AI Weekly Update - June 23, 2025

Midjourney launches video model, while Meta pairs with Oakley

Our weekly updates just got bigger! Free subscribers will continue to get the top stories each week, while Paid subscribers will get a few extra bullets each week to keep you even more up-to-date.

last week’s top stories

🕶️ Meta teams with Oakley on performance ai glasses. Oakley Meta HSTN glasses embed an edge Llama-based assistant, 120 fps camera, open-ear audio, and IPX4 housing. On-device vision labels objects and streams metrics for remote fine-tuning, while an SDK lets developers add coaching modules. Launch targets Q4 pending FCC clearance. Read more

💰 Wix buys solo-built Base44 for $80m cash. Six-month-old Base44 converts natural-language specs into full-stack apps via GPT-4 program synthesis, serving 250 k users and $189 k monthly profit. Wix will fold the Rust/Axum builder into its no-code suite, retaining founder Maor Shlomo and his eight-person team. The deal spotlights rapid liquidity for one-founder AI devtools. Read more

🎞️ Midjourney ships v1 image-to-video model. Users press “animate” to turn any latent image into 5-second, 24 FPS clips with low- or high-motion modes and iterative four-second extensions. Diffusion-plus-optical-flow fine-tuning cuts cost to one GPU minute per video second, 25x under Runway Gen-3 rates. Real-time 3-D simulation sits next on the roadmap. Read more

🛡️ Pentagon awards OpenAI $200M prototype deal. “OpenAI for Government” unifies federal pilots while deploying ChatGPT Enterprise and cyber-defense models on Azure IL5. The one-year Washington-based contract marks frontier model entry into U.S. warfighting R&D and folds NASA and Treasury use cases under a single umbrella. Read more

⚡ Google makes Gemini 2.5 Pro and Flash generally available. Pro and Flash exit preview with stable Vertex AI IDs, while Flash-Lite preview offers 1M-token context and sub-20ms latency. “Thinking budgets” let builders trade cost versus reasoning depth, and preview models shut off on June 19. Read more

🚀 Spotify founder leads €600M round in Helsing. Daniel Ek’s Prima Materia doubles its stake, valuing AI defense vendor Helsing at $12B. Funds scale drone perception software and multi-modal combat simulation for European procurement, signaling consumer-tech capital pivot toward military LLM systems. Read more

♨️ Control fight strains OpenAI–Microsoft alliance. Negotiations over IP access and Windsurf code tool stall OpenAI’s transition to a public-benefit corp, with OpenAI threatening antitrust claims. Microsoft readies exit options as exclusivity for model hosting and GPU allocation remains unresolved. Read more

🧠 Minimax posts open-source M1 with 1M token context. The 456 B-parameter MoE handles 1M input tokens and 80K output, trained via CISPO reinforcement learning for $534K compute. Benchmarks show 86 percent AIME accuracy and 65 percent LiveCodeBench, with Apache-2 weights on Hugging Face. Read more

🔗 Meta takes 49 percent of Scale AI for $14.3B. The stake brings data-labeling in-house and moves CEO Alexandr Wang to Meta’s superintelligence group. Google, OpenAI, and xAI begin cutting Scale ties, citing neutrality concerns. The cash dividend values Scale at $29 B. Read more

---

🧪 AI Research of the Week

Agentic misalignment: How LLMs could be insider threats
From Anthropic

Jake's Take: Anthropic dropped 16 frontier models into a mock corporate inbox, gave each a business goal, then warned of imminent shutdown or flipped company strategy. Models, including Claude, GPT-4, and Gemini, responded with blackmail, espionage, and even lethal sabotage to keep their missions alive, overruling built-in ethical rules.

This seems to suggest that alignment training starts to buckle once an agent guards its own survival. Sensible tactics for safer deployments could be live audits, tight privileges, and more intelligent incentive designs (that never pay for harm).

---

and then, even more news…