- Agents Made Simple
- Posts
- 👾 AI Nuggets #15: Global AI Race - EU $200B vs China Models
👾 AI Nuggets #15: Global AI Race - EU $200B vs China Models
Plus: The EU challenges US/China dominance, ByteDance's Goku slashes video ad costs, Perplexity sonar cuts response times and YouTube unveils Veo 2 video tools.
Tobias Wupperfeld
February 17, 2025
Welcome to edition #15 of the MadebyAgents newsletter. In this issue:
AI action summit in Paris: tensions between US, UK and EU
Latin America is working on LatamGPT
Impressive video model “Goku” from China
Perplexity sonar model with super low latency
YouTube releases AI video tools for shorts
How to convert videos to human-like articles
CursorAI pro tips (developers must know)
Latest News
Source: MadebyAgents via Grok2
The Paris AI Action Summit revealed growing divides between Western powers. US, UK, and EU leaders clashed over regulatory approaches while racing to establish dominance in AI development.
What’s New
The EU pushed its strict AI Act framework. US delegates countered with voluntary industry guidelines. Britain positioned as mediator, proposing a new "third way" regulatory model. No joint agreement emerged. France announced €500M in new AI startup funding.
Business Impact
Diverging rules create compliance headaches for international firms. EU-based companies face stricter transparency requirements. US/UK startups gain short-term flexibility but risk future alignment costs. Tech professionals should monitor:
EU's mandatory AI system audits
US tax incentives for AI R&D
UK's new AI safety institute grants
Source: latamgpt.org
Nine Latin American countries launch LatAmGPT - an open-source AI trained on local languages, dialects, and cultural data. Public release planned for mid-2025.
What's New
Chile’s CENIA leads 30+ institutions in developing Spanish/Portuguese models that understand regional slang, indigenous terms, and cultural contexts. The system uses 8TB of local texts - equivalent to 4 million books. Unlike closed AI like ChatGPT, code will be freely modifiable. The supercomputing hub in northern Chile handles the processing.
Business Impact
Lower AI adoption costs for SMEs using local-language tools
Startups can build custom solutions without licensing fees
Content creators gain culturally aware writing assistants
Source: Goku
ByteDance's new Goku model creates marketing videos from text prompts. The system specializes in human avatars and product showcases, claiming a 100x cost reduction over filming crews.
What's New
Goku handles complex 20-second sequences where competitors stall at 5 seconds, maintaining stable hand motions and facial expressions. It converts static product photos into rotating 360° videos automatically. Unique cultural adaptation lets marketers insert regional references, from Chinese poetry motifs to Mumbai street scenes. University of Hong Kong researchers contributed core architecture.
Business Impact
Marketing teams can create spokesperson videos with tiny budgets. Think of localized ads using neighborhood backdrops and regional dialects.
Technical teams require prompt engineering skills to direct AI actors. Ethical debates emerge about AI presenters replacing human influencers. China's strict AI regulations may delay global access despite open-source foundations.
Source: Perplexity
Perplexity's Llama3 Sonar 70B processes queries 58% faster than previous models while maintaining accuracy. The online version prioritizes real-time performance for time-sensitive applications like trading alerts and emergency response systems.
What's New
The model achieves 127ms average latency through optimized token streaming - faster than human blinking (300-400ms). It handles 28,000 context tokens without speed loss, letting users analyze lengthy documents in one go. Unlike offline AI, it combines fresh web data with core knowledge up to July 2024. Developers report 40% lower API costs versus similar-scale models.
Business Impact
Customer service teams could resolve tickets 3x faster using instant fact-checking during live chats. Startups could prototype real-time translation tools without cloud computing bills. But it comes with tradeoffs - the speed focus limits complex math tasks.
Source: Youtube blog
YouTube now lets creators generate entire video scenes through text prompts using Veo 2 - DeepMind's latest AI model. The upgrade cuts video creation time while improving realism in human movements and physics simulations.
What's New
Dream Screen's Veo 2 integration produces 1080p backgrounds from simple prompts like "busy Tokyo street at dusk." A new clip generator crafts standalone scenes without existing footage. The AI understands camera directions - specify "wide lens" or "film noir effect" for style control. Videos render 40% faster than previous versions while maintaining detail in hair strands and fabric movement.
Business Impact
Small teams can create product demos without filming crews - describe your item and get ready-to-post scenes. Two caveats: AI watermarks (SynthID) remain visible, and human review stays crucial - the model sometimes misinterprets complex prompts.
Currently live in US/Canada/Australia/NZ, with global expansion planned.
AI Investments - What to Watch1
Source: MadebyAgents via Grok2
Europe unveiled InvestAI - a €200 billion push to build four massive AI training centers. The plan targets "very large models" for medicine and science, offering startups access to 100,000 next-gen chips through public-private partnerships.
What's New
The €20 billion flagship fund creates CERN-like AI gigafactories - each housing four times more processors than current facilities. Unlike corporate-controlled US labs, these centers prioritize open innovation. A farmer's co-op could access the same supercomputers as pharma giants. The EU layers funding: low-risk grants cover basic infrastructure while private investors fund high-reward applications.
Business Impact
Startups gain free access to industrial-scale computing if they focus on climate, health, or manufacturing AI. The GenAI4EU program connects AI developers with 14 industries needing custom solutions - robotics firms get prioritized chip time. From 2025, SMEs can claim 30% tax rebates for using EU-made AI components. Language model builders access curated datasets from European health and transport systems.
Catch-Up Challenges
US firms still spend triple per AI project. The EU's ethics reviews add 6-8 weeks to product launches. Early applicants warn: proposals must use EU cloud services and avoid facial recognition tech. France's new gigafactory is about to open in 2026 Q3 - others follow through 2028.
AI Use Cases and Tools
🪄 Transform Video into Human-Like Articles: A Streamlined Process
1. Generate video transcript with NotebookLM → copy and save it
Output the entire transcript of the video.
2. Prompt to get a high-level outline/structure of the article → copy and save it
Act as an expert blog article writer specializing in SEO. Your task is to provide an outline and structure for an article based on the provided video transcript.
3. Go to gemini.google.com and select the 2.0 Flash Thinking Experimental with apps
It can access YouTube and Google search among other tools + has reasoning capabilities + free to use.
Act as an SEO expert specializing in keyword analysis. Your task is to search Google for the most relevant keywords that get good search volume and little to moderate competition. The keywords are for an article about <Article Title / Topic>. After you get the keywords rank them by potential and explain why you chose this order.
Here is a content outline of the article:
<Article Structure>
4. Copy the best keyword results
5. Choose an advanced LLM (e.g. Claude 3.5 sonnet, Deepseek R1 or ChatGPT 3o-mini)
6. Upload the context you copied previously
Act as an expert blog article writer with SEO specialization. Your task is to write an article based on the provided video transcript, following the content structure for the article and integrating keywords from the research results.
Try to keep most of the style from the video transcript.
The target audience is logistics business owners interested in AI and tech.
Maintain the tone of the video transcript.
Present the output as Markdown.
Organize content in sections using the provided content structure.
Try to reach a high readability inspired by Ernest Hemingway - short and clear sentences without oversimplifying the text and without replacing words that are typically used in that context.
Add placeholders in square brackets where I need to add content you don't get from the video transcript (for instance code examples).
I know you hit your output token limits. This is why I want you to split the output into several responses.
Split the article into 8000 token sections.
Add an overlap of 200 tokens between chunks.
Track page/section numbers.
Maintain table of contents.
Process section 1
7. After you get the first result send more prompts (”process section 2”, “process section 3”, etc.) to get the remaining sections.
8. Paste in WordPress or whatever CMS you use and add images, links, and code. Proofread and fine-tune.
9. Done!
This way I generated a pretty decent article within a few minutes with SEO optimization and high readability.
Notice the chunking I did in step 6? Necessary for Deepseek and Claude because they only support ~8000 output tokens. Otherwise you'll get a robotic summary with a lot of bullet points.
This way I converted my YouTube video to a blog article about AI web scraping.
MadebyAgents Updates
👨🏼💻 New Video Published: AI Coding Pro Tips (For Developers)
Whether you're using Cursor, Windsurf or VS Code with GitHub Copilot, without these 3 tricks the AI doesn't know what you really want. You end up fixing and correcting the code manually. I share my approach to slash those errors by 90+%.
That’s It for This Week
✨ Before You Go:
We’d love to hear what you think. Please share your opinion.
See you next time!
Tobias from MadebyAgents
 Tobias |
1 Disclaimer: The information shared reflects my personal opinions and is for informational purposes only. It is not financial advice, and you should consult a qualified professional before making any decisions.
Reply