Llama 4 is here and more (April 7, 2025 edition)

A wild weekend in AI — OpenAI pushed back GPT-5, Meta surprise-dropped three new Llama 4 models, and Microsoft decided it was time Copilot started remembering your dog’s birthday.

Let’s get into it.

🚨 GPT-5 Delayed (But There’s a Twist)

Sam Altman says GPT-5 is harder than expected — OpenAI is holding off a bit longer to integrate all the performance gains they uncovered mid-build. But here’s the upside: they’re splitting out two models originally meant to be internal — o3 and o4-mini — and releasing them soon.

Bonus: o3 has reportedly hit top-tier coder status in internal testing.

Also, GPT’s Deep Research tool may be coming to free users, and OpenAI’s legal battle with the NYT just got more serious — a judge is letting the copyright claims move forward after calling OpenAI’s defense a “straw man.” Yikes.

🦙 Meta’s Llama 4 Is Here (Yes, Three of Them)

Meta didn’t wait until Monday — they dropped Llama 4 over the weekend. Here’s the quick rundown:

• Scout – 10M token context, runs on one H100 GPU, best-in-class for code & summarization
• Maverick – 1M token context, beats GPT-4o & Gemini Flash in benchmarks
• Behemoth – Still training, but early results show it outperforms GPT-4.5 and Claude 3.7 in STEM

All of them use MoE (mixture-of-experts) architecture, but there’s a bit of controversy around benchmark results vs. public releases — some say Meta optimized specifically for test environments.

Notably: EU devs are blocked from using Llama 4 due to AI Act uncertainty.

🧠 Copilot Levels Up: Memory, Actions & Vision

Microsoft isn’t sitting still either. Copilot is getting major personalization features:

• Memory – Remembers past conversations and your preferences
• Actions – Can now book things online or complete tasks across sites
• Vision – Real-time camera input for mobile, plus desktop screen analysis
• Deep Research – Their answer to long-form AI investigations

It’s more like a personal assistant now — though it's still unclear if this makes users want to switch away from their current tools.

🎨 Midjourney V7, AI Brain Speech, & More Quick Hits

• Midjourney V7 – Entirely new image-gen architecture, now with Draft Mode for 10x faster ideation and voice input to generate visuals just by talking.
• AI Brain Speech – UC Berkeley + UCSF developed a model that turns brain signals into speech with just a 1-second delay. Game-changer for people with speech loss.
• AI Agent Cyberattacks? – Researchers warn that LLM-based agents are starting to mimic malware and adapt in real time. The line between helpful assistant and security threat is blurring fast.

✨ Tools Worth Checking Out

• Tana AI – Turn your notes into structured docs, tasks, or pages
• Runway Gen-4 – Finally some consistency in AI video!
• Clockwise – Smart scheduling that actually respects your focus time
• MoEngage Inform – AI-powered alert system to scale ops, not stress
• Readdy – Build websites by chatting, no drag-and-drop needed

That’s it for today — but if I had to leave you with one takeaway?

This week proved that benchmarks ≠ real-world results.
The best models on paper might not be the best in your product. Stay curious, test often, and don’t get swept up in the leaderboard hype.

Talk soon,
David

P.S. Anyone else feel like "GPT-5 delayed" is the AI version of “album pushed back but it's gonna be fire”? 🎧