Hello readers,
Welcome to the AI For All newsletter! Today, we’re talking about an AI second reader for routine chest CTs, agentic AI reducing manufacturing downtime, and more!
AI in Action: Pushing past what the doctor ordered
A chest CT gets ordered for the lungs, but the scanner captures the esophagus too, whether or not anyone had it in mind. A study in Nature Medicine describes EAGLE, a model built to examine it on routine noncontrast scans, a task its authors say was long considered impossible. The esophagus is a narrow tube that collapses on itself and shifts with the heartbeat, and early lesions stay in the superficial layers of its wall. EAGLE first locates the esophagus in the 3D scan, then scores any lesion for malignancy and draws a heat map of the regions behind its call.
The model trained on 6,813 patients from two centers and was validated on 80,612 people at 12 centers in China, the Czech Republic, and Australia. In external tests covering 11,466 patients at eight centers, it caught 90% of cancers at 98.5% specificity. Early disease is harder: sensitivity was 60.1% for stage 1 cancer and 52.5% for precancerous lesions. In a reader study, 17 radiologists went through the same 300 scans twice, alone and then with EAGLE, and their average sensitivity rose from 71.9% to 85.7% while specificity climbed from 79.6% to 91.7%.
The team also adapted EAGLE to the low-dose scans used for lung cancer screening, where it reached 88.4% sensitivity and 99% specificity on 1,607 scans. Among 10,959 people in routine low-dose exams, it flagged eight scans. One held a finding missing from the original report, and the patient was diagnosed with esophageal cancer eight days later. Run on old scans from 28 patients who were later diagnosed, it flagged 18, four of them at least nine months early. The caveats sit mostly in who was studied: the training data came entirely from China, where squamous cell carcinoma is more common, and the model did worse near the junction of esophagus and stomach, where Western adenocarcinoma clusters. It isn't an approved screening test, and the paper reports no data on survival.
🔥 Rapid Inferno 🔥
Commentary: Dead Money
Analysis: The Hater’s Guide to AI Debt (Part 2)
Anthropic lost $8 billion in 2025 on $4.6 billion in revenue
Net loss of $42 billion and operating loss of $8 billion
Operating loss widened from $2.98 billion in 2024
~25% of revenue came from two customers
“Anthropic said nearly a quarter of its revenue came from two customers last year, and as part of its risk factors, warned that many of its largest clients were not locked into long-term contracts and could cut or stop spending.”
Anthropic needs $518 billion to pay its compute deals
Anthropic has $252.5 billion in non-cancelable compute obligations across Microsoft, Google, and Amazon and $161.2 billion in Broadcom TPU lease obligations that are “largely non-cancelable”
In yet another flagrant case of circular financing, Broadcom agreed to lend Anthropic up to $42 billion to lease its chips
“A vendor who is fairly close to the company told me that they [Anthropic] underperformed their expectations for the third quarter.” — Economist Paul Kedrosky at 44:26
Financial Times: OK, so where are all these data centers?
“Morgan Stanley measured the gap earlier this week by estimating the shortfall in available power, concluding that more than half of the GPU servers sold between 2026 and 2028 might not have anywhere to be plugged in.”
“In the longer term, power availability is still the bottleneck — along with labour. And transformers. And cooling equipment. And backup generation. As Jefferies says: The gap between planned capacity and physical execution remains the central issue.”
Further reading (per FT): Where Are All The AI Chips?
Power delays put 2027 target at risk for Oracle data center
AI data centers are a $6 trillion time bomb
Cracks emerge in AI’s debt-fueled data center boom
“Some lenders are becoming more selective about the deals they’ll participate in. That narrows the pool of banks available to syndicate large multibillion-dollar loans needed for data center projects and other AI infrastructure. Société Générale and SMBC are among the firms that have become more selective, while MUFG is also stepping back.”
Businesses still don’t see AI returns per EY consultant
FTC opens probe into OpenAI and Anthropic
Addendum: AI agents are not going “rogue” — what’s actually happening
This anthropomorphized language attempts to remove responsibility from the AI labs for committing felony hacking. They are entirely in control of their software. The software is doing what it was intended to do.
Jacob Coxon worked with doomer PR firm despite claiming otherwise
The effective altruism cult actually believes LLMs can suffer
China has cracked down on AI relationships. Is it ahead of the game?
📖 What We’re Reading
Downtime remains one of manufacturing's most expensive problems. It can interrupt production, delay orders, increase labor pressure, and raise maintenance costs across the plant. The National Institute of Standards and Technology (NIST) estimates that U.S. manufacturers lost $18.1 billion to unplanned downtime tied to preventable maintenance issues.
