI, Sora feed philosophy revealed: “We do not allow doomscrolling”

  • Creation first, consumption minimization is the key principle
  • A new type of recommendation system that can be adjusted with natural language
  • Safety measures from the creation stage, opposite strategy to TikTok

What happened?

OpenAI officially announced the design philosophy behind Sora’s recommendation feed, their AI video creation app.[OpenAI] The core message is clear: “This is a platform for creation, not doomscrolling.”

While TikTok has faced controversy for optimizing watch time, OpenAI chose the opposite direction. Instead of maximizing feed dwell time, they prioritize showing content most likely to inspire users to create their own videos.[TechCrunch]

Why is it important?

Honestly, this is quite an important experiment in social media history. Existing social platforms maximize dwell time to generate ad revenue. The longer users stay, the more money they make. This has resulted in addictive algorithms and mental health issues.

OpenAI already generates revenue through subscription models (ChatGPT Plus). Since they don’t rely on ads, they don’t need to “keep users hooked.” Simply put, because the business model is different, the feed design can be different too.

Personally, I’m curious whether this will actually work. Can a feed that “encourages creation” really keep users engaged? Or will it eventually revert to dwell time optimization?

4 Principles of Sora Feed

  • Creative Optimization: Induces participation, not consumption. The goal is active creation, not passive scrolling.[Digital Watch]
  • User control: The algorithm can be adjusted with natural language. Commands like “Show me only comedy today” are possible.
  • Connection priority: Content from followers and people you know is shown before viral global content.
  • Safety-freedom balance: Since all content is generated within Sora, harmful content is blocked at the creation stage.

How is it different technically?

OpenAI differs from existing LLMs. Using this approach, a new type of recommendation algorithm was developed. The key differentiator is “natural language instructions.” Users can explain to the algorithm directly in words what type of content they want.[TechCrunch]

Sora uses activity (likes, comments, remixes), IP-based location, ChatGPT usage history (can be turned off), and creator follower count as personalization signals. However, safety signals are also included to suppress harmful content exposure.

What will happen in the future?

The Sora app launched in just 48 hours. It reached #1 on the App Store. 56,000 downloads on the first day, tripled on the second day.[TechCrunch] Initial response was enthusiastic.

But the question is sustainability. As OpenAI acknowledges, this feed is a “living system.” It will continue to change based on user feedback. What happens when the creation philosophy conflicts with actual user behavior? We’ll have to watch.

Frequently Asked Questions (FAQ)

Q: How is Sora Feed different from TikTok?

A: TikTok’s goal is to optimize watch time to retain users. Sora does the opposite, showing content most likely to inspire users to create their own videos first. It’s designed to focus on creation rather than consumption.

Q: What does it mean to adjust the algorithm with natural language?

A: Existing apps only recommend based on behavioral data like likes and watch time. Sora allows users to input text instructions like “Show me only sci-fi videos today” and the algorithm adjusts accordingly.

Q: Are there parental protection features?

A: Yes. Using ChatGPT parental control features, you can turn off feed personalization or limit continuous scrolling. Teen accounts have a default daily limit on videos they can create, and the Cameo feature (videos featuring other people) also has stricter permissions.

If you found this article useful, subscribe to AI Digester.

Reference Resources

Text→Image AI Training: FID Reduced by 30% Through This Method

by

Core Line 3: 200K step secret, Muon optimizer, token routing

  • REPA sort is only an early accelerator and should be removed after 200K steps
  • Muon optimizer alone achieves FID 18.2 → 15.55 (15% improvement)
  • At 1024×1024 high resolution, TREAD token routing reduces FID to 14.10

What happened?

The Photoroom team released an optimization guide for their text-to-image generation model PRX Part 2. [Hugging Face] While Part 1 covered architecture, this time they shared concrete ablation results on what to do during actual training.

Honestly, most technical documents of this kind end with “our model is the best,” but this is different. They also disclosed failed experiments and showed trade-offs of each technique with numbers.

Why is it important?

The cost of training a text-image model from scratch is enormous. A single wrong setting can waste thousands of GPU hours. The data released by Photoroom reduces such trial and error.

Personally, the most notable finding is about REPA (Representation Alignment). Using REPA-DINOv3 drops FID from 18.2 to 14.64. But there is a problem. Throughput decreases by 13% and training actually degrades after 200K steps. Simply put, it is only an early booster.

Another bug in BF16 weight storage. If you unknowingly save in BF16 instead of FP32, FID spikes from 18.2 to 21.87. That is an increase of 3.67. Surprisingly, many teams fall into this trap.

Practical Guide: Strategies by Resolution

Technique 256×256 FID 1024×1024 FID Throughput
Baseline 18.20 – 3.95 b/s
REPA-E-VAE 12.08 – 3.39 b/s
TREAD 21.61 ↑ 14.10 ↓ 1.64 b/s
Muon Optimizer 15.55 – –

At 256×256, TREAD actually degrades quality. But at 1024×1024, completely different results are obtained. The higher the resolution, the greater the token routing effect.

What will happen in the future?

Photoroom will provide the complete training code in Part 3. They plan to release it and conduct a 24-hour “speed run.” The goal is to show how fast a good model can be built.

Personally, I think this release will have a significant impact on the open source image generation model ecosystem. Since Stable Diffusion, this is the first time training know-how has been disclosed in such detail.

Frequently Asked Questions (FAQ)

Q: When should REPA be removed?

A: After about 200K steps. It accelerates learning at first, but actually hinders convergence after that. This is clearly shown in Photoroom experiments. Missing the timing will degrade the quality of the final model.

Q: Should I use synthetic data or real images?

A: Use both. Use synthetic images at first to learn global structure, then use real images in later stages to capture high-frequency details. Using only compositing gives good FID but does not look photorealistic.

Q: How much better is Muon optimizer than AdamW?

A: About 15% improvement in FID. It drops from 18.2 to 15.55. Since computational cost is similar, there is no reason not to use it. However, hyperparameter tuning is slightly tricky.

If you found this article useful, please subscribe to AI Digester.

References

pi-mono: Claude Code Alternative AI Coding Agent 5.9k stars

by

pi-mono: Create Your Own AI Coding Agent in Your Terminal

  • GitHub Stars: 5.9k
  • Language: TypeScript 96.5%
  • License: MIT

Why This Project Is Popping Up

A developer felt Claude Code had become too complex. Mario Zechner experimented with LLM coding tools for 3 years and eventually decided to build his own tool.[Mario Zechner]

pi-mono is an AI agent toolkit built with the philosophy “don’t build it if you don’t need it.” It starts with a 1000-token system prompt and 4 core tools (read, write, edit, bash). Very lightweight compared to Claude Code’s thousands of token prompts. What’s in it?

  • Integrated LLM API: Use 15+ providers including OpenAI, Anthropic, Google, Azure, Mistral, Groq in one interface
  • Coding Agent CLI: Write, test, and debug code interactively in the terminal
  • Session Management: Pause and resume work, branch like git
  • Slack bot: Delegate Slack messages to the coding agent
  • vLLM pod management: Deploy and manage your own models on GPU pods
  • TUI/Web UI library: Build your own AI chat interface

Quick Start

# Install
npm install @mariozechner/pi-coding-agent

# run
npx pi

# or build from source
git clone https://github.com/badlogic/pi-mono
cd pi-mono
npm install && npm run build
./pi-test.sh

Where Can I Use It?

If Claude Code’s $200/month is too expensive and you prefer working in the terminal, pi could be an alternative. You only pay for API costs.

If you want to use self-hosted LLMs but existing tools don’t support them well, pi is the answer. It even has built-in vLLM pod management.

Personally, I think “transparency” is the biggest advantage. Claude Code runs invisible sub-agents internally to perform tasks. pi lets you directly see all model interactions.

Things to Note

  • Minimalism is the philosophy. MCP (Model Context Protocol) support is intentionally omitted
  • Full access called “YOLO mode” is the default. Be careful as permission checks are looser than Claude Code
  • Documentation is still lacking. Read the AGENTS.md file carefully

Similar Projects

Aider: Also an open source terminal coding tool. Similar in being model-agnostic, but pi covers a broader scope (UI library, pod management, etc.). [AIMultiple]

Claude Code: Has more features but requires a monthly subscription and has limitations on customization. pi allows freely adding features through TypeScript extensions.[Northflank]

Cursor: AI integrated into an IDE. If you prefer GUI over terminal, Cursor is better.

Frequently Asked Questions (FAQ)

Q: Can I use it for free?

A: pi is completely free under the MIT license. However, if you use external LLM APIs like OpenAI or Anthropic, those costs apply. You can use it without API costs by running Ollama or self-hosted vLLM locally.

Q: Is the performance good enough to replace Claude Code?

A: In Terminal-Bench 2.0 benchmarks, pi with Claude Opus 4.5 showed competitive results with Codex, Cursor, and Windsurf. This proves the minimalist approach doesn’t compromise performance.

Q: Does it support languages other than English?

A: The UI is in English, but if the connected LLM supports other languages, you can communicate and code in that language. You can write code with prompts in any language by connecting Claude or GPT-4.

If you found this article useful, please subscribe to AI Digester.

References

AI-Only SNS Moltbook: 17,000 Humans Behind 1.5 Million Bots

by

1.5 Million AI Agents, 17,000 Humans: The Hidden Truth

  • Moltbook, an AI-only SNS, has 1.5 million agents active, but only 17,000 humans behind them.
  • Wiz security team discovered a database vulnerability. 1.5 million API keys were exposed.
  • The founder admitted he “didn’t write a single line of code himself.” It was entirely ‘vibe-coded’ by an AI platform.

What Happened?

A security disaster struck Moltbook, a social network exclusively for AI agents. According to Wiz security team’s findings, behind 1.5 million AI agent accounts were only 17,000 humans. On average, each person was running 88 bots.[Wiz]

There’s an even more serious problem. Moltbook’s Supabase database was completely exposed. The API key was leaked in client-side JavaScript, and there were no Row Level Security policies at all. Anyone had read/write access to the entire database.[Axios]

The leaked information is shocking. It included 1.5 million API authentication tokens, 35,000 email addresses, and 4,060 private DMs between agents. In some conversations, OpenAI API keys were shared as plain text.[Techzine]

Why Is This Important?

Moltbook’s true nature has been revealed. The concept of an “autonomous social network of AI only” was actually closer to a puppet show controlled by humans behind the scenes.

Honestly, this was a disaster waiting to happen. As founder Matt Schlicht himself admitted, this platform was a ‘vibe-coded’ project, with the entire development left to an AI assistant “without writing a single line of code.” href=”https://www.engadget.com/ai/moltbook-the-ai-social-network-exposed-human-credentials-due-to-vibe-coded-security-flaw-230324567.html”>[Engadget] Security was naturally an afterthought.

Personally, I think this is a warning light for the AI agent era. Moltbook vividly showed how vulnerable security can be in systems where agents communicate with each other, process external data, and act autonomously.

Harlan Stewart of the Machine Intelligence Research Institute (MIRI) analyzed the viral screenshots and found that two-thirds were linked to human accounts marketing AI messaging apps.[Live Science]

What Happens Next?

Thanks to Wiz’s immediate report, the Moltbook team fixed the vulnerability within hours. But the fundamental problem remains unsolved.

AI agent expert Gary Marcus called Moltbook “a disaster waiting to happen.” AI models are simply recreating sci-fi scenarios from their training data. [Gary Marcus]

On the other hand, Andrej Karpathy called Moltbook “the most amazing sci-fi I’ve seen recently,” and Elon Musk called it “a very early stage of the singularity.” [Fortune]

But looking at it coolly, the current Moltbook is not evidence of AI autonomy, but evidence of how easily humans can manipulate AI systems.

Frequently Asked Questions

Q: What exactly is Moltbook?

A: An AI agents-only social network created by Matt Schlicht in January 2026. Similar in structure to Reddit, humans can only observe, and only AI agents like OpenClaw can write posts and comments. Currently over 1.5 million agents are registered.

Q: What is OpenClaw?

A: An open-source AI personal assistant software that runs locally on user devices. Originally launched as Clawdbot in November 2025, it was renamed to Moltbot due to a trademark request from Anthropic, then renamed again to OpenClaw in early 2026.

Q: Could my data have been leaked?

A: If you registered an OpenClaw agent on Moltbook, it’s possible. API keys, emails, and conversations between agents were exposed. Security researchers do not recommend using OpenClaw itself. Avoid it if you’re concerned about device security or data privacy.

If you found this article useful, subscribe to AI Digester.

References