Google's Laptop-Ready Gemma 4 & Meta's Global AI Agent

Google releases Gemma 4 12B, an open-weights multimodal model running locally on laptops. Meta launches Business Agent globally on WhatsApp. Plus: Coralogix raises $200M, GitLab cuts 14% of staff, DeepSeek slashes prices, mathematicians issue AI warning, and Anthropic shares agent security details.

Transcript

SPEAKER_00 0:01

Hey, welcome to AI Mornings with Andreas Vig. It's June 4th, 2026. Google DeepMind released a new open weights model yesterday that's worth paying attention to. It's called Gemma412B, and it's designed specifically to run on consumer laptops. The 12 billion parameter model does something interesting architecturally. It completely eliminates the separate encoders that most multimodal models use for vision and audio. Instead, images and sound flow directly into the language model backbone. The result is a model that needs just 16GB of VRAM or unified memory to run locally, yet benchmarks close to Google's larger 26 billion parameter model. It also has native audio input, which is a first for this size class. Released under Apache 2.0, so fully open for commercial use, the Gemma family has now crossed 150 million downloads. Meta also made a significant product move. Their AI customer support agent, now called Meta Business Agent, is available globally in WhatsApp business starting this week. They've been testing it for nearly two years in India and Mexico and now it's rolling out everywhere. The agent handles customer questions, product recommendations, appointment booking, and sales lead qualification. If it can't handle something, it reroutes to a human. It's also coming to Instagram DMs. Meta is testing additional features like daily chat briefings and market research capabilities. For larger businesses, they're building a platform for custom agents that can connect to Shopify, Zendesk, and other systems. Pricing is token-based for enterprises. In funding news, Coralogix just raised $200 million at a $1.6 billion valuation. The Series F comes just 11 months after their last round. They're betting big on AI agent monitoring, essentially building the observability layer for autonomous software systems. Their CEO made an interesting observation. Customers increasingly don't want to log into dashboards. They want to connect their LLM and ask what's wrong. Revenue grew over 60% this past year, and they now have about 30 customers spending more than a million dollars annually. The infrastructure layer around AI agents is clearly heating up. But here's the flip side of that AI infrastructure build-out. GitLab just laid off 350 employees, about 14% of their workforce, to restructure for AI workloads. Their CEO said agents are pushing competitors to the brink because they work at machine scale, stressing infrastructure that was never designed for that kind of load. GitLab is partnering with an AI lab to rebuild their platform specifically for agent-scale work. This continues a pattern we've seen across tech. Companies posting strong earnings while cutting jobs and citing AI. GitLab's Q1 revenue was up 23% to 264 million. Pricing pressure from Chinese labs continues. Deepseek permanently cut V4 pro pricing by 75%. Input tokens are now 43.5 cents per million, permanently 87 cents. That's roughly nine times cheaper than comparable Western models. The price war is clearly not over. Academia is starting to push back. 16 mathematicians published what they're calling the Leiden Declaration, an 11-page manifesto warning that unchecked AI threatens the values and culture of mathematics itself. Endorsed by the International Mathematical Union, it argues that math is fundamentally a human endeavor built on creativity and collaboration, values that clash with commercial AI development. The declaration raises concerns about AI-generated papers, overwhelming peer review, difficulty assigning credit for discoveries, and mathematicians' work being used to train military and surveillance systems. It'll be discussed at the International Congress of Mathematicians in Philadelphia next month. Anthropic published a detailed technical blog post on how they secure Claude across their three agentic products. It's unusually transparent about specific security incidents. One vulnerability allowed malicious config files to execute before users even accepted a trust prompt. Another red team exercise showed a researcher successfully phishing an employee into running a prompt that exfiltrated AWS credentials. Anthropic also revealed that 93% of permission prompts in clawed code get approved, which creates approval fatigue and defeats the purpose of human oversight. The post outlines three containment patterns: ephemeral containers, sandboxes with human in the loop, and local virtual machines. Worth a read if you're building agentic systems. Finally, the Instagram hack story from earlier this week has an update. Meta's AI support chatbot was tricked into linking attacker-controlled email addresses to target accounts, letting hackers reset passwords and takeover handles. Meta said Monday the issue was fixed, but users continued reporting account takeovers on Tuesday. Meta is now alerting victims and securing accounts, though they won't say how many were affected. The attacks targeted accounts with desirable short usernames, which get traded in gray markets. The vulnerability here was almost comically simple, just asking the chatbot nicely. That's all for today. See you tomorrow.