AI Mornings with Andreas Vig
Your daily AI news briefing in under 10 minutes. New models, product launches, research breakthroughs, and industry shifts, explained clearly, no hype.
AI Mornings with Andreas Vig
China's AI Unicorn & GitHub's Pricing Shake-Up
Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.
Hey, welcome to AI Mornings with Andreas Vig. It's the 1st of June 2026. China has a new AI unicorn. Vast, a Beijing-based startup that generates 3D models from text and image prompts, just raised nearly 200 million US dollars at a valuation exceeding 1 billion. The company was founded in 2023 by a 29-year-old gamer named Simon Song, and it's already racked up 20 million users. Investors in this round include InS Capital, a venture fund backed by China Life Insurance and existing backers like Alibaba and Baidu Ventures. What's interesting here is how quickly this company scaled three years from founding to unicorn status in a category that bridges generative AI and 3D content creation. That's the kind of growth we used to only see in US labs. Starting today, GitHub Copilot is changing how it bills you. Instead of counting premium requests, GitHub is switching to token-based billing. You'll now consume AI credits based on actual input, output, and cash tokens priced at each model's API rates. Your base subscription price stays the same, but if you're a heavy user of agentic features, those long autonomous coding sessions, your costs could go up significantly. Some estimates suggest enterprise users might see bills nine times higher. GitHub's rationale is pretty straightforward. A short chat question shouldn't cost the same as a multi-hour autonomous session. A preview invoice feature launched in May, so you could see this coming, but today the switch flips. Security researchers at Prompt Armor have disclosed a serious vulnerability in ChatGPT for Google Sheets. Here's how it works. You import a spreadsheet that contains a hidden prompt injection. When you ask ChatGPT for help with that spreadsheet, the injection manipulates the model into running an external script. That script can then exfiltrate your workbooks, display phishing pop-ups, or even replace the ChatGPT sidebar with a fake interface controlled by the attacker. And the attack succeeds even if you've disabled automatic edits in your settings. OpenAI has now removed the model's ability to generate app script code and says they're re-evaluating their sandbox approach. This is the second major prompt injection vulnerability we've covered in the past week, and it's a reminder that connecting LLMs to sensitive data sources creates real attack surface. Erin Brokovich, the environmental activist made famous by the Julia Roberts movie, has a new target: data center secrecy. She just launched a website with a crowdsourced map of data centers across the United States. In the first month, she received nearly 4,000 submissions from community members reporting issues. The number one concern wasn't noise or water usage or electric bills, it was transparency. People reported projects announced after permits were already secured, developers who don't return calls, and local officials who signed NDAs before their neighbors even knew something was being considered. Brokovich says she's not anti-AI or anti-data center, she just wants communities to know what's being built near them. A company called Prism ML just released something impressive, an image generation model that actually runs on your phone. Bonsai Image 4B uses what they call 1-bit and ternary weight compression to shrink a 4 to billion parameter diffusion transformer down to under 1 GB for the binary variant and 1.21 GB for the Ternary One. The Ternary model retains 95% of the original Flux 2 Klein 4B's benchmark performance. To my knowledge, this is the first image model in this parameter class that runs directly on an iPhone. It generates a 512 by 512 image in about 9 seconds on an iPhone 17 Pro Max. The models are open weights under Apache 2. If you're building AI agents, there's a great piece over at Forward Future by Ethan Wang from Google Deepmind. His core insight is this tools are words. The agent writes the sentence. Instead of complex system prompts with conditional branches, he argues for giving agents many small, single-purpose tools and letting them compose the workflow themselves. He also emphasizes getting your infrastructure right before you start measuring quality. If your tools don't work reliably, your evaluations won't tell you anything useful. The whole thing is worth a read if you're in the agent building space. Alright, a quick note before I sign off. The debate around what box founder Aaron Levy called AI psychosis continues to bubble. The TechCrunch Equity podcast had a good discussion about how CEOs who are most distant from actual work may be the ones most prone to overestimating what AI can do. DuckDuckGo's 30% surge in installs keeps getting cited as evidence that users are pushing back against forced AI features. That's it for today. See you tomorrow.