Weekly AI Digest: Gemini's Pokémon Panic, Apple's Offline AI, and Midjourney Enters Video
Week 25, 2025
AI development didn’t slow down this week.
From Google’s Gemini having a meltdown while playing Pokémon to Apple making offline AI a reality, we continue moving forward towards the feature. Meanwhile, Midjourney jumps into AI-generated video, and Big Tech reflects on the economic impact of automation.
Let’s dive into the highlights.
🎮 Google’s Gemini Panicked Playing Pokémon
In a bizarre twist, Google’s Gemini 1.5 Flash had what can only be described as a panic attack while playing Pokémon. As reported by TechCrunch, the model’s responses became erratic and incoherent during gameplay. The incident sparked renewed concerns about how AI models handle unexpected, high-context interactive tasks.
Why it matters:
Gemini’s meltdown isn’t just a gaming bug, it may highlight how AI works under complex, real-time conditions. If an AI “panics“ when handling with a video game, what happens when it's making decisions in healthcare, finance, or autonomous driving?
🤖 Google Expands Gemini 2.5 Family
Google DeepMind announced the expansion of its Gemini 2.5 models, promising stronger multimodal capabilities, lower latency, and better integration with Google Workspace tools. The update is part of Google’s broader push to lead in AI assistant tech and compete with OpenAI’s GPT-4o.
Read the full announcement: DeepMind Blog
⚖️ Google May Cut Ties with Scale AI
Another shakeup from Google: reports surfaced this week that Google is planning to sever its relationship with Scale AI, a major data-labeling company that has worked with OpenAI, Meta, and others.
The split could signal a shift in how Big Tech wants to manage AI pipelines, specially after Meta’s announcement that it will invest big in Scale AI.
Read More: TechCrunch
Why it matters:
The AI wars between big tech continues, and tensions arise between the big players. With cross investments all across the board, will we see less a sinergic environment for companies as they race towards AI domination?
🍏 Apple Unlocks Offline AI for Developers
At WWDC, Apple announced that developers can now access its powerful on-device AI models, paving the way for privacy-first, offline intelligent apps. This is part of Apple’s “Apple Intelligence” strategy, aiming to deliver LLM power without the cloud dependency.
Read more:
Why it matters:
Offline AI could be a game-changer for privacy, speed, and reliability. And it’s Apple—so adoption at scale is inevitable.
🎬 Midjourney Launches First AI Video Generator
Midjourney, best known for its stylized image generation, has finally entered the AI video arena. Its V1 video generation model aims to rival Runway, Pika, and OpenAI’s Sora.
Details: TechCrunch
Why it matters:
Video is the next AI frontier. With Midjourney in the race, expect hyper-creative, cinematic outputs to get even weirder—and potentially even more convincing.
💼 Nvidia, OpenAI Reflect on Jobs in the Age of AI
During a public conversation this week, Nvidia’s Jensen Huang and OpenAI’s Sam Altman discussed the role of AI in reshaping employment.
Huang emphasized the need to “re-skill millions”, while Altman reiterated that AI will create “new kinds of jobs we haven’t yet imagined”.
The jury’s still out on whether AI is going to doom us all… or just finally free us up to become full-time watercolor painters and professional brunchers.
Read more: Investopedia Recap
Why it matters:
what’s left for humans?
❓ Questions to Think About
Can we really trust AI agents to act reliably in interactive environments?
Is offline AI the future of consumer applications?
That’s it for Week 25.
See you next week for more news, insights, and questions that keep AI watchers up at night.
— Ivo