📺 Original Video: AI News: Everyone’s Leaving ChatGPT! by Matt Wolfe
📅 Duration: 38:02 | Published: March 6, 2026
AI News: Pentagon Drama, a ChatGPT Exodus, and Meta Being Creepy
The Week in 30 Seconds
- OpenAI dropped two models in one week (GPT 5.3 Instant and GPT 5.4). Most people won’t notice.
- Anthropic told the Pentagon “no” on surveillance and weapons. Got blacklisted for it. Then something wild happened.
- ChatGPT uninstalls jumped 295% in a weekend. Claude hit #1 in the app store. Yeah, really.
- Google shipped a fast new model and Notebook LM now makes actual animated videos
- Meta‘s AI glasses are sending footage to human reviewers in Africa. Lawsuits incoming.
The Stories
OpenAI Ships Two Models Nobody Asked For
OpenAI released GPT 5.3 Instant on March 3rd and GPT 5.4 two days later. Both landed in ChatGPT. ▶00:14
5.3 Instant is basically a “vibes update.” Fewer unnecessary refusals, less robotic tone. Matt’s take: for daily use, you probably won’t feel it. ▶02:21
5.4 is the bigger deal. Computer use built in, better coding, 1M token context window. But Matt thinks it’s really built for agents, not humans: “this model feels like it was made more for the agents than anything else.” ▶08:10
Google’s Gemini 3.1 Flash Light
Google put out a fast, cheap model for quick tasks and API calls. Matt built a YouTube thumbnail analyzer with it and found it “really fast” and “really cheap to use.” Regular users won’t see a huge difference, but developers will like the speed and pricing. ▶10:57
Notebook LM Makes Real Videos Now
This one’s actually cool. Google added cinematic video overviews to Notebook LM using Gemini 3, Nano Banana Pro, and VO3. We’re talking real animated motion graphics, not slideshow garbage. Matt calls it “one of my favorite features” and says it could “replace After Effects“ for quick animations. ▶12:54
The Pentagon Blacklist Bombshell
Here’s where the week got interesting. ▶16:48
The Pentagon told Anthropic to drop their red lines on US surveillance and autonomous weapons. Anthropic said no. The Pentagon designated them a “supply chain risk“ and blacklisted them.
Then OpenAI immediately swooped in and took the contract. Same type of work. And people noticed.
Matt thinks the blacklisting won’t stick: “this is like one giant crazy overblown negotiation tactic between the Pentagon and Anthropic.“ ▶27:25
But the fallout was real.
The Great ChatGPT Exodus
This is the big one. ▶22:23
After the Pentagon story broke, users started leaving ChatGPT in droves. The numbers are insane: Claude became the #1 most downloaded app. ChatGPT uninstalls surged 295% over a single weekend. Claude’s revenue is approaching a $20B run rate.
Businesses flipped from OpenAI to Anthropic almost overnight. Claude made it easy by offering free memory imports from ChatGPT. The complete market reversal happened in days, not months.
Turns out people care about where their AI company draws the line.
Meta’s Glasses Are Watching You
An investigation found that Meta‘s AI glasses send footage to human reviewers in Africa when privacy settings aren’t configured right. We’re talking bathroom footage. Credit card details. Matt’s reaction: “weird and creepy.” ▶31:00
Meta‘s now facing lawsuits in the UK, New Jersey, and California. Earned every one of them.
Quick Hits
Alibaba dropped Qwen 3.5, which can run on an iPhone offline. XAI updated Grok 4.20 beta 2. Microsoft released a 15B parameter reasoning vision model. Nvidia’s GTC conference is coming up. The pace right now is relentless. ▶29:02
Matt on the hype cycle: “I kind of don’t want to be in the game of trying to hype up every new model that comes out as the absolute best, amazing, coolest, rad, most impressive model you’ve ever seen that’s practically a step away from AGI.” ▶12:02
Respect.
Every Tool and Product Mentioned
Models:
- GPT 5.3 Instant (OpenAI) – “less cringe” vibes update
- GPT 5.4 (OpenAI) – computer use, 1M token context
- Gemini 3.1 Flash Light (Google) – fast and cheap
- Gemini 3, Nano Banana Pro, VO3 (Google) – powers Notebook LM videos
- Qwen 3.5 (Alibaba) – runs on iPhone, 800M-9B parameters
- Grok 4.20 beta 2 (XAI)
- Microsoft 54 reasoning vision model – 15B parameter, open-weight
Apps and Platforms:
- Box AI – content management with AI features
- Notebook LM – Google’s research tool, now with cinematic video
- Claude – Anthropic‘s chatbot, currently #1 in the app store
- Codex app (OpenAI) – simple IDE, now on Windows
- Spectre 1 (B&Audible) – anti-recording device, $1,000
What Matt Actually Thinks
On cherry-picked demos: “Obviously, I always think these demos are most likely the cherrypicked demos. They probably have some demos that look like complete crap, and these are the ones that came out the best.” ▶07:53
On the Pentagon drama: He thinks it’s political theater that’ll blow over. But the user reaction? That was real and immediate.
On model fatigue: Most new releases don’t matter for normal users. The real upgrades are happening in agent capabilities and developer tools, not chat quality.
The Takeaways
Ethics moved markets this week. The Pentagon situation proved that people care about more than benchmarks. When OpenAI appeared to compromise on safety, users left. Fast.
Model updates are hitting diminishing returns for regular users. Matt keeps saying it and he’s right. GPT 5.3, 5.4, Gemini 3.1 — they’re better on paper, but most people using ChatGPT for daily tasks won’t feel the difference.
Privacy fights are just starting. Meta‘s glasses scandal is a preview. AI devices that collect data without clear consent are going to face more lawsuits, more regulation, and more backlash.
Brand trust beats technical specs. Claude didn’t become #1 because it’s smarter than ChatGPT. It happened because Anthropic drew a line and stuck to it. That mattered more than any benchmark.
We try hard to get the details right, but nobody’s perfect. Spot something off? Let us know