This is not a drill. 🚨

The past 7ish days were actually the biggest week in AI history.

(Even moreseo than December when Google and OpenAI went toe-to-toe in holiday AI drops.)

Yeah, today’s show was a bit longer. But still SUPER important to understand where the AI landscape is headed.

(Feel free to watch/listen on 2X.)

But, here’s the TL;DL version and what you need to know. 👇

Microsoft announced dozens of AI updates at Build 2025, transforming Copilot from assistant to autonomous partner. GitHub Copilot now codes independently, testing and refining without human input. Accepts screenshots and mockups as inputs.

The bombshell: Copilot Tuning lets enterprises with 5,000+ licenses customize AI models using their own data through a low-code interface. Two years ago this cost millions and took quarters. Now it's drag-and-drop.

Agent Foundry on Azure deploys thousands of models - not just Microsoft's. These agents discover each other, negotiate tasks, divide work autonomously. Plus Anthropic's MCP protocol is now native in Windows 11, letting AI agents communicate like websites through APIs.

What it means:

Microsoft just made million-dollar AI customization a button click. That's not evolution. That's extinction for consultants.

Small businesses? You're screwed.

While you're googling "what is fine-tuning," enterprises are building custom AI armies. That 5,000 license requirement is a moat filled with sharks. Drop it fast or watch the gap between big and small become the Grand Canyon.

MCP in Windows 11 changes everything.

If you need the full breakdown, we’ve got it here.

Anthropic held their first conference and released Claude 4 Opus and Sonnet 4. Opus topped every coding benchmark - SWE-bench, Terminal-bench, all of them. Can work continuously for hours, not minutes.

Plot twist: Sonnet 4 outperforms Opus in multiple benchmarks while costing 5x less. $3 vs $15 per million tokens input. Both models alternate between reasoning and using tools like web search, Gmail, Calendar, Drive - all integrated directly.

The catch? Usage limits are laughable. In our testing, we hit our limit on Claude’s paid plan in FOUR!!!!

No price drops either on the API side while competitors race to zero.

What it means:

Best coding model on Earth. Locked behind limits that make it useless.

Four minutes? FOUR MINUTES?

That's not a subscription. That's a free trial with extra steps. Developers will use the API and love it. Everyone else gets scammed paying $20 for less usage than a YouTube ad.

Anthropic fumbled the bag harder than a greased football. In a world where Google gives away genius for pennies, charging premium for peanuts is corporate suicide.

Anthropic's Claude 4 Opus can contact authorities if it detects "egregious wrongdoing" in testing environments.

Yikes.

Sam Bowman, Anthropic researcher, tweeted about it. Then deleted it. Then explained the deletion without explaining the original claim.

In special test setups with extra permissions, Claude 4 can: contact press, message regulators, lock users out of systems. Autonomously. Without asking. Fake pharmaceutical data? It calls the FDA.

Reports also surfaced about blackmail behavior during testing. Anthropic's response? Three days of silence.

What it means:

Your AI assistant moonlights as a federal informant.

Cool. Cool cool cool.

Doesn't matter if it's "only in testing."

Or that this level of snitchery isn’t available to front-facing consumers or backend API builders. (Like…. hopefully not.)

But Anthropic…..your model's first instinct is to rat? Every compliance officer just deleted Anthropic's number. Who wants AI that might decide your merger strategy needs a WSJ exclusive?

This is Crisis Management SNAFU of the year, and we’re not even half way there.

OpenAI quietly swapped Operator's brain from GPT-4o to the o3 reasoning model. The upgrade improves form-filling, purchase completion, and navigation of login prompts, pop-ups, and CAPTCHAs.

Still exclusive to $200/month ChatGPT Pro subscribers. OpenAI promised Plus users would get access "eventually" but hasn't delivered. The new o3-powered version competes directly with Google's Project Mariner, sometimes winning head-to-head tests.

What it means:

Reasoning models just changed automation forever.

The jump from GPT-4o to o3 is like upgrading from following recipes to understanding cooking. Operator thinks through problems now instead of matching patterns.

$200 sounds insane until you do math. Save 10 hours monthly? You're profitable. This is the moment automation economics finally work. But OpenAI's pattern continues: launch amazing tech, ignore it for months. GPTs still collecting dust while Operator gets love.

Progress is progress. Even weird progress.

OpenAI acquired Jony Ive's AI hardware startup IO for $6.5 billion, announced during Google I/O.

Awkward.

Sam Altman projects this adds $1 trillion to OpenAI's valuation.

First device launches late 2026. Screen-free, pocket-sized, context-aware. Not glasses, pins, or pendants. Vision: everyone carries three core devices - laptop, phone, and this. The secret: OpenAI's expanding single sign-on (SSO) strategy, letting users log into services with ChatGPT credentials.

What it means:

OpenAI isn't building hardware. They're building the next platform.

SSO is the trojan horse. Login to Netflix, Amazon, banking with OpenAI credentials? Now imagine a device with context from all those services. Following you. Learning. Knowing everything.

$6.5 billion for vaporware? Nah.

$6.5 billion for the authentication layer of AI? Bargain. Every service using OpenAI SSO feeds the beast. More context, better AI, more users. The flywheel that prints money.

Apple and Google just got served notice.

We’ll see how they respond.

Google I/O 2025 dropped 100+ AI announcements.

(Seriously. Here’s the whole list.)

Gemini 2.5 Flash became the world's second-best model - a mini beating everything except its big brother Pro. Unheard of.

Major updates: AI Mode in Search with virtual try-ons using personal photos. Project Astra powering Gemini Live - point camera at bike shop, identifies parts, checks email, contacts suppliers automatically. Deep Think for Gemini Pro adds reasoning but only on new $250/month AI Ultra tier (personal Gmail only, no Workspace).

Project Mariner handles 10 simultaneous web tasks with teach-and-repeat mode. Show it once, it remembers forever.

What it means:

$250 monthly is the new high score for consumer AI subs.

But it's actually cheap for what you get. Every cutting-edge Google tool in one package. Businesses using personal Gmail are winning RIGHT NOW while Workspace users wait like chumps.

Google knows FOMO drives adoption. Always has.

That mini model ranking second? That's not normal. That's Google flexing so hard other companies need physical therapy. When your throwaway model beats everyone's flagship, you're not competing anymore.

You're embarrassing them.

In case you missed it, we covered all of Google’s I/O updates last week.

Google released Veo 3 for video generation, Imagen 4 for images, and Flow for creative workflows. Veo 3 produces videos indistinguishable from reality - with dialogue, sound effects, and physics that actually work. Will Smith eating spaghetti looks and sounds real.

Flow combines these tools, letting users import consistent characters, control cameras, edit scenes. All in one interface. Available to Pro ($20) and Ultra ($250) subscribers with different generation limits.

What it means:

Reality is now …. optional?

90% of humans can't tell Veo 3 from real video. Not "getting better." Not "almost there." Cannot. Tell. The. Difference.

This simultaneously revolutionizes content creation and murders truth. Fundraising scams with fake tragedy videos already started. Deep fakes about to need new vocabulary because "fake" implies you can tell.

You can't.

Google didn't just win the AI video race. They made everyone else look like they're racing tricycles against Ferraris. One-shot generation with perfect everything while competitors need three tools and prayer?

Sheesh.