What I Learned Last Week #8
China makes AI education mandatory. Google launches prompt engineering guide, A2A protocol, and Firebase Studio. GPT-4.5 passes Turing. Vibe coding security issues. GDPR to be less stupid. And more...
It’s impossible to keep up with everything new in AI. Tell me if I missed anything impactful or engaging, and I will add it to the next issue. For now, here’s what I learned last week:
China will win long-term by making AI education mandatory in schools.
Google published a damn good 69-page guide to prompt engineering.
Shopify’s CEO declares AI skills mandatory for all employees.
Google launches agent-to-agent protocol, complementing MCP.
GPT-4.5 passed the Turing test when given a persona.
How AI is rewriting startup funding and growth models.
AI adoption is hindered by mediocre user experience.
Vibe coding presents security risks because AI invents nonexistent packages.
Google is on fire: launches Firebase Studio in response to vibe coding tools.
GDPR to be revised to be less stupid and allow more innovation.
Meta’s latest model, Llama 4, is considered quite lame by users.
Despite AI taking over coding, Microsoft wants to fire a lot of non-coders.
China Makes AI Education Mandatory in Schools
Tariffs won’t bring world supremacy. The real game for getting ahead is education. Starting this fall, China is making AI education mandatory in Beijing schools, from six-year-olds to high school students. Schools can choose to integrate AI into existing subjects or teach it separately. The goal is to foster innovation and tech skills, positioning China as a leader in AI. This comes at a time when many other educational institutions are banning LLMs for student exams and assignments. Read more →
Google's Masterclass in Prompt Engineering
Google just dropped an extensive 69-page guide to prompt engineering that's worth its weight in tokens. The document breaks down how to effectively communicate with LLMs through techniques like zero-shot, few-shot, and chain-of-thought prompting. What caught my eye is their emphasis on iterative testing and task decomposition — approaches I've found critical when working with these models. For anyone serious about squeezing better results from AI systems, this white paper offers both theoretical foundations and hands-on best practices. No more guessing what makes AI tick. Read more →
Shopify Makes AI Fluency Mandatory for All Employees
Shopify CEO Tobi Lütke has officially announced that AI skills are a non-negotiable expectation for all employees. In a leaked internal memo, Tobi states bluntly: 'I don't think it's feasible to opt out of learning the skill of applying AI in your craft.' The company is requiring teams to explore AI solutions before requesting additional resources and providing extensive AI tools across departments. When do you think your company will be ready to send out an email like that? Read more →
Google Launched A2A, A Universal Language for AI Agents
In parallel with Anthropic’s MCP, Google has introduced Agent2Agent (A2A), an open protocol designed to enable AI agents to communicate and collaborate seamlessly across different platforms and vendors. A2A promises a significant shift towards AI systems that can work together autonomously, regardless of their underlying framework or vendor. Read more →
GPT-4.5 Passes Turing Test, When Given a Persona
A few years ago, while working on Eternime, I realized that the passing of the Turing test would not be a specific moment but a series of disputed events. This past week, one of the ignored pieces of news was that OpenAI's GPT-4.5 has passed a three-party Turing test with a 73% success rate when instructed to adopt a specific persona. What's more striking? The AI was deemed more human-like than actual humans in the experiment. However, without persona prompting, its success rate dropped to 36%. Read more →
How AI is Rewriting Startup Funding Playbooks
Traditional funding models are being disrupted as AI reshapes how startups scale. I've been tracking an emerging trend called 'seed-strapping'—where founders raise one strategic round and then scale profitably—gaining momentum as AI enables smaller teams to generate revenue faster with less capital. The data is striking: while 75% of VC-backed startups never return capital and founders typically retain just 15% ownership by Series C, AI-augmented teams are demonstrating more sustainable growth paths. Read more →
The Importance of User Experience to AI Adoption
While Shopify's CEO recently told employees to use AI or face poor performance reviews, AI adoption is lagging despite clear productivity benefits. It isn't about intelligence; it's about usability. We're dismissing a crucial lesson: successful tools make complex technology accessible through thoughtful design. True mass adoption won't come from CEO mandates or more powerful models—it'll emerge from interfaces that make machine intelligence truly accessible. Read more →
The Dark Side of Vibe Coding: When AI Invents Software Dependencies
AI code assistants have a troubling habit: inventing software packages that don't exist. Researchers found commercial models hallucinate non-existent packages 5.2% of the time, while open-source models do it 21.7%. What's alarming is that attackers are now exploiting this by creating malicious packages with these hallucinated names and uploading them to repositories like PyPI or npm. Security experts call this 'slopsquatting.' Even more disturbing—sometimes Google's AI search results validate these fake packages, creating a perfect storm of AI systems reinforcing each other's mistakes. Read more →
Firebase Studio Is Google's All-in-One Workspace for AI App Development
Everyone was worried that Google was too late to the AI game when OpenAI launched ChatGPT. And they did struggle to catch up, but now they are on… fire! Google just unveiled Firebase Studio, a cloud-based development environment. What caught my attention is how it streamlines the entire AI app workflow — developers can sketch ideas with prompts, refine with Gemini-powered chat, dive into code when needed, and deploy with one click. I find it particularly interesting that they've designed it for both no-code enthusiasts and experienced developers who want VM-level customization. Check it out →
GDPR Will Be Revised to Be Less Stupid
The European Commission President Ursula von der Leyen wants to simplify GDPR rules to help companies compete better globally, especially against the U.S. and China. While a landmark privacy law, GDPR has been criticized for being costly and complex for businesses to navigate. Read more →
Meta’s New Llama 4 is Lame
While Meta’s benchmarks brag that their latest model outperforms Gemini 2.0 Pro, DeepSeek V3, and Claude 3.7 Sonnet, early community feedback is not that exciting. People are disappointed with Llama 4’s bad performance in areas like reasoning, coding, and long context comprehension. Meta is also blamed for shady practices, like mixing benchmark test sets in the post-training data or announcing models that nobody can test yet. Read more →
Microsoft Targeting Non-Coders in Latest Layoffs
Microsoft is eyeing more job cuts this May, specifically targeting middle managers and non-coding staff—a stark contrast to predictions that AI would replace programmers. The company aims to increase its ratio of product/program managers (PMs) to engineers—from around 5.5 engineers to one PM to ideally reach 10:1. Read more →
I hope you find the above things as interesting as I did.
Cheers to a new week and new adventures!
Mulțumesc, Marius!