Heya and welcome back to Five Things AI!
I had coffee with a business angel this morning and spent twenty minutes losing an argument about whether OpenAI survives the decade, which is a fitting start to a week where every story I read pointed in the same direction: the expensive stuff is getting harder to justify. Ed Zitron does the arithmetic on OpenAI's 800 billion in compute obligations, a research report suggests small models can already handle four out of five use cases at a seventh of the energy, and almost nobody seems to be using Fable 5 because the premium simply isn't worth it. Meanwhile Flock built the driver-tracking tool it spent years insisting it could never build, and a journalist worries that AI writing and AI detectors together are dragging his own prose toward the median. As a non-native English speaker who reads AI output daily for hours, I do wonder how this will change my own command of the language. I guess I have to pick up some pre-2022 writing again…
Read Five Things AI mostly about the gap between what things cost and what they are actually worth! 🤔
What Happens If OpenAI Dies?
The reason that OpenAI (and Anthropic, for that matter) wants you to think about things in terms of “annualized revenue” is because its actual revenues look a little tame compared to its commitments and burn rate. The Information reports that in Q1 2026, OpenAI burned $12.1 billion on “cost of revenue” and training on $5.7 billion in revenue, though it left out the sales and marketing segment where OpenAI burned $5.73 billion in 2025 — or, put another way, OpenAI spent $12.1 billion on compute to lose $6.4 billion, and that doesn’t include things like data costs or salaries or, well, anything. OpenAI (and by proxy The Information) somehow rationalizes this to only be a burn of $3.7 billion, likely using the same accounting bullshit that it did in the financials I saw.
Now, some of you might read that and say “wow, $5.7 billion is a lot of money!” but it doesn’t matter, because the more money OpenAI makes, the more its services cost. This is not difficult mathematics, but it is something that continues to escape the vast majority of coverage of the company, I assume because all of this feels a little insane when you think about it.
Many people hope that OpenAI is too big too fail and that Sam Altman has developed a reality distortion field just like Steve Jobs. It currently doesn’t look like that as OpenAI continues to burn through money and there is no end in sight. More focus and less side-projects for Sam Altman are one thing, but the other strategic flaw is building a large user base that doesn’t want to pay.
If this is true, the hyperscalers are toast
This already means that one can replace data centres and their expensive cutting-edge semiconductor infrastructure in four out of five use cases. The research report estimates that the addressable market in the US for SLMs has grown to about $10tn or one-third of the entire US GDP of $30tn. There isn’t much left for LLMs to thrive in, and every year, their advantage over SLMs is shrinking.
There clearly are areas where LLMs are still way ahead, particularly in agentic AI applications, where SLMs currently only achieve accuracy and success rates of less than 50%. Similarly, it is difficult to run these SLMs on smartphones so far. The models that can be run on an iPhone are significantly worse than the models that can be run on a desktop PC.
But – and this is important – the models run on a desktop PC, and even more so, the ones run on a smartphone are much more energy efficient than the ones run in the cloud. The inference per Watt of these SLMs is typically seven times larger than that of LLMs. And that means that when you encounter a task that can be solved on a desktop or even a mobile device, it is cheaper to do so locally than send it to a data centre.
So, while we focus on the big and mighty frontier models, the midpack models are getting so good, they can run on our laptops.
Ooops.
Almost Nobody Is Using Anthropic’s Fable 5
“Enterprises today are thinking through the best performance per dollar, or intelligence per watt,” says Gupta.
Gupta also shares her own experience using Fable 5, describing its outputs as increasingly difficult to read and follow. She says the model often uses dense language, unfamiliar terminology and long, meandering explanations that can make the user feel as though they are struggling to keep up with the model rather than benefiting from it.
The problem, she argues, is not simply that the model is highly capable, but that its communication style can create an “intelligence delta” between the model and the person using it. In practice, this can make Fable 5’s answers more cognitively demanding to interpret, particularly for users who do not share the specialised knowledge or vocabulary embedded in its responses.
I’m not sure that Fable is really too smart, it is just too expensive and for the premium I have to pay, I get results that might be better than with Opus, but not so much better that I am willing to burn through tokens that quickly.
Flock Has a Powerful New AI Tool for Police. We Got Its Code
Vehicle surveillance giant Flock Safety has told the public for years that its technology “cannot recognize, identify, or track individuals.” It has now built a system that does both, an artificial intelligence tool for police that can identify drivers and track vehicles by their patterns of movement alone, WIRED has learned.
Drawing on a network of cameras that logs the movements of drivers in more than 6,000 communities, the tool can pick out potential witnesses by how often their cars pass through a neighborhood, or surface a driver’s “associates” from the cameras they pass together. Because the system also reaches police case files, 911 dispatch logs, and commercial identity records, those plates can be turned into names, home addresses, and relatives. It can search for people in an area drawn on a map based on nothing more than a physical description.
Oooops. I don’t think this improves the confidence citizens have in this product family.
A Controversial Technology Is Making Me a Worse Writer. No, Not That One.
Whatever the future holds for the accuracy of A.I. detectors, I am unsettled by what both the A.I. writing and the detectors are doing to me right now. I worry about how A.I. writing and efforts to sniff it out are changing fully human writing. I write all week, every week, and can feel this story’s arsonists (LLMs, particularly when used the wrong way) and firefighters (A.I. detectors) weighing on my process in negative ways.
Like many journalists, my writing style comes from the books, news articles, blogs, and (yes) social media posts that have latched on to my brain over the years. A.I. writing is now in my water supply, whether I can identify each new gulp or not. It feels inevitable that I’ve started to internalize its tics, even the most infamous ones, like incessant em-dashing and overreliance on the “It’s not x but y ” structure. For one thing, I fear this will make me a worse writer, dragging a talent that was good enough to get me hired here (and elsewhere!) back toward the A.I.-generated median. Scarier yet, it’s possible that someone—or, heaven forbid, some A.I. detector—might mistake my text for ChatGPT’s.
I really am not so sure where I stand on this issue. I think it makes sense to flag AI generated content, even though most of us can detect AI slob from a mile away. It also makes sense that students produce their own work and do not turn in papers written by agents. But I think otherwise about my own work. I communicate with Claude daily and quite often let Claude write summaries. I hate writing summaries. And then Pangram comes along and claims that there is AI generated text. Well. Do the readers really care? It’s not that I let the AI fabricate whatever it wants, I still do the research and most of the writing here…
