Sections
00:00 What we're covering today
00:30 Bank of America added 1,000 contractors to fix flaws found by AI
01:33 Gartner warns that over $3.5 trillion in AI infrastructure spending is at risk
02:35 Google opens Deep Review for technical documents
03:37 Researchers correct recurring errors in weather forecasts weeks in advance
04:40 Singapore and DeepMind build a phone-based scam detector
05:35 Visit Hot Tea

Disclosure
Narration uses an AI-generated voice.

Transcript
Welcome to Hot Tea for Thursday, October 15, 2026. This briefing uses an A I-generated voice.

Bank of America's software patching workload leads today's news. Its chief executive says a thousand contractors joined earlier this year, and they're still working. We'll also cover a forecast for A I infrastructure costs and Google's document reviewer. Later, we'll look at a method for correcting weather forecasts and Singapore's planned scam detector for phones.

Bank of America added 1,000 contractors to fix flaws found by A I. CEO Brian Moynihan described the staffing in an October 14 CNBC interview. He said the contractors joined in the second quarter and are still working alongside about 3,000 bank employees in that area. Their work covers the bank's own software and code from vendors. Moynihan said advanced A I models, including Anthropic's Mythos, found the flaws. CNBC reports that software fixes and anti-money-laundering compliance added hundreds of millions of dollars to third-quarter costs. That's a combined figure, not a separate bill for Mythos. The staffing figures come from the CEO, not an independent audit of security gains. Moynihan expects remediation costs to decline. The disclosure doesn't establish an attack or breach, and it doesn't show that all 1,000 contractors work only on Mythos findings. Now the useful measures are how long verified flaws stay open, whether fixes cause failures, and whether the contractor workload falls.

Gartner warns that over $3.5 trillion in A I infrastructure spending is at risk. The firm's October 14 forecast puts that risk at the end of 2030. It assumes providers keep spending at the current pace. Gartner estimates that data center owners would need $8.5 trillion in cumulative revenue through 2030. That revenue would cover spending, financing, and a 12% return on invested capital. This is a forecast, not a measured loss. Gartner is not predicting that all exposed assets will be written down. It forecasts no more than $3.6 trillion in related A I revenue over that period. The assumed return and the way Gartner divides the revenue shape the gap. Gartner also predicts that specialist A I cloud providers will consolidate. That does not mean every company that leaves the market will close. Compare actual revenue, financing costs, and equipment replacement schedules with the forecast. Those figures would test whether owners can earn back their investment.

Google opens Deep Review for technical documents. Google opened the service to the public on October 14. Each user gets two free document reviews per week, and the first version accepts PDFs and plain text. It checks arguments and technical claims. Deep Review builds on a pre-submission assistant tested at four conferences with authors of more than 16,000 manuscripts. Google reports feedback from those authors, not proof that the tool can replace expert review. The pilots were also separate from official conference decisions. The International Conference on Machine Learning published a March survey of authors, but it didn't establish an independent error rate. Google says it keeps files uploaded to the public service for one year and doesn't use them to train or improve its models. Those terms differ from the earlier pilots, so organisations still need to decide which documents they may upload. The practical test is whether the tool's objections hold up and whether corrected documents contain fewer substantive errors.

Researchers correct recurring errors in weather forecasts weeks in advance. A Nature paper published on October 14 describes a method called Probabilistic Bias Correction. It learns the recurring errors in existing weather forecasts. In tests of an operational system, researchers report improvements across 98% of precipitation targets, 92% of temperature targets, and 91% of pressure targets. Those figures show how many targets improved, not the percentage increase in accuracy. The main tests cover weekly conditions for days 19 to 25 and days 26 to 32. They use a grid resolution of 1.5 degrees. The results do not promise precise local weather six weeks ahead. Microsoft says Kenyan forecasters are testing the method. Its commercial six-week service remains in private preview. The paper first appeared as a preprint in April. Wednesday marked its journal publication and Microsoft’s account of the deployment. Local performance and farmers’ decisions would show whether better probability forecasts reduce losses. Benchmark gains alone do not establish that result.

Singapore and DeepMind build a phone-based scam detector. Google DeepMind and Singapore's GovTech announced the on-device model on October 14. A small classifier checks messages first. A Gemma model then examines suspicious conversations for signs of manipulation. Minister Josephine Teo confirmed the collaboration in a speech that day. DeepMind says users would have to consent, and their conversations would remain on the device. The company plans to add the model to applications in phases, so the announcement doesn't establish nationwide availability. DeepMind provided no figures in this account for accuracy, false alarms, battery use, or fraud losses. Keeping the analysis on the phone limits what leaves the device, but it doesn't prove that every warning is correct. App availability, clear consent controls, and measured false alarms will test the proposal. Prevented losses matter more than the number of warnings.

That is the signal before the noise. This briefing was produced from Hot Tea's verified daily edition. For the complete briefing and every source link, visit Hot Tea dot A I.