Hey friends,
Today’s newsletter is packed. Between Amazon clawing back relevance, Stanford and NVIDIA building minute-long AI cartoons, and Google turning Gemini into a legit research assistant, it feels like we’re watching the next AI wave take shape — one that’s less about novelty, more about actual utility.
Let’s jump in.
🎥 AI Learns to Tell a Story (Finally)
After years of choppy, forgettable AI video clips, NVIDIA and Stanford just raised the bar with Test-Time Training — a new technique that stitches together consistent, minute-long animations.
The model uses neural networks as a kind of working memory, and it shows. In demos with Tom & Jerry-style shorts, scenes flow logically and characters stay recognizable. It’s still early, but this is a real step toward actual AI filmmaking.
🔊 Amazon’s New Audio + Video Stack = Not Playing Around
If you thought Amazon was sitting out the GenAI race, think again.
They just dropped:
- Nova Sonic — A voice model that beats OpenAI on latency and clarity in noisy settings
- Nova Reel 1.1 — A video model that now generates 2-minute clips, with multi-shot control
Both models are live in Bedrock, and they’re cheap — Sonic clocks in at around 80% less than OpenAI’s equivalents. Combine that with their agentic browser tool (Nova Act) and Alexa+ rollout, and Amazon suddenly feels like a very real player again.
🔎 Gemini 2.5 Pro Becomes a Full-Fledged Researcher
Google’s Gemini Advanced users can now leverage an AI researcher by trying out “Deep Research” — a new mode that synthesizes sources, adds audio summaries, and integrates cleanly with NotebookLM and AI Mode. It’s positioned less as a chatbot and more as an end-to-end knowledge worker.
Other new upgrades:
- Visual input recognition (you can upload images for context-aware answers)
- Geospatial reasoning (think satellite imagery analysis + urban planning insights)
- Audio Overviews (podcast-style summaries of long-form research)
It’s one of the first AI tools that actually feels like it could replace some human workflows instead of just summarizing them.
🧠 Murati’s Secret Lab Assembles the OGs
Mira Murati’s startup, Thinking Machines, now has nearly half its team composed of OpenAI alumni. New additions include:
- Bob McGrew, ex-Chief Research Officer at OpenAI
- Alec Radford, co-creator of GPT
No one knows exactly what they’re building yet — but with this roster, it’s either going to be amazing… or a very expensive ghost ship.
📊 Quick Hits
- Meta’s Llama 4 is great at math and reasoning, but struggles with long-context tasks.
- NVIDIA’s new open model (Nemotron Ultra) beats DeepSeek R1 and Llama 4 at half the size.
- Together AI’s DeepCoder-14B is open-source and rivals top closed-code models.
- Waymo might use in-car camera footage for training + ads, prompting privacy alarms.
- ElevenLabs’ MCP server now enables real-time AI voice agents across platforms (we use them too!).
- Perplexity just launched a $5K credit program for early-stage startups.
⚒️ 5 New Tools Worth Checking Out
- Stepsailor – Turn your product into a guided learning experience.
- Recall – Your AI-powered second brain for organizing scattered ideas.
- ezsite – One-click AI website builder (zero-code).
- Midjourney v7 Alpha – Better coherence, smoother image generation.
- Supaboard – Secure, no-code dashboard builder.
If this helped you cut through the noise, forward it to someone who’s been asking “What’s real in AI right now?”
Or hit reply and tell me: what’s one tool or idea from today’s newsletter you’re actually going to try?
Talk soon,
David