27 Aug 2025

Photo Editing Banainas!; Plug-In Agent; Teachers & AI

🍌 Google's Gemini 2.5 Flash Image is Bananas!

This week, the AI world went absolutely bananas for... well, nano-banana! This model, which has been dominating the image editing leaderboards in secret, was finally unveiled as Google's Gemini 2.5 Flash Image. And ya, you can believe the hype.

For what feels like an eternity, AI image generators have struggled with a simple task: keeping a character's face consistent from one image to the next. You'd ask for a picture of a CEO in a boardroom, and then in a coffee shop, and you'd get two completely different people. Those days seem to be over.

Gemini 2.5 Flash Image has cracked the code on character consistency. You can now create a character and place them in a variety of settings - a desert, underwater, a disco - and they'll look like the same person every time. This is a boon for everything from marketing campaigns to creative storytelling.

But wait, there's more! This model is packed with features that are set to completely change the way we create and edit images:

Multi-step Editing: You can layer on changes with natural language prompts, and the model will maintain consistency throughout the process.
World Knowledge: Unlike other image models, Gemini 2.5 Flash Image has a grasp of the real world, allowing it to make smart choices about the content it generates.
Multi-image Fusion: You can seamlessly merge multiple images and drag and drop objects between them.

And the best part? It's not too pricey at just $0.039 per image. This competitive pricing is already shaking up the market, with major players like Adobe integrating it into their Firefly platform.

This move signals a shift in the AI landscape, as it seem we're moving away from the "one model to rule them all" mentality and into an era of multi-model ecosystems. As The Neuron puts it, Adobe is becoming the "Switzerland of AI creativity," offering a buffet of models for creators to choose from.

Why this matters: The future of AI isn't about a single, monolithic model. It's about having the right tool for the right job. And with Gemini 2.5 Flash Image, Google has just given us a very powerful new tool.

Read more: The Rundown, AI Secret, The Neuron

🏢 Enterprise AI: A Reality Check

An MIT study (flawed, in our view; misread by media) highlighted by AI Secret, found that a 95% of enterprise GenAI deployments don't actually move the P&L needle.

So, what's the problem? It's not the models themselves, but the clumsy way they're being integrated into old, clunky workflows. As AI Secret puts it, many companies are just "stapling ChatGPT-like tools onto old workflows". This has led to a lot of expensive demos dressed up as strategy, and not a lot of real-world results.

The 5% of projects that are seeing "rapid revenue acceleration" are the ones that laser-focus on a single pain point and execute ruthlessly.

Trouble here of course is we're not thinking "AI Native" and also expecting too much too soon. Companies should be experimenting like mad, and with that will come many dead ends. But in product development the learnings from reaching those dead ends are in some ways the point.

Get a grip, MIT.

Read more: AI Secret

🤖 Claude Plug-In Supercharges Chrome

Get ready for a hands-free future, because AI is coming for your browser. Anthropic is leading the charge with its new "Claude for Chrome" extension, a research preview that gives the AI assistant agentic control over your browser.

Claude can navigate websites, click buttons, fill out forms and even handle tasks like scheduling meetings and drafting emails. The goal is to create a seamless and intuitive user experience where you can simply tell your AI what you want to do, and it does it.

Rightly so, Anthropic is taking a cautious approach, with a strong focus on security. The main concern is prompt injections, where malicious instructions can be inserted into web content to hijack the AI. To combat this, Claude is being equipped with a range of safety mitigations and permissions.

As these agentic systems become more sophisticated, they have the potential to fundamentally change the way we work and live. But as the saying goes: proceed with caution.

Read more: The Neuron

💼 AI and Employment: Disruption Continues

The debate over AI and jobs is heating up, with two major studies painting a complex and somewhat contradictory picture.

On the one hand, a Stanford study highlighted by The Neuron found that AI has caused a 13% job decline for young workers since 2022. This may be a stark reminder of the disruptive power of AI, and it underscores the need for proactive measures to support workers who are being displaced.

On the other hand, the MIT study mentioned earlier suggests that the impact of AI on the workforce may be more nuanced at the moment. While some jobs are being automated, the study also found that many companies are struggling to get real value from their AI investments. This suggests that the transition to an AI-powered economy may be slower and more complex than some have predicted.

All this could add up to our being in a period of profound transition. Some jobs will be lost, but new ones will be created. The key for workers will be to adapt, adapt, adapt.

Read more: The Neuron

📚 The Other Side of the Desk: How Teachers Are Really Using AI

While everyone's been talking about students using AI to cheat on homework, Anthropic just dropped some fascinating insights into what's happening on the other side of the classroom. Their new report analyzed 74,000 conversations from educators using Claude, and the results are eye-opening.

Here's the breakdown of how professors are actually putting AI to work:

Curriculum Design Takes the Crown (57%): Teachers are using AI most heavily for creating lesson plans, designing course materials and structuring educational content. This makes perfect sense - it's creative work that benefits from AI's ability to generate ideas and organize information.

Research Support Comes Second (13%): Academic research is getting a boost from AI, helping educators dig deeper into their subjects and stay current with developments in their fields.

Student Work Evaluation (7%): This is where things get interesting—and controversial. While only 7% of conversations focused on evaluating student work, it's become the most polarizing use case.

The grading controversy is particularly fascinating. Despite being rated as AI's weakest capability, a whopping 49% of assessment conversations showed heavy automation. Teachers are clearly struggling with workload, but they're also pushing AI into areas where it might not be ready to deliver reliable results.

The Custom Tools Revolution: Perhaps most impressively, professors are building sophisticated custom tools using Claude's Artifacts feature. We're talking interactive chemistry labs, automated grading rubrics and visual dashboards. These educators aren't just using AI - they're becoming quasi AI developers.

The Automation Paradox: Teachers are happy to automate the boring stuff -financial planning, record-keeping administrative tasks. But when it comes to the core of education - teaching and advising - they want to keep the human touch. This suggests a mature understanding of where AI adds value versus where human judgment remains essential.

Why This Matters: This research flips the script on the usual "students vs. AI" narrative. It shows that educators are thoughtfully integrating AI into their workflows, but they're also grappling with the same challenges everyone faces: figuring out where AI helps and where it hurts.

The classroom of the future won't be about humans versus machines - it'll be about humans and machines working together. And based on this data, teachers are already figuring out how to make that partnership work.

Read more: The Rundown

🌪️ AI Storms Into Weather Forecasting

Google DeepMind's Weather Lab model didn't just predict Hurricane Erin - it nailed the path and intensity of this Category 5 monster, outperforming both the National Hurricane Center's official forecasts and the gold-standard physics models that meteorologists have relied on for decades.

This isn't just about getting tomorrow's weather right. This is about fundamentally disrupting an entire industry that's been doing things the same way since the dawn of modern meteorology.

The Old Way vs. The New Way: For decades, weather forecasting has been all about physics-based simulations. Massive supercomputers crunching through atmospheric equations, burning through computational cycles like there's no tomorrow. It's expensive, it's slow, and apparently, it's not as accurate as we thought.

Enter Google's AI model, trained on historical storm data. No complex physics equations. No supercomputer farms. Just pattern recognition on steroids. And it's faster, cheaper and more accurate than the traditional approach.

Who should be paying attention?

Forecast Agencies: The National Weather Service and similar organizations worldwide suddenly have competition from tech companies
Insurance Companies: More accurate predictions could revolutionize risk assessment and pricing models
Energy Traders: Better weather forecasts mean better predictions of energy demand and renewable energy output
Disaster Response: Emergency management agencies could get more reliable advance warning

The Bigger Picture: This isn't really about weather ... it's about predictive infrastructure. As AI Secret points out, "the moat isn't supercomputers, it's data pipelines and model trust". Google just demonstrated that with the right data and the right AI model, you can outperform decades of specialized expertise and billions of dollars in traditional infrastructure.

What This Means for Other Industries: If AI can beat meteorologists at weather prediction, what other "expert domains" are ripe for disruption? Financial forecasting? Medical diagnosis? Supply chain optimization? The implications are remarkable.

The question isn't whether AI will transform predictive modeling across industries. The question is: who will own that transformation?