/ / THEFUTURE /
In this issue: MIT’s questionable 95% AI failure report. Google’s “Nano Banana” makes faking influencers stupidly easy. Nieman Lab’s Andrew Deck on why aggregation jobs are disappearing faster than anyone realizes. Plus: How I built a web app with GPT-5 in one hour.
What we’re talking about: Attention-grabbing headline alert! There’s a project called Nanda at MIT that imagines trillions of AI agents collaborating and transacting on the web. Out of this project comes a report claiming 95% of organizations get zero return on their GenAI investment. “Wall Street’s biggest fear was validated,” writes Axios.
Before you say “told you so,” let’s actually look at the report. It’s based on 52 interviews with “stakeholders” and 183 “leaders” who answered a survey. The bottom line: Large companies spent money on AI pilots that didn’t deliver measurable ROI in the first six months.
The report blames organizational barriers. While people use AI for personal tasks, they struggle with enterprise tools that don’t learn, integrate poorly, or clash with company workflows. Throughout the report, the authors keep mentioning agents and their Nanda project as solutions. Sounds more like a sales pitch than science.
Liking ChatGPT is now a hot take: After the botched launch of the new ChatGPT version, there were plenty of comments. The AI hasn’t gotten better, just the packaging. Now Ezra Klein jumps in with a lengthy opinion piece. He doesn’t really know where AI is heading either. It’s all very complicated, better not commit to anything. His wildest take: He finds the new ChatGPT useful. He writes:
This is the first A.I. model where I felt I could touch a world in which we have the always-on, always-helpful A.I. companion from the movie “Her.”
The dystopian movie from 2013 shows a lonely man desperately in love with a chatbot. If you ask me, this is not something we should be aiming for.
Heads up: Faking influencers is about to get more easy. Until now, making an AI persona look consistent across different angles and images took some manual work. That’s changing with a new image model that’s been testing under the codename “Nano Banana.” Word is Google’s behind it, and a Google DeepMind employee is teasing something for this week. So, here we go.
What else I’ve been reading:
Displacement reality check: While we debate whether AI will replace journalists, Nieman Lab’s Andrew Deck has been documenting what’s actually happening: thousands of AI-generated local newsletters, fully automated news roundups, and the quiet erosion of aggregation jobs. His perspective on the real threat and the tools that could win you a Pulitzer.
Three Questions with Andrew Deck
Andrew Deck is the AI Staff Reporter at Nieman Journalism Lab.
What's the most important question right now?
How is my newsroom going to weather the erosion of news aggregation?
Editorial roles centered on aggregation — curating and synthesizing existing reporting — are among the jobs most at risk of displacement by generative AI tools and by startups operating outside traditional news organizations.
Over the past decade, an era often defined by social media traffic and SEO gamesmanship, newsrooms invested heavily in editorial teams that aggregated rather than producing original journalism. News sites were flooded with stories recycling other outlets’ reporting for their own audiences, or with articles made entirely of reworked press releases and curated social posts. The business models of digital journalism incentivized this kind of aggregation. Even now, many editorial staffers spend an entire day producing aggregations without ever speaking to a source.
Generative AI tools cannot displace shoeleather reporting. Whether they should be or not, these tools are increasingly being used to automate significant parts of the aggregation workflow.
I’ve been reporting on networks of hundreds (and even tens of thousands) of AI-generated local newsletters spreading across the U.S. At Good Daily and Patch, daily roundups that were previously produced by dozens of writers and editors are now fully automated. Startups like NoahWire and Open Mind are experimenting with automating parts of the discovery, drafting, and even fact-checking for aggregated articles. In all the chatter about newsroom displacement, I think aggregation still feels sorely underestimated as a space on the cusp of a major reinvention.
What will we be shaking our heads about a year from now?
We’ve devoted so much of our strategic energy to how AI tools might deliver minimal gains in newsroom productivity — without properly sandwalling our editorial operations against the rise of generative search.
In my view, the so-called “traffic apocalypse” hasn’t been borne out in the numbers… yet. That said, I welcome the urgency and attention recent coverage has brought to how generative AI tools are changing our audience landscape. There is good reason to be sounding the alarm.
Until recently, though, most newsroom leaders I’ve spoken to have been more concerned with whether or not a ChatGPT Enterprise account would shave hours off an investigation, than about whether generative search products will kill the subscription conversions that ultimately fund that reporting once it’s published.
What future are you looking forward to?
When machine learning techniques long pioneered by data journalists are accessible to the average reporter.
Over the past couple years, I’ve profiled a handful of Pulitzer Prize winners and finalists that disclosed using AI in their work. These adoption use cases run the gamut, including creating complex data visualizations and training custom object detection models. One trend that stood out: some awardees were using off-the-shelf tools, or using conversational AI models that required no knowledge of Python or R.
The barrier to entry for using machine learning is lowering by the day, and I can’t wait to see more journalists without engineering backgrounds get their hands on these powerful reporting tools.
Hands on: There is one thing I found GPT-5 to be really good at. You might have heard about AI coders that work with a dozen agents and build elaborate systems for software development. Not me! I just gave GPT-5 one simple instruction and got a working prototype in about one hour.
What I wanted to build: this. A map of running clubs, but for my city. This was my prompt:
Let's build a website. First, let's plan. I have a Google Spreadsheet or a public Notion page with the weekday, time, website, notes, and location of running clubs in Hamburg. The website header has the weekdays from Monday - Sunday. List all running clubs sorted by time for the given day. Then there is a nice map of Hamburg that displays the locations for the current weekday. Show labels with time. You can click the labels with time and should see in the list which run club it is. Any questions so far?There were some minor questions, and just like that, I had a web app. Which I uploaded to GitHub so that I could deploy it to Vercel. If you’re not familiar with these tools, don’t worry, neither was I. But you already know who helped me learn it. Just did it. Cold plunge.
Don’t you worry: The API key for the map server is stored securely and limited to that particular domain. My bottom line: Just keep it simple.
This is THEFUTURE.
