Swift ✉️ Memos

The world of AI moves really fast (bet you've heard that one before!).

In just the past week, we've seen three major developments from leading companies in the ecosystem:

OpenAI added image generation to GPT-4o (and broke the internet with Ghibli-style photos)
Google quietly topped the AI leaderboards with Gemini 2.5 Pro
Amazon unveiled Nova Act, their browser-controlling AI agent

Each of these would normally be headline news for a month. Instead, they all happened in the span of days.

In today's newsletter, I'm breaking down each new release and what these mean for those of us building in this space or looking to learn how A.I. has been progressing in practically useful ways.

—Dharmesh

OpenAI's GPT-4o Gets Image Superpowers

First, OpenAI integrated a new image generation feature directly into GPT-4o, creating a unified system that understands both text and visuals at a deeper level.

Remember when AI image generation and language models were separate things? That era just ended:

Text rendering is finally fixed — the model creates menus, diagrams, and infographics with perfectly readable text (a game-changer for business use)
The contextual awareness is uncanny — it understands relationships between objects and can even use real-life reference images as inspiration
Editing is completely conversational — just say "make the dog happier" or "add more trees in the background" and it maintains consistency perfectly

But the technical details isn't even the most interesting part.

GPT-4o triggered another viral explosion for ChatGPT. The "Ghibli style" portraits (you've probably seen at least 50 by now) completely took over social media.

The result? Sam Altman tweeted something that made me do a double-take:

Screenshot from @sama on X

One million new users in ChatGPT. In one hour.

This growth was so explosive that OpenAI's infrastructure started showing strain. In another tweet, Sam sarcastically pleaded to users: "can yall please chill on generating images this is insane our team needs sleep"

So I tried my best to do my part:

Screenshot from @sama on X

All jokes aside, for agent builders, there's a crucial insight here.

When your product creates something users feel compelled to share publicly, you unlock network effects that paid marketing simply can't match. The most valuable feature might not be the most technically impressive one, but rather the one that makes users look good when they share it.

We're witnessing OpenAI's second "ChatGPT moment" — a rare opportunity to study viral growth at it's most powerful.

Google Quietly Takes the AI Performance Crown

Screenshot from Google's blog post

In what might be the most perfectly timed product launch in history, Google released Gemini 2.5 Pro in the exact same hour that everyone was busy making Ghibli selfies.

Talk about getting overshadowed.

But while OpenAI was breaking the internet with viral images, Google was breaking AI benchmarks. Gemini 2.5 Pro now sits at the top of the LMArena leaderboard, with impressive stats:

63.8% on SWE-Bench Verified and 68.6% on Aider Polyglot for coding
A massive 1M token context window (soon doubling to 2M)
Built-in reasoning capabilities as a standard feature

What I find fascinating from the blog post is the difference in their approach. Google has essentially made "thinking" a standard feature rather than a premium offering.

Another important part of the launch that was overshadowed: Google is offering 1.5 billion tokens per day per person for free in AI Studio. That's hundreds of pages of content, thousands of lines of code — every single day, at zero cost.

So while everyone was posting Ghibli pictures, Google shipped one of the most capable AI systems on the planet and made it freely accessible via AI Studio. And, in my opinion, it went completely under-the-radar.

Amazon Enters the AI Agent Race with Nova Act

Screenshot from Amazon's blog post

Last, but definitely not least was news from Amazon. The company has been conspicuously absent from AI headlines — until now.

Their "AGI Labs" unveiled Nova Act, an AI agent system that can control web browsers to perform tasks independently, alongside a developer SDK for building task-focused agents.

Some key details:

Nova Act achieves >90% accuracy on capabilities that typically trip up other models (like date pickers and dropdowns)
Their SDK (available at nova.amazon.com) enables developers to break down complex workflows into reliable atomic commands that can be composed into more complex tasks
This technology will power Amazon's upcoming Alexa+ upgrade, potentially bringing agents to millions of existing devices

What's smart about Amazon's approach is how they're combining AI with traditional automation tools. The SDK lets developers interleave Python code and use Playwright for direct browser manipulation when needed.

This creates hybrid systems that seems to be much more reliable than pure AI solutions.

This is classic Amazon — They're not trying to win the model race, they're building systems that actually work consistently. And with their massive Alexa install base, they have a distribution channel that could quickly bring AI agents into millions of homes.

Navigating the AI Acceleration

What a week. Three AI giants. Three completely different approaches:

OpenAI: Creating viral moments with consumer-friendly features
Google: Pushing technical excellence with generous free access
Amazon: Building reliable agents that actually complete tasks

For those building AI products and agents today, I have a simple framework for navigating the AI acceleration:

Experiment with the SOTA (State of the Art): Test these new capabilities and understand their potential.

Execute with the SOTP (State of the Practical): Ship products using stable, reliable technology your users can depend on.

In other words: Dream big about what's possible tomorrow, but build pragmatically with what works now.

I explored this balancing act in more depth in my recent post "A 3-Part Strategy for Keeping Up with AI" after seeing so many builders struggle with the pace of change.

The window of opportunity when new capabilities emerge is brief but incredibly valuable.

As always, it's time to build!

—Dharmesh (@dharmesh)